xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001994145
007 cr mnu|||uuuuu
008 090330s2008 flu s 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0002429
Remsen, Andrew Walker.
Evolution and field application of a plankton imaging system
h [electronic resource] /
by Andrew Walker Remsen.
[Tampa, Fla] :
b University of South Florida,
Title from PDF of title page.
Document formatted into pages; contains 128 pages.
Dissertation (Ph.D.)--University of South Florida, 2008.
Includes bibliographical references.
Text (Electronic dissertation) in PDF format.
ABSTRACT: Understanding the processes controlling the distribution and abundance of zooplankton has been a primary concern of oceanographers and has driven the development of numerous technologies to more accurately quantify these parameters. This study investigates the potential of a new plankton imaging sensor, the shadowed image particle profiling and evaluation recorder (SIPPER), that I helped develop at the University of South Florida, to address that concern. In the first chapter, results from the SIPPER are compared against concurrently sampling plankton nets and the optical plankton counter (OPC), the most widely used optical zooplankton sampling sensor in the field. It was found that plankton nets and the SIPPER sampled robust and hard-bodied zooplankton taxa similarly while nets significantly underestimated the abundance of fragile and gelatinous taxa imaged by the SIPPER such that nets might underestimate zooplankton biomass by greater than 50%.Similarly, it was determined that the OPC misses greater than a quarter of resolvable particles due to coincident counting and that it can not distinguish between zooplankton and other abundant suspended particles such as marine snow and Trichodesmium that are difficult to quantify with traditional sampling methods. Therefore the standard method of using net samples to ground truth OPC data should be reevaluated. In the second chapter, a new automated plankton classification system was utilized to see if it was possible to use machine learning methods to classify SIPPER-imaged plankton from a diverse subtropical assemblage on the West Florida Shelf and describe their distribution during a 24 hour period. Classification accuracy for this study was similar to that of other studies in less diverse environments and similar to what could be expected by a human expert for a complex dataset.Fragile plankton taxa such as larvaceans, hydromedusae, sarcodine protoctists and Trichodesmium were found at significantly higher concentrations than previously reported in the region and thus could play more important roles in WFS plankton dynamics. Most observed plankton classes were found to be randomly distributed at the fine scale (mm-100 m) and that greatest variability within plankton abundances would be encountered vertically rather than horizontally through the water column.
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
Advisor: Thomas L. Hopkins, Ph.D.
Gulf of Mexico
x Marine Science
t USF Electronic Theses and Dissertations.
Evolution and Field Application of a Plankton Imagi ng System by Andrew Walker Remsen A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy College of Marine Science University of South Florida Major Professor: Thomas L. Hopkins, Ph.D. Douglas C. Biggs, Ph.D. Frank E. Mller-Karger, Ph.D. Joseph J. Torres, Ph.D. John J. Walsh, Ph.D. Date of Approval: January 28, 2008 Keywords: zooplankton, Gulf of Mexico, machine lear ning, sampling systems, optics Copyright 2008, Andrew Remsen
i Table of Contents List of Tables iv List of Figures vi Abstract xi Preface xiii Chapter 1: What You See Is Not What You Catch: A Co mparison of Concurrently Collected Net, Optical Plankton Counter (OPC), and Shadowed Image Particle Profiling Evaluation Recorder (SIPPER) Data from t he Northeast Gulf of Mexico. 1 Introduction 1 Methods 3 Net Sample Treatment 4 SIPPER Data Analysis 5 OPC Data Analysis 9 Results 11 Hydrography 11 Mesozooplankton and Mesozooplankton-Sized Particl e Abundance 11 OPC Abundance Correction and Coincidence Frequenc y 16 Size Frequency Distributions 17 Sample Biovolume Estimates 20 The Problem of Trichodesmium 21 Taxonomic Composition, SIPPER vs. Nets 22 Total Biomass 29 Taxonomic Differences in Size Distribution 29 Discussion 32 Occurrence of Unidentified Particles in Plankton Image Datasets 32
ii Importance of Trichodesmium in Subtropical Systems 33 Marine Snow and Large Phytoplankton as Signal Rat her than Noise 33 Comparison between Different Sampling Methods 33 Implications for Subtropical Oceanic Biology 35 Effects of Formalin Preservation on Sample Size D istribution 36 Comparison of SIPPER Performance with Other Imagi ng Systems 36 Conclusions 37 Chapter 2: Describing plankton distribution and abu ndance in neritic sub-tropical waters using SIPPER-2 and an automated classification system. 38 Introduction 38 Methods 41 Description of SIPPER Sensor 44 Detection and Extraction of SIPPER Images 46 Manual Classification of Images and Development o f Training Library 47 Automatic Classification of Images 49 Classification Error Analysis 50 Analysis of SIPPER-2 Spatial Data 50 Results 55 Hydrography 55 Total SIPPER-2 Particle Abundance and Size Distri bution 56 Selection of Image Groups and Classifier Performa nce 56 Human Expert Classification Error Estimation 62 Example Images 63 Taxonomic composition of Imaged Dataset 74 Abundance and distribution of SIPPER-2 Image Clas ses 75 Other and Unidentified Class 75 Zooplankton 77 Trichodesmium 87
iii Protoctists 90 Marine Snow 92 Net Sample Analysis and Comparison with Concurren t SIPPER Data 94 Fine Scale Distribution of SIPPER-2 Image Classes 96 Discussion 106 Classifier Performance and Comparison with Other Field Deployed Classifiers 106 Limitations of SIPPER Imaging System 107 Human Classification Error 108 Comparison with Plankton Net Data 109 Ecological Implications for the WFS 110 Possibility of Diel Vertical Migration of the WFS Plankton Assemblage 113 Observations of Plankton Behavior 114 Fine Scale Distribution of SIPPER Particle Classe s 114 Conclusions 116 References 117 About the Author End Page
iv List of Tables Table 1: Dry Weight Biomass Regressions for zooplan kton from Oceanic Waters of the Gulf of Mexico. 7 Table 2: Abundance (number m-3) of mesozooplankton sized particles as estimated by the 162 m m net system, SIPPER and the OPC. 15 Table 3: OPC normalized particle abundance, theoret ical coincidence factor, SIPPER estimated coincidence percentage, count los s and corrected OPC particle abundance. 19 Table 4: Net estimated abundance (Number m-3) and biomass (mg m-3) of zooplankton groups 25 Table 5: SIPPER estimated abundance (Number m-3) and biomass (mg m-3) of zooplankton groups 26 Table 6: List of features extracted from every SIPPER-2 imag e and those used for the classifier after running the Â“wrapperÂ” feature optimization algorithm. 49 Table 7: Image classes used during development of t he multiple-class SVM classifier used for this study. 60 Table 8: Confusion matrix of 31 class SVM classifie r for the SIPPER test set (109122 particle images). 64 Table 9: Confusion matrix of 21 class SVM classifie r for the SIPPER test set (109122 particle images). 65 Table 10: Confusion matrix of classification perfor mance between the primary human classifier of SIPPER images (TH1) and a secondary human expert (TH2) in sorting a subset (N=2178) of the SIPPER test set. 66
v Table 11: Average minimum size of net collected zoo plankton used for comparison of the sampling effort between plankton nets and t he SIPPER-2. 95 Table 12: Comparison of day and night SIPPER zoopla nkton abundance estimates versus the abundance estimates of similarly sized zooplankton collected by concurrently sampling 162 m m plankton nets. 100 Table 13: Results of statistical analysis of spatia l patterns in the shallow daytime test set sample. 101 Table 14: Results of statistical analysis of spatia l patterns in the deep daytime test set sample. 102 Table 15: Results of statistical analysis of spatia l patterns in the shallow nighttime test set sample. 103 Table 16: Results of statistical analysis of spatia l patterns in the deep daytime test set sample. 104
vi List of Figures Figure 1: Schematic of the USF High Resolution Samp ler indicating location of the described zooplankton sampling systems 4 Figure 2: Diagram illustrating the general concept of the SIPPER line-scan camera system. 6 Figure 3: An example 2048 2048 pixel frame illust rating the large sampling area of SIPPER and its image quality. 10 Figure 4: Representative SIPPER images of the thirt een enumerated plankton groups. 12 Figure 5: Temperature, salinity, fluorescence and e xtracted chlorophyll (denoted by *) profiles (0-200 m) from the study site 13 Figure 6: Numerical abundance estimates of the three zooplank ton sampling methods 14 Figure 7: Cumulative size-frequency distribution spectra for the three sampling Sensors (300-5000 m m ESD). 18 Figure 8: Proportion of SIPPER identified plankton to the SIP PER image total per 100 m m ESD size class. 20 Figure 9: Cumulative sample biovolume ( mm3 10m-3) versus size (300-5000 m m ESD) distribution. 21 Figure 10: Zooplankton abundance (numbers m-3) profile determined from the net and SIPPER after Trichodesmium abundance was separated from the SIPPER i.d. dataset. 22 Figure 11: SIPPER (solid line) and net (dotted line) numerical abundance estimates of zooplankton groups with significant differences be tween the two sampling systems (paired t-test, p <0.05). 27
vii Figure 12: SIPPER (solid line) and net (dotted line) numerical abundance estimates of zooplankton groups with no significant differences between the two sampling systems (paired t-test, p <0.05). 28 Figure 13: Cumulative abundance of mesozooplankton sized particles or mesozooplankton determined by the three sampling m ethods separated into 500 m m size classes (left graphs) up to 1500 m m ESD 30 Figure 14: Cumulative abundance of mesozooplankton sized particles or mesozooplankton determined by the three sampling m ethods separated into 500 m m size classes (left graphs) greater than 1500 m m ESD 31 Figure 15: Photograph of SIPPER-2 mounted on the HR S. 41 Figure 16: Depth profiles from the seven deployment s of the HRS collected during this study. 42 Figure 17: Comparison of a similarly sized salp ima ge from the SIPPER-1 binary imaging sensor and the SIPPER-2 3-bit grayscale im aging sensor demonstrating the increased detail visible with SI PPER-2. 46 Figure 18: Examples of regular (a), random (b), and aggregated distributions with their cumulative histograms superimposed on the cumulati ve histograms generated by Monte Carlo simulations. 54 Figure 19: Vertical profiles of A: temperature, B: salinity and C: sigma-t collected during the 24 hour sampling period. 55 Figure 20: Vertical profiles of A: fluorescence and B: extracted chlorophyll collected during the 24 hour sampling period. 56 Figure 21: Distribution and abundance of all partic les (no bubbles and scanlineartifacts) imaged by SIPPER greater than 550 m m ESD in size over the 24 hour sampling period. 58 Figure 22: Size distribution in equivalent spherica l diameter of the total SIPPER Particle images collected during this study (N=139 1227). 59
viii Figure 23: SIPPER images of A: the cladoceran Penilia avirostris and B: The ostracod Euconchoecia chierchiae. 67 Figure 24: SIPPER images of chaetognaths. Several b ehaviors are visible in these images: beginning clockwise from the top, reproduct ion, parasitism, defecation, cannibalism and predation. 67 Figure 25: SIPPER images of the trachymedusa Aglaura hemistoma. 68 Figure 26: SIPPER images of hydromedusae and narcom edusae making up the other cnidarian class. 68 Figure 27: SIPPER images of various calanoid copepo d species. 69 Figure 28: SIPPER images of A: the cyclopoid copep od genus Oithona B: The poecilostomatoid copepod species Macrosetella gracilis and C: the poecilostomatoid copepod genus Oncaea 69 Figure 29: SIPPER images of eumalocostracan crustac eans. 70 Figure 30: SIPPER images of A: echinoderm plutei an d B: Elongate phytoplankton colonies 70 Figure 31: SIPPER images of doliolids 71 Figure 32: SIPPER images of larvaceans. 71 Figure 33: SIPPER images of marine snow. 72 Figure 34: SIPPER images of the other and unknown p article class. 72 Figure 35: SIPPER images of various forms of sarcod ine protoctists. 73 Figure 36: SIPPER images of A: Elongate trichomes a nd linear colonies of Trichodesmium and B: Tuft and Puff colonies of Trichodesmium 73 Figure 37: Composition of the SIPPER dataset as det ermined through the three multiple class SVMs run on the dataset 74 Figure 38: Size distribution in ESD for the non-zoo plankton image classes. 76 Figure 39: Distribution and abundance of the other and unidentified particles image class collected during this study. 76
ix Figure 40: Distribution and abundance of the 13 SIP PER imaged zooplankton classes chosen for the final classification. 77 Figure 41: Distribution and abundance of larvacean images collected during this study. 78 Figure 42: Size distribution in ESD for the non-cru stacean zooplankton image classes. 79 Figure 43: Distribution and abundance of calanoid c opepod images collected during this study. 80 Figure 44: Size distribution in ESD for the crustac ean zooplankton image classes. 81 Figure 45: Distribution and abundance of ostracod i mages collected during this study. 82 Figure 46: Distribution and abundance of Aglaura hemistoma images collected during this study. 83 Figure 47: Distribution and abundance of Oithona sp. images collected during this study. 84 Figure 48: Distribution and abundance of chaetognat h images collected during this study. 84 Figure 49: Distribution and abundance of echinoderm plutei including both echinopluteus and ophiopluteus images collected during this study. 87 Figure 50: Distribution and abundance of unidentifi ed hydromedusae and narcomedusae making up the other cnidarian class. 87 Figure 51: Distribution and abundance of the poecil ostomatoid copepod genus Oncaea images collected during this study. 88 Figure 52: Distribution and abundance of eumalocost racan crustacean images collected during this study. 88 Figure 53: Distribution and abundance of doliolid i mages collected during this study. 89 Figure 54: Distribution and abundance of the poecil ostomatoid copepod Macrosetella gracilis images collected during this study. 89 Figure 55: Distribution and abundance of Trichodesmium colony images collected during this study. 91
x Figure 56: Distribution and abundance of sarcodine protoctist images collected during this study. 93 Figure 57: Distribution and abundance of the elonga te phytoplankton images collected during this study. 93 Figure 58: Distribution and abundance of marine sno w images collected during this study. 94 Figure 59: ANND cumulative histogram of the observe d shallow daytime bubble class compared to the 1000 random distributions. 105 Figure 60: ANND cumulative histogram of the observe d deep daytime Oithona class compared to the 1000 random distributions. 105
xi Evolution and Field Application of a Plankton Imagi ng System Andrew Walker Remsen ABSTRACT Understanding the processes controlling the distri bution and abundance of zooplankton has been a primary concern of oceanographers and ha s driven the development of numerous technologies to more accurately quantify these para meters. This study investigates the potential of a new plankton imaging sensor, the shadowed imag e particle profiling and evaluation recorder (SIPPER), that I helped develop at the University o f South Florida, to address that concern. In the first chapter, results from the SIPPER are compared against concurrently sampling plankton nets and the optical plankton counter (OPC), the most wi dely used optical zooplankton sampling sensor in the field. It was found that plankton net s and the SIPPER sampled robust and hardbodied zooplankton taxa similarly while nets signif icantly underestimated the abundance of fragile and gelatinous taxa imaged by the SIPPER such that nets might underestimate zooplankton biomass by greater than 50%. Similarly, it was dete rmined that the OPC misses greater than a quarter of resolvable particles due to coincident c ounting and that it can not distinguish between zooplankton and other abundant suspended particles such as marine snow and Trichodesmium that are difficult to quantify with traditional sam pling methods. Therefore the standard method of using net samples to ground truth OPC data should b e reevaluated. In the second chapter, a new automated plankton classification system was utiliz ed to see if it was possible to use machine learning methods to classify SIPPER-imaged plankton from a diverse subtropical assemblage on the West Florida Shelf and describe their distribut ion during a 24 hour period. Classification accuracy for this study was similar to that of othe r studies in less diverse environments and similar to what could be expected by a human expert for a complex dataset. Fragile plankton taxa such as larvaceans, hydromedusae, sarcodine protoct ists and Trichodesmium were found at significantly higher concentrations than previously reported in the region and thus could play
xii more important roles in WFS plankton dynamics. Mos t observed plankton classes were found to be randomly distributed at the fine scale (mm-100 m ) and that greatest variability within plankton abundances would be encountered vertically rather t han horizontally through the water column.
xiii Preface This research concludes a long and circuitous fora y into the worlds of technology and marine biology in an attempt to develop and apply a new research tool to better understand the important role that zooplankton play in the ecology of the world oceans. I was lucky enough to arrive at the USF CMS during a time of dynamic grow th and innovation and become involved with a lab that was pushing the boundary towards applyin g new biological sampling methods to its research. This work was one of the primary drivers behind the development of what would later become the Center for Ocean Technology (COT). My preliminary research objective was to investiga te the oceanic zooplankton assemblage response to different water mass conditi ons in the Gulf of Mexico using what was then the state of the art optical plankton counter (OPC). Results from the OPC were to be calibrated using organisms collected from the multi ple net system sampler described by Tracey Sutton in his dissertation. However, we could never find any predictable relationship between what was counted and sized by the OPC and the zoopl ankton we collected in the nets. Fortunately, Dr. HopkinsÂ’s tenure as the first dire ctor of the Center for Ocean Technology (COT) gave me the opportunity to become closely associate d with COT engineers in researching and designing new technologies to augment and or replac e our use of the OPC in studying zooplankton. I initially collaborated with a gradua te student from Florida Atlantic University, Tom Wilcox, to develop and test a high-frequency broadb and sonar to count, size and image individual zooplankton and that would be mounted on our plankt on sampling platform. By acoustically imaging particles that would later be sampled by th e OPC, we hoped to be able to identify these particles and determine why there was a discrepancy betwe+en the nets and the OPC (Remsen et al., 1996). However, the necessary size resoluti on of the sonar precluded it from having any measurable range and therefore only imaged zooplank ton and other particles directly in front of it. When I presented preliminary data from the sonar at an ONR sponsored bioacoustics workshop,
xiv one of the conveners, Dr. Peter Wiebe from WHOI, re marked that with that range, I should think about using an imaging system instead of a sonar as the main reason for using acoustics is for its longer range. Soon after, I became deeply involved in the develo pment and field testing of the shadowed image particle profiling and evaluation re corder (SIPPER), a joint project between our lab and COT. The SIPPER was developed from an initi al idea between Dr. Hopkins and Larry Langebrake of COT as an imaging analog to the OPC. Data from the SIPPER gave us the opportunity to investigate planktic distribution at the scale of the individual zooplankter and observe plankton in-situ. Preliminary field work wi th the SIPPER provided evidence that both plankton nets and the OPC had serious problems in q uantifying the zooplankton assemblage in the Gulf of Mexico. This corroborated my observati ons from analyzing concurrently collected plankton net and OPC data collected during three ye ars of sampling at an offshore sampling station and finding no relationship at all between the two datasets. This research then focuses on work after the preliminary field deployment of the SIPPER. It does not include mention of the numerous cancelled research cruises, flooded pressu re vessels, short-circuited electronics, broken connectors, cables and winches that have mad e this such a special experience. The first chapter was published in Deep Sea Research I in Jan uary, 2004 (Remsen et al., 2004) and information regarding the distribution of Trichodesmium colonies determined in chapter two was included in Walsh et al., 2007. Finally, this project would not have been possible without the counsel, patience and support of my major professor, Tom Hopkins. His for esight in integrating new technologies with traditional zooplankton sampling methods made this work possible and he was instrumental towards the foundation of the Center for Ocean Tech nology. My labmates Dr. Tracey Sutton and Dr. Scott Burghart provided invaluable assistance, advice and much needed humor to my endeavors. Much of this work would have been imposs ible without the engineering ingenuity and skill demonstrated by Larry Langebrake, Dr. Scott S amson, Mike Hall, Bill Flanery, Eric Kaltenbacher, Chad Lembke, Jim Patten, Randy Russel l, Ray Carr and Gino Gonzales of COT. Additional help in maintaining the HRS and SIPPER w as provided by COT personnel Charlie
xv Jones and Joe Kolesar and the USF CMS shop crew, Ji m Mullins, Jim Mulholland and Rich Shmid. The automatic plankton recognition system us ed in Chapter 2 was developed in collaboration with Dr. Dmitry Goldgof and Dr. Larry Hall of the USF College of Computer Science and Engineering with their students Dr. Tong Luo an d Kurt Kramer. Assistance at sea was gratefully accepted from Eric Nelson, Bill Husar, D r. Jose Torres, Graham Tilbury, Jen Jarrell, Chris Simoniello and Richard OÂ’ Driscoll. Funding f or this research was provided for by the Office of Naval Research (Grants # N00014-94-0963, N00014 -04-1-0421, N0014-02-1-0266, N0001407-1-0802). I am especially grateful for my committ ee members Doug Biggs, Frank MllerKarger, Jose Torres and John Walsh and to Scott Sam son for agreeing to chair my defense.
1 Chapter 1: What You See Is Not What You Catch: A Co mparison of Concurrently Collected Net, Optical Plankton Counter (OPC), and Shadowed Image Particle Profiling Evaluation Recorder (SIPPER) Data From the Northeast Gulf of Mexico. Introduction Zooplankton are key mediators of particle flux, fis heries recruitment and biomass production within the world oceans (Lenz 2000). Inf ormation on their abundance and distribution in space and time are required to accurately predic t their contribution to these processes. Field observations of zooplankton indicate that they oper ate along a continuum of spatial and temporal scales leading to heterogeneous or Â“patchyÂ” distrib ution patterns (Haury et al., 1977; Omori and Hamner, 1982; Gallienne et al., 2001). Traditional methods such as plankton nets, pumps and bottles are limited in sampling zooplankton over th e entire distribution spectrum, especially at the fine scale (meters to hundreds of meters, seconds t o hours) because of their integrative nature and the time consuming task of analyzing individual zooplankton samples. Additionally, a significant fraction may be under-sampled by plankt on nets because of extrusion through the net mesh, retention within the net, and destruction of fragile forms such as gelatinous zooplankton when physically captured (Gallienne and Robins, 200 1; Halliday et al., 2001; Hopcroft et al., 2001; Warren et al., 2001). To address these limitations, alternative instrumen ts for sampling zooplankton in situ have been developed over the last twenty years (Sch ulze et al., 1992; Skjoldal et al., 2000; Wiebe and Benfield, 2003). These new devices provid e the increased spatial and temporal resolution necessary to study the coupling between physical processes and zooplankton distribution patterns and for modeling zooplankton population and tropho-dynamics. One of the most widely used of these new instruments is the op tical plankton counter (OPC), with approximately one hundred units in use throughout t he world (Zhou and Tande, 2002). The OPC provides quantitative measurements of abundance and size of mesozooplankton-sized particles
2 (250 m to 2 cm) and can be deployed from a diverse array of platforms (Foote, 2000). However, the taxonomic resolution of the OPC is limited exce pt in low diversity assemblages where separable peaks in a OPC generated size distributio n might be attributable to a specific species or developmental stage (Herman, 1992). Consequently the OPC is most often used to complement net data by providing high-resolution in formation on the spatial distribution patterns of the net-identified zooplankton. While many inves tigators have found rough agreement between net counts and OPC estimates of zooplankton abundan ce (Foote 2000, Zhou and Tande, 2002), there have been instances where the OPC and net abu ndance estimates have differed significantly (Grant et al., 2000; Halliday et al., 2001; Sutton et al. 2001). These differences have been attributed to extrusion of zooplankton through the plankton net mesh, counting of detrital aggregates and or large phytoplankton colonies, and coincident counting where the OPC counts multiple particles in the light path as a single la rger particle (Woodd-Walker et al., 2000; Zhang et al., 2000; Halliday et al., 2001). Advances in zooplankton imaging technology could he lp make sense of these conflicting results. Results from instruments such as the Video Plankton Recorder (VPR, Davis et al., 1992) and the Shadowed Image Particle Profiling and Evalu ation Recorder (SIPPER, Samson et al., 2001) indicate that they can provide both high qual ity taxonomic information and high resolution in the temporal and spatial domains. Previous compa risons between the VPR and nets have indicated that they describe similar distributions for abundant zooplankton groups (Benfield et al., 1996; Gallager et al. 1996), are more effective at sampling fragile and gelatinous forms than nets (Norrbin et al., 1996; Dennett et al., 2002), and c an assess the contribution of detrital aggregates or Â“marine snowÂ” to particles in the mesozooplankto n size range (Ashjian et al., 2001). This paper compares the abundance and size distribu tion of mesozooplankton and suspended particles sampled by nets and the OPC aga inst data concurrently collected by the SIPPER in offshore waters of the Gulf of Mexico. I hypothesized that the SIPPER should image all particles within the mesozooplankton size range that would be resolvable by either the net or the OPC and act to independently verify the other t wo sampling systems. The composition of the mesozooplankton assemblage sampled by the SIPPER an d plankton nets was then compared.
3 Methods Zooplankton were sampled at a station in the oceani c waters of the eastern Gulf of Mexico (27 N 86W, 3 km water depth) with the High Resolution Sampl er (HRS), a comprehensive towed marine particle analysis platfo rm (Figure 1; see Sutton et al. 2001 for a full description). The HRS samples zooplankton through a square 9.6 cm sampling tube (92.16 cm2 mouth area) leading to a 20-position cod-end net ca rousel fitted with 162 m m plankton nets. The nets have an open filtering area to mouth area rati o of 11:1 and the aluminum sampling tube precludes retention of organisms. Mounted inline wi th the sampling tube was the SIPPER zooplankton-imaging sensor. The sampling tube proj ects past the frame of the HRS and has a knife-edge to minimize any pressure wave that might develop in front of the aperture to reduce possible avoidance of the sampler by zooplankton. A n optical plankton counter (OPC) with a rectangular 2 22 cm sampling aperture (44 cm2 mouth area) was mounted within the frame of the HRS and a half-meter below the sampling tube. T he OPC was positioned so that it would sample water that was not influenced by the frame o f the HRS. Both net and electronic zooplankton sampling are co mputer controlled onboard ship via a custom designed software interface such that new SI PPER and OPC files are created and the sensors begin sampling when a net is triggered open When a particular net sample is ordered closed, the OPC and SIPPER files corresponding to t hat net sample are also closed, thereby creating two independent samples that can be compar ed against the net sample. Environmental and diagnostic information (CTD, fluorometer, trans missometer, inclinometer, and flow-meter data) are continuously recorded on a separate HRS d ata file for later analysis. A single deployment sampling 10 discrete depths (10 -100 m in 10 meter increments) beginning two hours after local sunset on July 21st, 2000, was chosen for this study. Each depth was sampled for 10 minutes. The SIPPER and the net system both collect zooplankton through the sampling tube at the front of the HRS and there fore sample the exact same volume of water. 37.9 m3 of water was sampled by these two systems for this study, averaging 3.79 (+ 0.18) m3 per depth stratum. The OPC, which was situated below an d slightly aft of the HRS sampling tube, has a sampling aperture approximately half that of the other two systems, and filtered a total of 18.39
4 m3 of seawater, averaging 1.84 (+ 0.09) m3 per depth stratum. Tow speed was determined with a calibrated flow meter (TSK Inc.) mounted at the fro nt of the sampler. It registered a near constant tow-speed of 0.75 m s-1. Inclinometer data indicated that the HRS maintai ned a near-perfect horizontal attitude at each depth stratum. Figure 1. Schematic of the USF High Resolution Sampler indica ting location of the described zooplankton sampling systems. Only one net is shown attached at the carousel to reduce confusion in the figure. Net Sample Treatment Zooplankton collected in the nets were fixed immedi ately in 5% v:v buffered formalin upon recovery of the sampler and stored for later a nalysis in the laboratory. Net samples were split into subsamples, when necessary, with a Motod a splitter for analysis of approximately 1000 individual organisms per sample. Identifications we re carried out to species when possible for copepods and to major group for the other taxa with a dissecting microscope. The cyanobacteria
5 Trichodesmium was noted if present but was not enumerated, as it tends to be difficult to wash off the net mesh making quantitative analysis difficult Zooplankton were measured to total length by method s described in Hopkins (1981). Equivalent spherical diameter (ESD) was calculated for individuals of each taxon with Optimas (Media Cybernetics, version 6.5) image analysis sof tware and a video camera connected to the microscope for comparison with OPC and SIPPER size measurements. Optimas determines ESD by measuring the area of an object and then calcula ting the diameter of a sphere with the same area. Regressions for calculating ESD from total l engths for each taxon were then calculated and applied to each net sample. Sample biovolume was ca lculated for each sensor according to the equation SBV= k = n i iESD1 3) ( 6p, where k is the sub-sample ratio, n the number of in dividuals, and ESDi the ESD of the ith individual (Labat et al., 2002). Biomass for net an d SIPPER samples was calculated from regressions determined for zooplankton from the Gul f of Mexico by our laboratory (Table 1) by measuring 50 individuals from each selected taxon. SIPPER data analysis The SIPPER is a continuously imaging zooplankton se nsor that records two dimensional, high resolution images of zooplankton and other sus pended particles prior to sampling by the HRS plankton nets (for a full description of SIPPER see Samson et al., 2001). The SIPPER works by projecting a collimated laser light sheet perpendicular to seawater flow through the sampling tube of the HRS and continuously imaging t he outlines and shadows of particles as they pass through the sheet onto two line-scan camera sy stems mounted orthogonal to each other. In this manner, pairs of digital images are created fo r each particle passing through the sensor. These cameras are capable of recording images at up to 8-bit (256) color grayscale, but we used a real time thresholding step to reduce the recorde d data to single bit or black and white images. This dramatically reduced the instantaneous data ra te of the system, increased data storage
6 capability, and made particle detection easier as p articles were comprised mostly of black foreground pixels against a white background. Because zooplankton will be randomly oriented as th ey pass through SIPPER, the use of two orthogonally mounted cameras significantly incr eases the possibility that at least one camera will capture an image of a zooplankter in a recogni zable orientation. Line-scan cameras build an image one line at a time, and because the particle flow through the SIPPER/net sampling tube is unidirectional, each particle can only be imaged on ce. Figure 2 demonstrates the concept of a line-scan camera system using a single camera for e asier conceptualization. Figure 2.Diagram illustrating the general concept o f the SIPPER line-scan camera system. A cumulative image of the sample volume is built of m any individual single scan lines. In this figure, multiple scan lines are included in the Â“single sca n lineÂ” frames in order to minimize the number of single scan line frames illustrated.
7 Table 1. Dry weight biomass regressions for zooplan kton from oceanic waters of the Gulf of Mexico (N=5 0 for each taxon). Group Dry weight regressions (mg DW) Correlation (r2) Notes Copepod DW=.0085(ML)3.1007 0.988 ML=metasomal length Combination of 38 copepod taxa in the NE Gulf of Mexico Chaetognath DW = 0.0002(TL)3.1612 0.971 TL=total length Cnidaria DW = 0.0029(ESD)2.28 0.868 Decapod and Euphausiid DW = 0.001(TL)3.1331 0.977 Doliolid and Salp DW = 0.0108(TL)1.6307 0.986 Larvacean DW = 0.0164(HL)2.0922 0.995 HL=head length Meroplankton DW= 0.0041(ESD)2.31 0.881 Mostly echinoderm bipinnaria and plutei Mollusc DW=.0296(ESD)1.5646 0.878 Mostly pteropods and some heteropods Other crustaceans DW=0.0505(TL)1.8223 0.982 Mostly ostracods and some amphipods Polychaete DW=0.0091(TL)1.801 0.977 Combination of unid. polychaetes and Tomopteris sp. Siphonophore DW= 0.0088(TL)0.0414 0.981
8 Image resolution is determined by the line scan cam era pixel array size in one dimension (in this case 9.6 cm divided by 2048 pixels in the camera array allows for 47 m m resolution) and the flow speed through the sample tube divided by t he line scan rate of the camera system in the other (in this case ~0.75 m s-1 divided by 15,000 line scans a second allows for a n average pixel dimension of 50 m m). Therefore, apparent pixel dimensions used for S IPPER imaging were almost equal, measuring 47 by 50 m m. This was confirmed by comparing the size of larg e unique organisms from the SIPPER dataset with the actual o rganism measurements from the concurrent net sample. The near uniform apparent pixel dimens ions also ensured that the organisms were being imaged without significant distortion. Because the SIPPER is a high-resolution continuousl y imaging sensor, a significant amount of data are generated and recorded every sec ond. Black and white image data were recorded at approximately 8 megabytes s-1, with each raw SIPPPER file averaging 4.8 gigabyte s total. Fortunately, SIPPER data are perfectly suite d for run-length encoding compression algorithms whereby long runs of identical binary da ta (such as the white background of particle free water) can be represented by much shorter bina ry descriptions resulting in significant data storage savings (up to 277). When decompressed, ea ch SIPPER file can be thought of as a Â“strip chartÂ” the length of each sampling run (~0.5 km) and a width of 9.6 cm with images of every organism and particle that passed through the sampl e tube recorded in the approximate spatial distribution that existed in situ but expressed in two dimensions. SIPPER imaged par ticles are colored black while the background is colored white This makes the particle detection and extraction phase of SIPPER processing very simple. Custom region-of-interest (ROI) extraction software was developed with Lab Windows/CVI (National Instruments) and used to dete ct, extract and create bitmap format images of zooplankton and other particles. The routine fir st divided the SIPPER data into 2048 by 2048 pixel Â“framesÂ”, each of which is equivalent to appr oximately 1/7th of a second of sampling or approximately 10 cm (Fig. 3) of travel and computat ionally not too large to process. ROIs were then located in each frame by finding foreground pi xels (black) and extracting those groups of
9 black pixels that contiguously were comprised of a user defined number of black pixels or greater. For this study, ROIs were extracted that were large r than 250 m m ESD. A preprocessing digital dilation step was used to connect black pixels that were within 3 pixels of other black pixels to ensure that organisms and particles with non-contig uous boundaries due to imperfect thresholding or illumination were included in the R OI extraction and counted as a single particle image. If a ROI spanned more than a single frame, t he next frame would be added to the current frame so that the contiguous particle image could b e extracted. These steps ensured that almost all particles greater than 250 m m ESD were extracted from the SIPPER file. Extracted particle images were then viewed with a t humbnail browsing program (Thumbs Plus, Cerious Software). I manually sorted the imag es into 13 recognizable plankton groups (Fig. 4) and one unidentified particle class. Recognizabl e images of marine snow were included in the unidentified class as it was difficult to identify marine snow once it approached 1 mm in size. Trichodesmium colonies were included as a plankton class because of their high abundance in the SIPPER dataset and large individual size. Becau se of the high diversity of the zooplankton assemblage in the deepwater Gulf of Mexico (e.g. Or tner et al., 1989 identified 133 separate zooplankton species) I did not attempt to use SIPPE R to identify organisms to species even though some species with characteristic features we re easily recognized. Images from each class were then analyzed with Optimas image processing so ftware that collected size and morphological information from each ROI including E SD. The location of each ROI within the SIPPER sampling transect was recorded on a separate data file. Locations were then checked against each other to ensure that no particle was c ounted more than once. SIPPER data were not sub-sampled. Because of the large number of images that had to be manually classified, only images from one of the two orthogonal views were us ed for this study. OPC data analysis A detailed explanation of the design and operation of the OPC is found in Herman (1992). Basically the OPC measures the amount of light bloc ked by the area of a particle as it passes through a collimated light sheet between the transm itter and receiver. The blocked light signal is
10 digitized and converted into a size measurement in the form of an equivalent spherical diameter. The OPC is capable of resolving particles 250 m m ESD and greater in size (Herman 1988,1992), but is vulnerable to coincident counting at high pa rticle concentrations (undercounting multiple particles that pass through collimated light sheet at the same time) and has difficulty accurately describing the size of translucent organisms (Zhang et al. 2000, Grant et al. 2000). OPC determined particle size data were binned into 100 m m ESD groups (300 to 5000 m) for comparison with the net and SIPPER zooplankton data Figure 3. An example 2048 2048 pixel frame illust rating the large sampling area of SIPPER and its image quality. The dashed line represents the d imensions of the Â“pseudovolumeÂ” used to estimate OPC coincidence (4.6 9.6 0.4 cm). Sca le bar is equivalent to 1 cm.
11 Results Hydrography Summer conditions in the northeast Gulf of Mexico a re usually stable with the exception of quasi-annual intrusions of the Loop Current and associated eddies (Maul and Vukovich, 1993; Mller Karger, 2000) and even rarer incursions of l ow salinity surface plumes from the Mississippi River outflow (Mller Karger et al., 1991; Mller K arger, 2000) that might influence the zooplankton assemblage (Ortner et al., 1995). Tempe rature and salinity profiles collected during this study (Fig. 5) indicated the water being sampl ed as Gulf Common water (GCW, Vidal et al., 1994). This is differentiated from Subtropical Unde rwater (SUW) being transported by the Loop Current from the depth of the 22o C isotherm. In GCW the 22o C isotherm is found between 50 and 100 m whereas in the Loop Current it is found b elow 150 m (Austin and Jones, 1974). The seasonal thermocline was located between 25 and 30 m depth. Salinity profiles (Fig 5) and satellite data indicated no influence from the Miss issippi River. The deep chlorophyll maximum (DCM) was at approximately 65 m near the salinity m aximum, and there was very low chlorophyll biomass throughout the rest of the epipelagic zone based on both in situ fluorescence and extracted chlorophyll (Fig. 5). Mesozooplankton and Mesozooplankton-Sized Particle Abundance The vertical distribution pattern of mesozooplankto n and mesozooplankton-sized particles was described similarly by all three samp ling methods (Fig. 6). There was a peak in abundance at 10 m and a secondary maximum at 40 m. The major difference was in the total number of particles sampled by each method. The SIP PER recorded the highest numbers of mesozooplankton-sized particles at all depths sampl ed when compared against the results of the nets and the OPC. The SIPPER data were separated in to two abundance estimates: (1) total number of extracted ROI images with a greater than 250 m ESD Â“particleÂ” (SIPPER total) and (2) those ROIs that could be identified as planktic organisms (SIPPER i.d.). A total of 174,699 SIPPER ROIs were extracted and manually examined fr om the 100 minutes of SIPPER data.
12 Most of these images contained unrecognizable parti cles and only 28% of the total (48,931 plankton images) could be classified into one of th e 13 plankton groups. The proportion of SIPPER i.d. to SIPPER total ranged from 41% at 10 m eters to 17% at 100 meters (Table 2). Plankton net estimates of mesozooplankton abundance were the lowest at each depth sampled relative to the other sampling methods. Net counts on average equaled only 13% of the SIPPER total, 24 % of the OPC total and 49% of SIPPER iden tified plankton abundance. The OPC consistently sampled approximately one-half the num ber of particles that the SIPPER imaged at all depths. Figure 4. Representative SIPPER images of the thirt een enumerated plankton groups. Groups AF are from left to right, with group code in parent heses: A. other crustaceans (Crus), B. copepods (Cope), C. larvaceans (Larv), D. Trichodesmium sp. (Tric), E. protoctists (Prot), F. echinoderm larvae (Echi). Scale bar for these groups is equiva lent to 2.5 mm. Groups G-M are, from left to right: G. chaetognaths (Chae), H. cnidarians and ct enophores (Cnid), I. euphausiids and decapods (Euph), J. Polychaetes (Poly), K. Molluscs (Moll), L. other tunicates (Tuni) and M. siphonophores (Siph). Scale bar for these groups is equivalent to 5 mm.
13 Figure 5. Temperature, salinity, fluorescence and e xtracted chlorophyll (denoted by *) profiles (0200 m) from the study site
14 Figure 6. Numerical abundance estimates of the thre e zooplankton sampling methods. All abundances were normalized to volume filtered.
15 Table 2. Abundance (number m-3) of mesozooplankton sized particles as estimated b y the 162 m m net system, SIPPER and the OPC. SIPPER counts are split into total ROIs with a greater tha n 250 m m ESD particle within it, and those images that cou ld be classified into one of the 13 plankton groups. Performance of the net system, SI PPER identified plankton and the OPC are all compar ed against the SIPPER total. Depth (m) Net counts SIPPER total* SIPPER identified images OPC counts Net counts/ SIPPER total SIPPER identified/ SIPPER total OPC counts/ SIPPER total 10 1537 7397 3508 4335 21% 47% 59% 20 1117 6634 1591 3320 17% 24% 50% 30 374 5849 1404 3497 6% 24% 60% 40 756 5855 1487 3529 13% 25% 60% 50 431 5428 1106 2983 8% 20% 55% 60 664 4811 1372 2640 14% 29% 55% 70 371 3864 742 2139 10% 19% 55% 80 279 2428 727 1365 11% 30% 56% 90 359 1877 591 942 19% 31% 50% 100 171 1657 276 831 10% 17% 50%
16 OPC Abundance Correction and Coincidence Frequency Assuming particles were randomly distributed within the water column; Sprules et al. (1992) derived a formula describing the probability of two or more particles occurring within the sampling beam of the OPC at the same time given a k nown particle concentration. They determined that coincidence would be a significant source of error for all but the lowest zooplankton densities. To determine if coincidence was responsible for the low OPC counts relative to the SIPPER total, I modified the formul a of Woodd-Walker et al. (2000) using SIPPER total counts normalized to volume sampled as the kn own concentration of OPC detectable particles in the OPC light beam. The average number of particles in the OPC beam ( m ) is determined by: m =CV, where C was the concentration of particles greater than 250 m m in the SIPPER total for each depth and V was the volume of the OPC beam (220 mm x 20 mm x 4 mm or 17.6 ml). The average number of particles recorded by the OPC (av no.) is calculated by the equation: OPC av. no. = 1-em (Woodd-Walker et al., 2000). The coincidence factor can then be calculated by dividing the average number of particles in the OPC beam by the average number of particles recorded by the OPC (coincidence factor = m / OPC av. no.). For this study, the coincidence fac tor ranged from 1.01 to 1.06 indicating that coincidence should hav e been a rare occurrence within the OPC if the particles were randomly distributed at the concentr ations sampled by SIPPER (1-8 particles l-1). To investigate further, I created a Â“pseudovolumeÂ” within the SIPPER image dataset equivalent to the volume sampled by the OPC at any one instant. Because the sampling area of the SIPPER is approximately twice that of the OPC a nd of a different geometry, I used a subsample of the SIPPER imaging window that was 4.6 cm 9.6 cm 0.4 cm to create a sampling volume of 17.6 ml, equivalent to that of the OPC. I calculated the distance between each particle from its neighbors within the Â“pseudovolumeÂ” to det ermine which particles would be affected by coincident counting (this can be visualized in Fig. 3, where a copepod and two Trichodesmium
17 colonies occupy the dotted box representing the Â“ps eudovolumeÂ” dimensions), if the SIPPER were to sense particles like the OPC. On average, 2 9% of SIPPER imaged particles of OPC detectable size had a neighbor closer than 4 mm and therefore would not be individually counted by the OPC (Table 3). The large number of close-tog ether particles suggests that their distribution was not random. By correcting for the estimated coi ncidence frequency, OPC abundance estimates were recalculated and corresponded more c losely with the SIPPER totals. However, corrected OPC counts still only accounted for 61-78 % of the SIPPER imaged particles. Size Frequency Distributions The cumulative size-frequency distributions of the three sampling methods demonstrated large differences in sampling performance (Fig. 7). While each sampling method was able to discern the same exponential decrease in abundance with increasing size, the SIPPER was able to detect far more particles than either the net or OPC for a given size class. Much of the discrepancy between the SIPPER total and SIPPER i.d abundances could be attributed to the large number of less than 0.5 mm ESD particles that could not be identified. This was most likely due both to the large numbers of small-suspended pa rticulates in the water column and also the minimum size resolution of identifiable objects in the SIPPER dataset. While the SIPPER can detect and image very small particles, the capabili ty to identify them is made difficult by the 50 m m pixel size. Smaller plankton will be imaged, but no t with enough pixel definition to determine their identity. As plankton images grow larger, there ar e more pixels available to define their shape and aid recognition. This concept is illustrated in a graph plotting the proportion of identifiable plankton images versus the SIPPER total (Fig. 8). This proportion rose steadily with particle size, such that at ESDs over 2 mm, over 80% of the SIPPER ROIs were of identifiable plankton. The inability of the OPC and nets to detect and sam ple the smallest size classes in the same magnitude as the SIPPER was most likely due to the inefficiency of the net in sampling the smallest zooplankton and suspended particulates due to extrusion through the net mesh (Gallienne et al., 2001; Hopcroft et al., 2001) and to approaching the 250 m m ESD detection limit for the OPC. Both the net and OPC additionally disp layed a systematic abundance difference of
18 up to an order of magnitude less at each size class compared to the SIPPER datasets, indicating that these differences were not size dependent. Figure 7. Cumulative size-frequency distribution spectra for the three sampling sensors (300-5000 m m ESD).
19 Table 3. OPC normalized particle abundance, theoretical coin cidence factor, SIPPER estimated coincidence percen tage, count loss and corrected OPC particle abundance. Depth (m) OPC counts m-3 Coincidence factor SIPPER-estimated coincidence percentage Counts lost to coincidence Corrected OPC counts m-3 10 4335 1.06 33.6% 1458 5793 20 3320 1.05 38.3% 1270 4591 30 3497 1.05 33.0% 1154 4651 40 3529 1.05 30.1% 1063 4591 50 2983 1.04 27.1% 808 3791 60 2640 1.04 23.6% 624 3264 70 2139 1.03 24.7% 539 2668 80 1365 1.02 21.9% 299 1664 90 942 1.02 22.2% 209 1152 100 831 1.01 23.3% 194 1025
20 Figure 8. Proportion of SIPPER identified plankton to the SIPPER image total per 100 m m ESD size class. Sample Biovolume Estimates Differences in both the abundance and size distribu tion of the three datasets led to large differences in the sample biovolumes (SBV) estimate d by each sensor (Fig. 9). Because the main discrepancy in the SIPPER i.d. to SIPPER total abun dance estimates was from the smallest size classes, the biovolume difference between the two w as much less than the abundance difference due to the minimal contribution small particles or organisms make to the biovolume total. Thus, while the identified plankton of the SIPPER i.d. da taset made up only 28% of the SIPPER total particle abundance, they made up 79% of the SIPPER total biovolume. Net and OPC SBV were only 11% and 23 % of the SIPPER total respectively and 13% and 29% of the SIPPER i.d. biovolume.
21 Figure 9. Cumulative sample biovolume (mm3 10m-3) versus size (300-5000 m m ESD) distribution. A logarithmic scale is used and the values transfor med (n+ 0.1). The Problem of Trichodesmium Ideally, the taxonomic composition of the net and SIPPER i.d. datasets should be identical. However, this was not the case, as the a dvantages and disadvantages of each system in sampling different components of the plankton we re manifested in significantly different descriptions of the assemblage. Firstly, the coloni al cyanobacteria Trichodesmium sp. formed the most abundant plankton class in the SIPPER dataset, especially at a depth of 10 m where it was found at concentrations greater than 1800 colonies m-3, but were not quantified in the net samples. Trichodesmium is a filamentous phytoplankton that is difficult t o enumerate in zooplankton net samples because of its fragility an d tendency to stick to the net mesh. The opacity and large size (0.5-4 mm ESD) of Trichodesmium colonies made them readily detectable
22 by SIPPER and, most likely, the OPC. Those colonies that might have been fragmented or disrupted in the SIPPER sampling tube, and single t richomes (which image as long strands), were not counted as Trichodesmium in the SIPPER dataset. Removing the Trichodesmium images from the SIPPER i.d. dataset yielded zooplan kton counts ~50% higher than that from the nets (Fig. 10). Figure 10. Zooplankton abundance (numbers m-3) profile determined from the net and SIPPER after Trichodesmium was separated from the SIPPER i.d. dataset. Taxonomic Composition Â– SIPPER vs. Nets Differences between the taxonomic composition of th e SIPPER and net samples varied even more considerably than the differences in abun dance. Data comparing zooplankton composition, abundance and biomass from the nets an d SIPPER are presented in Tables 4 and 5 respectively. Copepods dominated the net samples, c ontributing 63.7 % of the abundance and 36.4 % of the biomass. Larvaceans (17.4 %) were the only other significant contributor to the net
23 abundance total, with the other tunicate class, com prising doliolids and salps, and the protoctista class, comprising mainly acantharians, radiolarians and tintinnids, contributing between 4 and 5 % each. No other zooplankton class contributed more than 2.5 % to the net-sample abundance total. Because of their large individual size, euph ausiids and decapods were the second largest biomass component in the net samples, contributing slightly less (31.8 %) than the copepods. Other crustaceans (comprising amphipods, cladoceran s and ostracods), the other tunicates class, and siphonophores all contributed between 6 and 9 % to the total net collected biomass, mostly based on their larger individual size. No ot her zooplankton group contributed more than 3.5 % to the total biomass in the nets. Six zooplankton groups were found to be significant ly more abundant in the SIPPER i.d. dataset than the concurrent net estimates (Fig. 11) according to paired t-tests (Zar, 1984). These taxa can be broadly grouped as fragile and or gelat inous zooplankton that are easily damaged or disrupted on encountering a net. These included 4 o ut of 5 of the most important numerical contributors to the SIPPER i.d. assemblage. Larvace ans were the numerical dominant, contributing 35.6% to the total and were more than 3x more abundant than the net total. Analysis of net sample collected larvaceans indicated the ma jority were predominately oikopleuridae (mostly Oikopleura dioica) with the fritillariidae also present in noticeable numbers. Three other fragile plankton groups (protoctista, other tunicat es and cnidarians/ctenophores) were important numerically, each contributing between 5 and 14% to total zooplankton abundance. Copepods were the only important non-fragile zooplankton gro up and were the second highest contributor (27.7 %) to the SIPPER i.d. abundance total. The di fferences in abundance between SIPPER and the nets for the fragile and gelatinous zooplankton ranged from just over 200% for the siphonophores to over 1400% for the cnidarians and ctenophores. However, the difference in siphonophore abundance was probably much higher, as individual bracts or nectophores found in net samples were counted as individuals whereas SIP PER imaged and counted whole organisms. This was also the case for polychaetes, which often were found broken up into segmented parts in the nets. The six other zooplan kton groups, comprising mainly more robust
24 taxa such as crustaceans, were sampled similarly by both the net and SIPPER and showed no appreciable difference in abundance (Fig. 12). Doliolids and salps, which made up the other tunica te class, were the biomass dominant in the SIPPER dataset, contributing 37 % to the tot al. Examination of the SIPPER imagery and concurrent net samples revealed that this group was dominated by Dolioletta gegenbauri or a similar congener that made up more than 90% of the total. Interestingly, the large biomass difference (12.2) between the SIPPER and net for t his class was due not just to significant loss of individuals through the net mesh from extrusion or disruption (3.8 more of this class were found in the SIPPER imagery), but also to loss of r eproductive tissue. In the upper 60 meters, a large proportion (33% mean, range: 22-39 %) of the doliolids were of the asexually reproducing oozoid stage (Fig. 4L), which have a lengthened dor sal process or appendix that bears budding blastozooids that form the next stage in the doliol id life cycle. Very few doliolids observed in the net samples bore an intact dorsal process. Copepods (18%) and the euphausiids and decapod class (14 %) each contributed over 10 % to the SIPP ER biomass total, while siphonophores (9%), larvaceans (7%) and other crustaceans (6%) contribu ted more than 5%.
25 Table 4. Net estimated abundance (Number m-3) and dry weight biomass (mg m-3 DW) of zooplankton groups. Plankton group names abb reviated to four letter codes as in figure 3. Depth (m) Chae. Cnid. Cope. Euph. Tuni. Larv. Echi. Moll. Cru s Poly.* Prot. Siph.* 10 19 (1.23) 4 (-) 1039 (13.15) 9 (6.73) 31 (3.39) 280 (0.99) 8 (0.01) 42 (0.62) 16 (1.70) 5 (0.01) 66 (NA) 17 (4.45) 20 15 (0.66) 5 (0.02) 775 (6.12) 3 (1.32) 15 (1.54) 161 (0.68) 4 (-) 63 (0.31) 13 (0.56) 2 (0.58) 51 (NA) 9 (0.81) 30 15 (0.7) 3 (0.02) 206 (2.65) 9 (12.12) 14 (0.43) 74 (0.31) 5 (-) 11 (0.14) 17 (1.92) 3 (0.01) 15 (NA) 5 (0.22) 40 22 (1.21) 3 (0.01) 462 (7.48) 8 (5.45) 24 (0.29) 159 (0.54) 3 (0.01) 6 (0.06) 21 (1.43) 1 (0.01) 37 (NA) 11 (0.73) 50 6 (0.94) 4 (0.02) 246 (5.36) 7 (1.19) 30 (0.37) 82 (0.29) 2 (-) 9 (0.18) 10 (0.49) 9 (0.28) 17 (NA) 10 (1.38) 60 8 (0.08) 4 (0.01) 289 (8.23) 3 (12.52) 133 (2.10) 162 (0.62) 3 (-) 14 (0.11) 18 (2.17) 2 (0.02) 21 (NA) 8 (0.63) 70 3 (0.17) 4 (0.01) 237 (2.99) 4 (0.53) 13 (0.52) 64 (0.21) 3 (-) 8 (0.04) 11 (0.93) 3 (0.01) 21 (NA) 3 (1.99) 80 4 (0.07) 3 (0.01) 199 (3.23) 2 (0.54) 4 (0.22) 41 (0.16) 2 (-) 2 (0.08) 9 (0.65) 4 (0.02) 8 (NA) 1 (0.02) 90 4 (0.10) 4 (0.01) 293 (2.88) 6 (6.00) 6 (0.11) 14 (0.05) 3 (-) 1 (0.05) 16 (2.41) 3 (0.01) 8 (NA) 2 (0.47) 100 2 (0.06) 2 (-) 121 (1.19) 3 (5.19) 1 (0.03) 16 (0.06) 8 (-) (-) 13 (1.24) 1 (-) 4 (NA) 1 (0.24) Total 97 (5.21) 35 (1.24) 3866 (53.29) 54 (51.61) 271 (9.00) 1052 (3.92) 42 (0.04) 155 (1.59) 143 (13.49) 33 (0.94) 247 (NA) 65 (10.93) indicates that body segments found in net counted as individuals indicates less than 0.01 mg m-3 NA indicates biomass not determined for that group
26 Table 5. SIPPER estimated abundance (Number m-3) and dry weight biomass (mg m-3 DW) of zooplankton groups. Plankton group names abbreviated to four letter codes as in figure 3. Depth (m) Chae. Cnid. Cope. Euph. Tuni. Larv. Echi. Moll. Cru s. Poly. Prot. Siph. 10 22 (1.78) 117 (2.66) 537 (12.79) 11 (7.76) 81 (24.12) 608 (4.06) 16 (0.03) 16 (0.37) 16 (1.54) 5 (1.57) 257 (NA) 12 (3.7) 20 22 (1.75) 59 (2.11) 446 (7.58) 3 (2.77) 51 (11.89) 407 (3.37) 12 (0.02) 10 (0.11) 10 (1.24) 5 (0.97) 140 (NA) 23 (3.02) 30 27 (1.15) 73 (3.06) 258 (4.25) 13 (8.93) 122 (18.42) 466 (3.66) 8 (0.01) 9 (0.17) 15 (1.82) 4 (1.93) 145 (NA) 18 (3.93) 40 31 (0.89) 53 (0.75) 303 (7.90) 10 (5.27) 203 (18.95) 398 (2.83) 11 (0.02) 12 (0.09) 27 (3.10) 1 (0.17) 180 (NA) 7 (1.45) 50 8 (0.37) 50 (0.87) 284 (8.02) 4 (1.05) 180 (16.88) 265 (1.79) 10 (0.02) 13 (0.17) 14 (1.90) 5 (2.91) 137 (NA) 5 (0.63) 60 10 (0.37) 79 (1.01) 239 (8.61) 3 (8.39) 288 (19.20) 436 (2.77) 6 (0.02) 9 (0.12) 19 (2.50) 2 (1.14) 94 (NA) 10 (1.99) 70 9 (0.27) 32 (0.39) 195 (8.12) 3 (0.24) 61 (6.63) 173 (1.15) 3 (-) 5 (0.06) 9 (1.38) 1 (0.05) 112 (NA) 7 (1.05) 80 7 (0.43) 17 (1.06) 125 (3.65) 4 (1.53) 26 (1.99) 231 (1.36) 4 (-) 2 (0.25) 11 (2.16) 1 (0.43) 87 (NA) 8 (1.94) 90 7 (0.47) 17 (0.59) 88 (2.79) 3 (4.39) 12 (0.70) 143 (0.61) 6 (-) 4 (0.12) 11 (1.49) 1 (0.01) 83 (NA) 23 (6.48) 100 3 (0.10) 8 (0.35) 57 (1.32) 4 (6.31) 5 (0.17) 71 (0.36) 54 (0.04) 1 (0.01) 4 (1.78) 1 (0.21) 54 (NA) 22 (5.69) Total 145 (7.59) 505 (12.85) 2532 (60.25) 56 (46.63) 1029 (119.07) 3199 (21.95) 130 (0.18) 80 (1.47) 135 (18.91) 23 (9.38) 1289 (NA) 134 (29.87) indicates less than 0.01 mg m-3 NA indicates biomass not determined for that group
27 Figure 11. SIPPER (solid line) and net (dotted line ) numerical abundance estimates of zooplankton groups with significant differences bet ween the two sampling systems (paired t-test, p <0.05).
28 Figure 12. SIPPER (solid line) and net (dotted line ) numerical abundance estimates of zooplankton groups with no significant differences between the two sampling systems (paired ttest, p <0.05). Vertical distribution patterns of most specific zoo plankton groups sampled by both the nets and SIPPER were similar, even when the abundan ce estimates were quite different. For example, while doliolids and salps were significant ly under-sampled by the nets compared to SIPPER, both instruments sampled an abundance maxim um at the DCM much higher than at any other depth. However, for some zooplankton grou ps, SIPPER proved far more capable in describing both abundance and vertical distribution patterns than the nets. For example, cnidarians and ctenophores were extremely abundant in the SIPPER dataset and demonstrated a strong bimodal distribution with maxima at 10 and 6 0 meters, but were virtually absent within the net samples. Additionally, most cnidarians and cten ophores collected in nets were unidentifiable, especially after fixation. In contrast, within the SIPPER dataset, many individual cnidarian and ctenophore taxa could be enumerated to genus or eve n species. For example, we were able to
29 determine that the narcomedusae Solmundella bitentaculata was the most abundant identifiable cnidarian in the depth ranges sampled. Total Biomass Total biomass (0-100 m) determined from SIPPER data was more than twice that determined from the net data (3417 mg m-2 DW vs. 1592 mg m-2 DW), but the vertical biomass distribution pattern was similar. Most of the biom ass difference was explained by the under representation of the fragile taxa in the net sampl es. The fragile and gelatinous zooplankton groups enumerated from the SIPPER dataset contribut ed greater biomass (~1937 mg m-2 DW) than the entire zooplankton assemblage sampled by t he nets. Biomass did not include the protoctista class, which consisted of organisms wit h mineral skeletons or tests that may have biased the results. Because the protoctista class was more than 4 more abundant in the SIPPER dataset than the nets, the true biomass diff erence was probably even greater. Taxonomic Differences in Size Distribution Generally, the three sampling systems showed very l ittle correspondence between particle abundance of a given size class (Fig. 12, left graphs), but some trends were apparent. Even though the absolute totals were very different between 60 and 70% of the total net, OPC and SIPPER unidentified particle abundance consiste d of particles between 250 and 500 m ESD and between 91and 96% of the total were made up of particles less than 1 mm ESD. In contrast, only 73% of the SIPPER identified plankto n were less than 1 mm ESD and the smallest size class contributed less than 26% of the total. The number of SIPPER identified plankton increased with size relative to the other three dat asets. For example, SIPPER i.d. abundance at the largest size class (>2500 m m ESD) outnumbered nets, OPC and the SIPPER unident ified datasets by 6.8, 7.5 and 15.5 respectively. Small zooplankton made up the majority of zooplankt on sampled by both SIPPER and the nets, although the relative importance of large r forms was much greater in the SIPPER i.d. dataset (Figs. 13 and 14, right graphs). Whereas on ly 8.4 % of the net collected zooplankton
30 were larger than 1 mm ESD, more than 25 % of the SI PPER imaged zooplankton were larger than this size. This difference was especially pron ounced in the fragile and gelatinous zooplankton groups that were under sampled by the n ets, such as the other tunicates class, where more than 74% of the SIPPER imaged organisms were larger than this size class compared to only 21% for those collected by the net s. However, this trend was also found for some of the zooplankton groups that showed no sampl ing bias in their abundance estimates. More than 90% of the net collected polychaetes were smaller than 0.5 mm ESD, while more than 80% of the SIPPER imaged polychaetes were larger th an 1 mm ESD. Similarly for planktic molluscs, more than 95% of the net collected indivi duals were less than 1 mm ESD compared to only 47% for SIPPER. Figure 13. Cumulative abundance of mesozooplankton sized particles or mesozooplankton determined by the three sampling methods separated into 500 m m size classes (left graphs) up to 1500 m m ESD. SIPPER data were separated into identified p lankton and unidentified particles. Numerical abundance of the thirteen plankton classe s was calculated for each size class for the net samples and the SIPPER i.d. dataset (right grap hs)
31 Figure 14. Cumulative abundance of mesozooplankton sized particles or mesozooplankton determined by the three sampling methods separated into 500 m m size classes (left graphs) for particles larger than 1500 m m ESD. SIPPER data were separated into identified p lankton and unidentified particles. Numerical abundance of the thirteen plankton classes was calculated for each size class for the net samples and the SIPPER i.d. dataset (right graphs)
32 Discussion The disparate results between the three sampling me thods is at first confusing given that the nets and the SIPPER sampled the exact same wate r volume and the OPC was sampling less than a meter away. This likely precludes the possib ility that micro-scale patchiness affected these differences. The SIPPER provided a picture of a zoo plankton assemblage both more numerous and diverse than either of the other methods. While the majority (~67%) of SIPPER extracted images could not be identified, those that were sti ll significantly outnumbered organisms collected by the nets at most depths. Occurrence of Unidentified Particles in Plankton Im age Datasets The number of unidentified particles in our dataset is comparable to that of other investigators using in-situ imaging sensors. For ex ample, Ashjian et al. (2001) were not able to classify 43% of VPR images collected during three c ruises to Georges Bank. Additionally, they included marine snow as a class that comprised over 71% of their classified images. In contrast, I did not separate marine snow from our unclassified group. While many of these particle images were of identifiable marine snow such as cast-off l arvacean houses, diatom mats and fecal pellet strings, the majority of the unclassified images we re of particles less than 1mm ESD that lacked any resolvable characteristics to aid in identifica tion. This was partly a problem of the coarse imaging resolution of the SIPPER (50 m m square) relative to the small size of the particl es. A small copepodite or copepod nauplius measuring 400 m m TL would be imaged by the SIPPER but would be comprised of such a few number of pixe ls that identifying it as such would be impossible with the present SIPPER configuration. H opkins (1981), studying zooplankton at the same station as this study, found that metazoan pla nkton under 1 mm total length sampled from bottle casts outnumbered metazoan plankton > 1 mm c aught in a 162 m m plankton net by 35x and were made up primarily of copepod nauplii and c opepodites. Thus, it is possible that a large percentage of the small-unidentified particles in t he SIPPER dataset were of small zooplankton such as copepod early life stages. While I did not use images from the second camera of SIPPER for this study (which imaged orthogonal to the firs t), it may have proven useful in identifying some
33 of these smaller particles by providing a second pe rspective that could present recognizable features. Importance of Trichodesmium in Subtropical Systems A substantial number of the larger classified image s extracted from the SIPPER dataset were of Trichodesmium colonies, especially at a depth of 10 meters where they outnumbered both the SIPPER and net zooplankton abundance estim ates. Trichodesmium is an important component of tropical and subtropical oceanic ecosy stems, contributing a significant amount of new nitrogen to otherwise impoverished waters (Capo ne et al., 1997; Karl et al., 1997). Furthermore, Trichodesmium has been implicated in contributing to the initiat ion of harmful algal blooms of the dinoflagellate Karenia brevis (Lenes et al. 2001) in the Gulf of Mexico. Typical ly, the abundance and vertical distribution of Trichodesmium are determined with water bottles or drift nets, which are limited in their ability to d etect it at low concentrations, can damage or disto rt specimens, and are prone to sampling error due to t he small volumes sampled (Chang, 2000). Because SIPPER samples a larger volume of water, it can detect Trichodesmium at much lower concentrations than these traditional methods. Marine Snow and Large Phytoplankton as Signal Rathe r than Noise Other investigators using optical methods to invest igate zooplankton distributions have also found that large colonial phytoplankton such a s diatoms can dominate the marine particle assemblage within the mesozooplankton size range (N orrbin et al., 1996; Grant et al., 2000) The dominance of marine snow and of Trichodesmium and other phytoplankton in the water column suggests that data collected with non-imaging optic al sensors such as the OPC must be interpreted with caution when converted to zooplank ton abundance and size distributions. Imaging instruments such as the SIPPER, on the othe r hand, can differentiate between these groups and allow for accurate determination of thei r contribution to the total marine particle assemblage. Comparison between Different Sampling Methods There have been a number of studies investigating t he performance of the OPC against plankton net catches (Sameoto et al., 1993; Grant e t al., 2000; Halliday et al., 2001), but few if
34 any comparisons against in-situ imaging systems. Herman (1992) suggests that with strict sample and analysis control, net and OPC counts can agree to within 30%. Most studies, however, appear to have much more trouble reconciling the ou tput of the OPC with what is collected within a net. Using the OPC mounted on the HRS to sample a 80 km transect on the West Florida Shelf, Sutton et al. (2001) found that the OPC grossly und erestimated mesozooplankton abundance, especially at high concentrations (>10,000 organism s m-3), but described the overall pattern of zooplankton fairly well. Herman (1988) and Sprules et al. (1998) have also reported OPC counts less than net counts, while others have reported la rge overestimates by the OPC relative to plankton nets (Grant et al., 2000; Halliday et al., 2001). Suggested causes of OPC underestimates are coincident counting (Sprules et al., 1992; Woodd Walker et al., 2000; Labat et al., 2002) and the presence of highly translucent o rganisms (Wieland et al., 1997; Beaulieu et al., 1999), while overestimates have been attributed to the presence of marine snow and detrital aggregates (Zhang et al., 2000), large phytoplankto n (Grant et al., 2000), small zooplankton that pass through the net mesh (Halliday et al., 2001; Z hou and Tande, 2002), and fragile organisms that are destroyed in the nets (Gallienne and Robin s, 2001). During this study, the OPC consistently sampled app roximately half the number of particles that the SIPPER imaged at all depths. Muc h of this underestimation could have been due to coincidence, as I showed within the sub-samp led SIPPER imaging volume. More than 29% of the particles occurred within 4 mm of each o ther within the SIPPER Â“pseudovolumeÂ” and, therefore, would have been counted as a single part icle if sampled by the OPC. By correcting for this, OPC abundance estimates approached between 60 -80% of the SIPPER imaged particles in the mesozooplankton size range. The occurrence of l arge numbers of highly transparent organisms such as cnidarians, ctenophores, doliolid s and salps that were imaged by SIPPER could be responsible for much of the remaining diff erence in abundance estimates between it and the OPC as earlier studies using the OPC have demon strated that it can miss detecting or underestimate the size of transparent zooplankton ( Labat et al., 2002). Additionally, the difference in the two sensors sampling areas could also contribute to the lack of correspondence between the OPC and SIPPER counts. Baumgartner (200 3) found that late copepodite stages of
35 Calanus finmarchicus could avoid the OPC at speeds similar to this stud y. Because the OPC has a wide but narrow sampling mouth (2 22 cm) and th e SIPPER has a larger square aperture (9.6 9.6 cm), zooplankton may have been able to escape out of the way of the OPC more often than the SIPPER. Implications for Subtropical Oceanic Biology The deepwater Gulf of Mexico biological community h as been sparsely sampled (Biggs and Ressler, 2001), but the general consensus is th at the biology of offshore waters in the Gulf is similar to that of other low-latitude tropical ocea ns, with low biomass and high diversity of zooplankton, ichthyoplankton and micronekton (Hopki ns, 1981; Hopkins et al., 1996; Biggs and Ressler, 2001). The abundance, composition, size an d vertical distribution and biomass of the mesozooplankton sampled during this study by the HR S plankton nets was similar to that found during an earlier study at this station (Hopkins, 1 981) and to other investigations in oceanic waters of the Gulf of Mexico (Cummings 1983; Ortner et al., 1989; Biggs and Ressler, 2001). Combined with an earlier study investigating the di stribution of mesozooplankton on the West Florida Shelf using the HRS (Sutton et al., 2001), these results suggest that the somewhat small plankton nets used on the sampler provide similar r esults to those of other investigators using larger plankton nets to describe the mesozooplankto n assemblage. Comparing the HRS net catches versus what SIPPER im aged in the same water yielded a significantly different picture of the mesozoopla nkton assemblage. While a number of investigators have begun to stress the need to use multiple nets of different mesh sizes to adequately sample the entire mesozooplankton size r ange (Gallienne and Robins, 2001; Hopcroft et al., 2001), our results suggest that nets still might miss a large numerical and biomass fraction. While copepods were both the numerical and biomass dominant in the net samples, larvaceans were the numerical dominant and doliolids and salps (forming the other tunicate class) were the biomass dominant in the SIPPER dataset. Small copep ods, which made up the majority of the net-caught zooplankton, such as the genera Calocalanus Oithona Paracalanus Oncaea and Temora, were difficult to identify in the SIPPER dataset b ecause of their small size even though they were likely imaged, and therefore were underes timated (in this case counted in the SIPPER
36 total). The large number of fragile and gelatinous organisms in the SIPPER dataset and their near absence in the nets obviously has implications on h ow a planktic ecosystem is described. For example, Hopkins et al. (1996) found that midwater shrimps and fish, the two dominant micronekton groups in this region, accounted for on ly 25% of the zooplankton daily production consumed in the eastern Gulf. It remains unresolved which ecosystem components are responsible for most zooplankton predation although they suspected large gelatinous predators. More work in this region with SIPPER might resolve that question. Effects of Formalin Preservation on Sample Size Dis tribution Fixation of zooplankton samples with formalin has b een shown to cause shrinkage of the preserved organisms (Postel et al., 2000). It is po ssible that shrinkage may have contributed to the observed differences between the size-frequency biovolume and biomass of the net samples with that of the SIPPER. For example, Beaulieu et a l. (1999) measured a 41% decrease in biovolume of the scyphozoan medusae Aurelia aurita and Nishikawa and Terazaki (1996) found that doliolids and salp body lengths shrank to appr oximately 86-93% of their live length after preservation. Omori (1978) observed that the size of copepods and other crustaceans were less affected by fixation than gelatinous organisms. Thi s likely explained some of the differences observed in the pteropod and polychaete size distri bution in the net samples compared to SIPPER. However, the large differences in biovolume and biomass of the gelatinous and fragile organisms between SIPPER and the nets were due to i ncreased abundance of these animals at all size classes. Therefore, the biovolume and biom ass differences were due more to a difference in total abundance rather than a shift in the size frequency spectrum. Comparison of SIPPER Performance with Other Imaging Systems Prior investigations comparing net and imaging syst ems have yielded similar results to this study. Parallel deployments of the VPR and the MOCNESS on Georges Bank have shown that the MOCNESS significantly under-samples echino derm larvae, larvaceans and medusae relative to the VPR (Benfield et al., 1996) and the VPR also sampled foraminifera, acantharians and other fragile protoctistan zooplankton more eff ectively than nets (Gallager et al., 1996; Norrbin et al., 1996; Ashjian et al., 2001). In tho se studies, however, copepods and other harder
37 bodied organisms were still the numerical and bioma ss dominant in both the net and VPR samples and thus the under-representation of the fr agile forms appeared to be less important. In the central North Pacific Ocean, Dennett et al. (20 02) reported that colonial radiolarian colonies averaged 380 more abundant in VPR samples than tho se collected in the nets and were an important but overlooked component of biomass in ol igotrophic waters. Similarly, our study in an oligotrophic central oceanic ecosystem indicated th at traditional net sampling might miss more than half the mesozooplankton biomass. Conclusions This study demonstrates the importance of in-situ imaging systems to accurately assess the abundance, size distribution and composition of a low-latitude mesozooplankton assemblage. These systems provide the capability to sample gela tinous and fragile organisms that otherwise may be overlooked, even though they may be importan t contributors to the ecology of an ecosystem. Similarly, the limitations of the OPC an d plankton nets in describing this assemblage were explored. The primary disadvantage of the SIPP ER is the current need to manually classify the large volume of images generated by the sensor. The ability to automatically identify images of zooplankton collected in the lab or field has re ceived considerable attention (Jeffries et al., 1984; Tang et al., 1998; Akiba, 2000; Iwamoto et al ., 2001) and an operable pattern recognition algorithm is in use for the VPR (Tang et al., 1998) In Chapter 2, I detail the first results from of a new grayscale imaging SIPPER and investigate the pe rformance in the field of a plankton identification software package I helped develop fo r the SIPPER. These improvements should allow for greater discrimination between particle g roups, especially marine snow and smaller plankton that were difficult to identify with the b inary imaging SIPPER. With these advances I believe the SIPPER can provide more accurate mesozo oplankton abundance and size measurements than nets and provide valuable insight into processes controlling zooplankton distributions at both the individual and community level.
38 Chapter 2: Describing plankton distribution and abu ndance in neritic subtropical waters using SIPPER-2 and an automated classification system. Introduction A comprehensive knowledge of the distribution and d iversity of zooplankton populations is critical to determine their influence on fisheri es recruitment, phytoplankton production via grazing, contribution to particle flux and their po ssible use as sentinels for global climate change (Lenz, 2000; Hays et al. 2005). This task is made d ifficult by the fact that these populations are distributed heterogeneously over a broad range of t emporal and spatial scales (Haury et al., 1977;Mackas et al., 1985). These heterogeneous patc hes are often the focus of increased production, feeding and reproduction for plankton g roups as resources within these patches are often greater than the ambient surroundings (Mullin and Brooks, 1976; Malkiel et al., 2006). To adequately sample these patchy distributions of zoo plankton requires a capability for intensive and high frequency sampling of the population so th at the entire range of variability can be observed and the processes responsible for the patt erns uncovered (Sutton et al., 2001;Yamakazi et al. 2002). Traditional methods of sampling zooplankton using nets or bottles are limited in this regard as high frequency sampli ng can generate hundreds to thousands of samples that are both costly and time consuming to process and analyze. Furthermore, these methods integrate spatial information so that finescale distribution patterns cannot be observed. These limitations have led to the development of a number of in-situ zooplankton imaging sensors such as the Video Plankton Recorder (VPR, D avis et al., 1992), Underwater Video Profiler (UVP, Gorsky et al., 1992), Shadowed Image Particle Profiling and Evaluation Recorder (SIPPER, Samson et al., 2001) and the Zooplankton V isualization and imaging system (ZOOVIS Benfield et al., 2003) that collect high resolution imagery of plankton and suspended particles. While currently unable to deliver the taxonomic det ail of traditional methods that physically capture zooplankton, these systems can provide cont inuous or near continuous measurements of
39 the distribution of these groups over a wide range of temporal and spatial scales (Wiebe and Benfield, 2003, Benfield et al., 2007). For example the VPR has been used to both investigate the nearest neighbor distances between copepods alo ng a sampling transect (Ashjian et al., 2005) and examine the vertical distribution of Trichodesmium across the Atlantic basin (Davis and McGillicuddy, 2006) Sampling by these systems is far less physically in trusive than traditional zooplankton sampling methods which often destroy or damage the more fragile forms. Whereas abundance estimates from imaging systems are comparable with net and bottle estimates of robust, non fragile zooplankton such as copepods and pteropods carried out in the same area (Benfield et al., 1996, Gallagher and Davis, 2003) or concurrent with the net sampling (Remsen et al., 2004, Broughton and Lough, 2006), abundance estimates of gelatinous or fragile forms such as hydromedusae or larvaceans collected by imaging sys tems are often many orders of magnitude higher than that from nets or bottles (Benfield et al., 1996; Dennet et al., 2002; Remsen et al., 2004) collected in the same area. Additionally, th e distribution and abundance patterns of marine snow aggregates (Ashjian et al., 2001, 2005b) and l arge algal (Sieracki et al., 1998, Pilskaln et al., 2005) and cyanobacterial colonies (Remsen et a l., 2004, Davis and McGillicuddy, 2006; Walsh et al., 2007) can also be determined using im aging systems. Until recently, a major bottleneck in the applicati on of plankton imaging systems to field studies was the necessity of manually identifying t he acquired images. While much of the postprocessing and measurement of imaged particles has been automated, the actual classification of imaged particles had to be done manually (Benfield et al. 1996; Ashjian et al., 2001; Remsen et al., 2004). As field deployments of these systems c an generate hundreds of thousands (Ashjian et al., 2001; Remsen et al., 2004) to millions of i mages (Hu and Davis, 2005) per cruise, manual classification of the collected dataset would be im practical. While methods to automatically classify plankton images have been under developmen t for over 25 years (Jeffries et al., 1984; Rolke and Lenz, 1984), it has only been until recen tly that they have shown promise for field collected images.
40 Early studies of automated plankton classification systems utilized only distinctive and recognizable images of plankton collected in the la b (Jeffries et al., 1984) or the field (Tang et al. 1998) separated into a relatively small number of g roups. While these resulted in accurate classifiers (90-92%), they were not applicable to f ield collected images that can be unrecognizable, projection variant, occluded, out o f focus, and unevenly illuminated (Davis et al., 2004; Luo et al. 2004). When unidentified SIPPER im ages were included in a test set, Luo et al. (2004) found that the accuracy of their multi-class support vector machine (SVM) classifier fell from 90 to 75% for 5 groups of plankton. Davis et.a l (2004) applied their neural network classifier to VPR images collected in the field including an Â“ otherÂ” or unidentified class, and achieved 61 % accuracy while classifying their dataset into 7 gro ups in real time. This was later improved to 72% using a new feature set and switching their classif ier to a SVM (Hu and Davis, 2005). Further improvements were made using a dual classification method where a classification was made only if both classifiers agreed (Hu and Davis, 2006 ). Otherwise an image would be labeled as unknown. Using the ZOOSCAN imaging system to count and identify preserved plankton from net samples, Grosjean et al. (2004) has achieved 75% ac curacy for a 29 group training set using a discriminate forest classification algorithm while classifying at rates of up to 2000 images per minute. Many systems now have a semi-interactive ma nual correction step in which images classified with low probability can be sent to a hu man expert for correct classification without significantly slowing down the classification effor t (Davis et al., 2004; Grosjean et al., 2004). This can improve overall classification accuracy to 80-8 5% (Grosjean et al., 2004) and is comparable to what humans achieve with complicated classificat ion tasks (Culverhouse et al., 2003) but with significantly greater speed. In this chapter I examine the results of a multiple class SVM classifier (Luo et al., 2004, 2005) identifying plankton and particle images coll ected from a new grayscale imaging SIPPER deployed in the subtropical waters of the eastern G ulf of Mexico. The new system is described and the performance of the classifier in describing the composition and spatial distribution of a diverse plankton assemblage is explored.
41 Methods A new grayscale-imaging SIPPER was deployed on the USF high resolution sampler (HRS) to investigate small scale plankton and suspe nded particle spatial distributions during a September 2002 cruise to the West Florida Shelf (WF S) in the Gulf of Mexico aboard the R/V Suncoaster. The HRS is a comprehensive marine part icle analysis platform (Sutton et al., 2001, Remsen et al., 2004) capable of concurrently collec ting electronic environmental sensor data along with discrete net and water bottle samples fo r verification and calibration of the sensors (Fig. 15). Figure 15. Photograph of SIPPER-2 mounted on the HR S. The verification capability of the HRS is comprised of (A) the 97 cm2 sampling tube, (B) the SIPPER imaging system, (C) the 20 position plankton net carousel, and (D) the 10 1.2 liter Niskin bottle array A series of seven deployments of the HRS were made at a station (27.2 N 83.5 W, also known as the Florida-Ecology of Harmful Algal Blooms (ECO HAB) station 10) along the 50 m isobath of
42 the WFS during a 24 hour sampling period spanning S eptember 19-20 2002. Each deployment consisted of a vertical profile of the water column and 10 discrete stops at ~5 m increments from 345 m in which discrete SIPPER and 162 m m plankton net samples were collected for 8 minutes each (fig. 16). At each depth stratum a 1.2 liter N iskin bottle on the HRS would also be opened and closed for later chlorophyll extraction and ana lysis. SIPPER sampling did not take place any shallower than 2-3 m due to concerns with imaging t he ships wake and or encountering shipinduced turbulence. Tow speed of the HRS during dep loyments averaged 1.5 knots (0.77 ms-1). Sampling was meant to continue for several days to investigate the diel distribution of planktic forms, but a failure of our custom designed block f or paying out HRS cable cut short our effort. Figure 16. Depth profiles from the seven deployment s of the HRS collected during this study. The circled areas indicate the depths where SIPPER and net sampling took place. A total of 70 discrete depth SIPPER and 69 plankton net samples (One net sample was lost) were collected for this study. The SIPPER and plankton nets utilized the same 9.6 9.6 cm (92.16 cm2) sampling tube to sample seawater. The nets and SIPP ER sampled the water column for a total of 560 minutes (9.3 hours) and sampled greater than 245 m3 of seawater. SIPPER data
43 files ranged in size from 105 to 688 MB and totaled over 20 GB of image data. We used plankton image classification and extraction software (PICES ), developed in collaboration with the USF College of Engineering (Luo et al. 2004, 2005), to extract and classify all of the images greater than a user specified size from each SIPPER data fi le. For this study, all particle images greater than 255 total pixels in area (0.23 mm2 or 0.55 mm equivalent spherical diameter (ESD)) wer e extracted from each SIPPER file. Although SIPPER ca n resolve and image smaller identifiable particles, the proportion of unidentifiable particl es rapidly increases below this size, reducing the effectiveness of the classifier. Zooplankton and other collected particles were gent ly washed out of the HRS plankton nets using a hand-held pump sprayer and into plasti c sample jars. They were immediately fixed in 5% v:v buffered formalin and stored for later analy sis at our lab. Samples were split using a Motoda box splitter into usable sub-samples of 1500 -2000 individual organisms. These were then analyzed under a dissecting scope where organisms w ere identified to species when possible. Up to forty individuals from each identified taxon for each sample were measured using an ocular micrometer and the appropriate size measurement was recorded. Nineteen samples comprising one nighttime (midnight) and one daytime (noon) ver tical profile were analyzed for this study. Because the nets collected smaller zooplankton than were extracted from the SIPPER dataset by PICES, we had to first identify the appropriate net zooplankton size threshold comparable to the SIPPER threshold of 550 m m ESD to allow for accurate comparison between the sampling effort of the plankton nets and the SIPPER. Direct areal c omparisons were not applicable as the SIPPER measured zooplankton in situ and the net collected zooplankton were measured af ter the rough handling in the net, wash down and preser vation. Additionally, because the SIPPER images zooplankton while in an undisturbed state, t he measured ESD will also include fine setae, antennae, tentacles, mucous, egg sacs, and any othe r body part or associated debris that will not be apparent in the same way in the preserved organi sm. Therefore I analyzed the hundred smallest extracted SIPPER images for each zooplankt on class found in the nets and collected appropriate measurements from each image to determi ne the minimum size of net collected zooplankton for comparison against the SIPPER datas et.
44 Description of SIPPER sensor The grayscale imaging SIPPER (informally known as S IPPER-2) is the successor to the prototype binary imaging SIPPER (Samson et al. 2001 Remsen et al. 2004). The SIPPER is comprised of five main components: a 9.6 9.6 cm ( 92.16 cm2) sample tube, a pair of collimated light sources, a pair of high-speed line-scan camer as and optical lens assemblies and a data storage device. Basically, the instrument operates by projecting a collimated light sheet across seawater passing through the sample tube and contin uously recording the outlines and silhouettes of suspended particles and plankton wit h a pair of high speed line-scan cameras mounted orthogonal to each other. While the standar d operating principles remain the same, there have been significant modifications to the SI PPER hardware to both improve the system and enable it to record in grayscale. Improvements to line scan camera technology have ma de faster, more sensitive and affordable digital cameras available for use. We re placed the two original cameras on the prototype SIPPER that each scanned a 2048 pixel lin e image at 15,000 lines s-1, with two new Dalsa Piranha 2 cameras capable of recording a 409 6 pixel line image at 21000 lines s-1. This achieved multiple benefits: it increased our image resolution, and it allowed us to tow SIPPER-2 at faster speeds while still capturing workable ima ges. These cameras were also smaller, more sensitive and used less power than the original cam eras. However, they also resulted in a nearly threefold increase in data rate due to their higher resolution and scan rate. To accommodate this increase in data rate, the data storage device had to be modified. Previously, 8-bit camera data were converted to bin ary (black and white) by applying a threshold with a field programmable gate array (FPGA) at the camera and then combined and compressed (30:1, real-time) using an FPGA processing board an d then buffered onto a stand-alone in-situ data storage system capable of recording up to 7.5 megabytes (MB) s-1. In the SIPPER-2 configuration, all image processing is done at the data storage system in a separate pressure vessel. Grayscale information is transmitted from t he camera pressure vessels to the data processing vessel via a low voltage differential si gnal (LVDS) parallel-serial data link. Since the grayscale information is still present, it can be s tored, assuming the storage system can maintain
45 the data rate. To achieve this, we developed a high er-speed (20 MB s-1), higher-density (112 GB) data storage device. It consisted of a custom FPGAprocessing board, a single-board PC, a National Instruments PCI-DIO-32HS digital input car d, a Boulder Instruments StreamStor 303 drive controller, and 112 GB RAID disk array. Even with this system, we limited ourselves to recording grayscale data from one camera at a time to reduce the potential for data loss due by approaching the maximum record rate. Image data were recorded in 3-bit grayscale (8 colo rs). This format was a compromise on being able to store the data quickly and also have time for image processing in 1 clock cycle of the FPGA while not filling the hard drives too rapi dly. After real time flat-field correction, the ra w 8-bit camera data were thresholded to 3 bits by com paring individual pixel intensity against a running average intensity value for each camera pix el measured every 50 ms. The average individual pixel intensity was divided into eighths with the first brightest division equivalent to white and the last eighth corresponding to black wi th six gray levels between the two and assigning the current pixel brightness to the corre sponding 3-bit grayscale value. The grayscale images collected by SIPPER-2 were qualitatively bet ter than that of the binary imaging SIPPER (fig. 17) and more amenable to automated classifica tion (Luo et al. 2005). The optical resolution of the SIPPER-2 was determin ed to be approximately 70 m m using a calibrated resolution target slide (USAF 1951 res olution target). The pixel resolution is controlled by two separate processes: The pixel res olution in one dimension is calculated by dividing the width of the sample tube by the number of pixels making up the sample line image (in this case 9.6 cm 3800 = 25.3 m m) and in the other dimension it is determined by d ividing the flow speed through the tube by the scan rate of the linescan camera (in this case 0.77 ms-1 21,000 lines s-1 = 37 m m). Particles and features smaller than 70 m m can be detected and imaged but their sizes will not be accurate.
46 Figure 17. Comparison of a similarly sized salp ima ge from the SIPPER-1 binary imaging sensor and the SIPPER-2 3-bit grayscale imaging sensor dem onstrating the increased detail visible with SIPPER-2. Detection and extraction of SIPPER images After sampling and between deployments of the HRS, SIPPER-2 data were offloaded to a shipboard PC. The SIPPER-2 data were then decompr essed and particle images were detected, extracted and classified using PICES. Ima ge detection is relatively straightforward with the SIPPER-2 as particles (grey or black pixels) ar e easily discernible from the background (white pixels) making segmentation relatively straightforw ard. Particle images may have white pixels within their boundary but must have a discernible e xternal edge made up primarily of foreground pixels to be detected and extracted. All resolvable particles within the sampling tube of SIPPER are assumed to be in focus. Particle detection and extraction begins with a sca nning of the entire continuous SIPPER image file that takes the form of a matrix 3800 pix els by 21,000 pixels the length of sample in
47 seconds, and an indexing of all occupied (foregroun d) pixel coordinates. A connectivity routine is then run on the index in which occupied foreground pixels within 2 pixels of another occupied pixel are connected and considered to be part of th e same particle and combined within the index. The perimeter and area of each particle in t he index are then calculated. Particle images at or above the user specified minimum-size are locate d within the index, extracted as bitmaps with a unique file name, and stored on disk. Concurrent with image extraction, morphological and texture features from each particle image were computed and written to a data file for use with the automated classification component of PICES. A total of 57 features were ext racted from each particle image for this study (Table 6). About half of these features were shapebased and initially used for classifying black and white images from the prototype SIPPER-1 (Luo e t al., 2004). These included the 8 invariant moments of the whole image and the image edge, 7 gr anulometric features (Tang et al., 1998), and domain specific features such as size, convex r atio, eigenvalue ratios and transparency ratio. An additional 20 texture and contour based features described in Luo et al., 2005 were also used for this study. These were comprised of 8 weighted moment invariants of the whole image, in which the grayscale intensity value of each pixel w eighted the calculation of the moments, 5 fourier descriptors of the image contour and textur e respectively, and a size and convex ratio weighted by the grayscale intensity of each pixel i n the image. Additionally we utilized 8 new features for this study. These included the graysca le intensity histogram of each particle image, in which the proportion of each 7 foreground colors is calculated and a height/width ratio of the bounding box of the image. For spatial statistics purposes, the two dimensional location of the centroid of each extracted particle within the SIPP ER sampling transect was calculated and included in the extracted data file. Manual classification of SIPPER images and developm ent of training library A subset of the extracted particle images were then classified manually to develop a training library for the classifier and to create a test set to gauge its accuracy. Prior sampling experience from the WFS using the HRS (Sutton et al ., 2001) indicated that the zooplankton community at the 50 m isobath was comprised of at l east two distinct communities: a calanoid
48 copepod dominated community in the well mixed surfa ce waters and a mixed copepod and ostracod layer below the pycnocline. Consequently, images from a shallow (3-5 m) and deep (4345 m) SIPPER daytime (~noon) sample were manually s orted to determine the initial composition of the training library for use with our classifier An additional two samples, comprising night (~midnight) shallow and deep tows were used to vali date the automated classification system as a test set and to compare the day-night performance of the classifier. These four SIPPER files comprised a total of 109122 images. An average SIPPER tow from this study sampled appro ximately 3.4 m3 of seawater in 8 minutes. We used this volume to determine the maxim um concentration an organism could be present in the water column and not be detected by SIPPER with a 5% probability using l = ln p/v where l is the concentration of the organism, p is the prob ability of detection and v is the volume of water sampled (Benfield et al. 1996). For SIPPER -2, l was ~4 individual m-3. Therefore, when building our preliminary training library we counte d as a class any taxonomic group recognizable by SIPPER that numbered more than 5 occurrences in any of our two training-set samples as a class to include in the first run of our automated classification system. This was done to ensure that we didnÂ’t overlook a group that might be more common in another sample even though it was relatively scarce but found in detectable numbers i n the test set. To ensure we wouldnÂ’t positively bias classifier performance, training image example s from a file being classified were not included in the classification of that SIPPER file.
49 Table 6. List of features extracted from every SIPP ER-2 image and those used for the classifier after running the Â“wrapperÂ” feature optimization al gorithm. Feature name Number of features Used after Â“wrapperÂ” Moment invariants of the original image 8 1 Moment invariants of the edge image after closing 8 2 Weighted moment invariants on the whole image 8 2 Fourier texture 5 5 Fourier contour 5 1 Granulometric features 7 3 Intensity histogram 7 6 Domain specific features Size 1 Convex area 1 1 Transparency ratio 2 1 Eigenvalue ratio 2 Head-tail ratio 1 Weighted size 1 1 Weighted convex ratio 1 1 Automatic classification of images PICES incorporates a multiple-class SVM to classify SIPPER particle images (Luo et al. 2004, Luo et al. 2005). Briefly, SVMs are a set of margin based linear classifiers that work by mapping the training data from two classes in multi dimensional space using the feature vectors extracted from the example images in the training l ibrary. A decision boundary or hyperplane is computed that maximizes the margin that separates t he two groups using example data known as support vectors that lie closest to the boundary The SVM classifies unknown images based on where they lie along the decision boundary descr ibed by the support vectors. Mapping of the feature data into the higher dimension feature spac e is accomplished by the use of a kernel function. The kernel function allows a linear class ifier to solve a non-linear classification problem. To extend SVMs from 2-class to multiple class class ification problems, we used the pairwise or one-versus-one approach (Luo et al., 20 04) where all possible pairs of classes are used to create binary SVMs so that the total number of classifiers is equal to n(n-1)/2 where n is equal to the number of classes chosen. Unknown exam ples are then voted on by each of the
50 SVMs; the one that receives the greatest number of votes becomes the predicted class. Each SVM classification is also assigned a probability b ased on its proximity to the decision boundary. These probabilities are used to break ties if an im age has an equal number of votes for more than one class. Classification error analysis Classifier performance was calculated by examining the classifier accuracy rate as determined by a confusion matrix built using a grad ing utility in PICES. By comparing the results of the manually sorted test-set and a computer clas sified test-set, each cell of a confusion matrix is occupied by the number of images classified as a specific image class. The diagonal of the matrix then contains the number of images labeled c orrectly by the computer classifier for a specific class. This number divided by the total h uman count for a specific class is called the classifier accuracy (Davis et al. 2004). This assum es the human classification is error-free, which is most often not the case (Culverhouse et al. 2003 ). To quantify that assumption, we had another plankton biologist classify a random subset of the training library so that we could determine potential human error in grading and comp are it against the computer classification accuracy. Analysis of SIPPER-2 spatial pattern data The SIPPER-2s capability to continuously image a re latively large volume of water makes it possible to investigate the spatial relationship s between individual particles within the sampled volume. Because SIPPER is sampling a three-dimensio nal environment and imaging it in twodimensions, it is not possible to determine the tru e linear distances between individual particles with total accuracy as we cannot determine where a particle is vertically within the ~10 cm height of the sample tube. However, by sampling enough par ticles, finescale relationships can still be explored. To investigate apparent nearest neighbor distances (ANND) between individual particle classes, we used a modified version of the Spatial Analysis Technique (SPLAT, Widder and Johnsen, 2000, Malkiel et al., 2006) that uses Mont e Carlo methods to compare observed ANND and those expected if the particles were randomly d istributed. Random distributions were created using a random number generator to create two-dimen sional coordinates of the particles within a
51 simulated volume the same dimensions as the SIPPER transect using the observed abundance and the length of the SIPPER sampling transect (Mal kiel et al., 2006). The simulations were further refined using the observed size distributio n of the particles to avoid biasing the distributions towards a regular distribution due to a mismatch between the size of the actual particles and the simulations (Widder and Johnsen, 2000). SIPPER estimates of ANND were calculated for individuals from each classified gro up in the manually classified test set from the two-dimensional centroid location of each particle in the SIPPER sampling path. ANND measurements were also calculated for the entire pa rticle assemblage within a sample. These observed ANND were then compared against thos e calculated from 1000 Monte Carlo simulations of each groupÂ’s distribution. Sim ulated ANND were determined in two steps: first, the simulated particle class was randomly po pulated in the SIPPER sample volume at the observed abundance and second, the centroid positio ns of each simulated particle were then determined and the nearest neighbor distance betwee n each particle class was calculated. Cumulative histograms of ANND from the simulations were used as models of complete spatial randomness and then compared against those from the observed data to determine whether the groups were randomly distributed. Histogram bin siz e was determined for each class individually by taking the maximum observed ANND and dividing it by 500. The number of ANND occurrences in each size bin was then calculated. This is illustrated in figure 18, where the y axis indicates the proportion of occurrences of a given NND equal to or less than the NND value indicated on the x-axis. Each histogram has four pl ots: the average from the simulations (dotted line), the minimum and maximum bounds from the rand om simulations (dashed lines) and the observed data (solid line). Observed data with a lo w proportion of small nearest neighbor distances would indicate a regularly spaced particl e distribution (fig. 18a), while an observed distribution with a large proportion of small NNDs would indicate an aggregated distribution (fig 18c). An observed distribution falling within the bounds of the simulated random distributions is considered to be randomly distributed (fig.18b). We tested whether the observed distributions were significantly different from the simulated ran dom distributions using a U2 statistic. This statistic is the sum of squares of the deviations f rom the average cumulative histogram from the
52 simulations and compared against each simulation an d the observed data. The rank of the U2 statistic for the real data relative to the simulat ions can be used as a probability value to determine if the observed NND can be explained by c hance. If the real datasets U2 value was greater than 950 or more of the simulations U2 values, there was less than a 5% chance that the distribution could be due to chance. The shape of t he distribution then tells you if the spacing was aggregated, random or regular. While measuring ANND is relatively straightforward, these measurements are subject to biases introduced by the finite spatial coverage of the sampling device. Particles at the edge of the sampling area have a high probability that thei r nearest neighbor is located outside the area sampled (DeRobertis, 2002). Replacing the true near est neighbor with the closest one inside the sampling area results in an overestimation of NND a nd biasing the resulting distribution towards regularity (called Â“edge effectsÂ”). We used the Â“Z-scoreÂ” to account for these edge eff ects and allow for direct comparison of NND of particles at different particle densities in whi ch the observed mean NND at a given target density was standardized to the NND expected under a random spatial distribution within the exact volume sampled (Malkiel et al., 1999, Malkiel et al., 2006). The Z score is calculated by N X X zr r ss= Where sX is the mean ANND of the sample, N is the number of particles in the sample, while rXand rsare the mean and standard deviation from the 1000 r andom simulations. The Z score indicates by how many standard deviations the mean ANND deviates from that expected under complete spatial randomness. A Z<0 indicates a tend ency towards aggregation and a score less than -2 is significant at the 0.05 probability leve l while a Z>0 indicates a tendency towards regularity and a score greater than 2 would be sign ificant. The Z score was compared against the results from the U2 statistic for both the manually classified and PIC ES classified test set.
53 While the U2 statistic and Z-score provide us with information on the ANND distribution of a particular particle class, it provides no context on the degree of spatial heterogeneity along the entire sample transect (~400 m in 8 minutes). Simpl y stated, the plankton of a particular class could be closer to each other than could be explain ed by chance but still be more or less homogenously distributed along the sampling path. T herefore, we used LloydÂ’s index of patchiness (Lloyd, 1967) to measure the degree of a ggregation at the meter scale for the 4 samples comprising the test set. This measurement is scale dependent, as it requires the researcher to make a decision on what scale to exam ine the potential variance along the sample transect. We counted the number of each particle im age class occurring along each meter of sampling in the 4 test set samples to calculate the index of patchiness at the meter-scale. LloydÂ’s Index of patchiness is determined using x x x s P1 ) 1 (2+ =, where x is the mean and s2 is the variance of the number of the selected part icle class in a series of SIPPER samples of 1 m length. P indicates the di stribution of individuals in a given sample relative to that expected in a Poisson distribution A P equal to 1 corresponds to the crowding expected in a Poisson distribution while a P equal to 2 indicates the individuals are twice as crowded at the meter-scale than if they were random ly distributed.
54 Figure 18. Examples of regular (a), random (b), and aggregated distributions with their cumulative histograms superimposed on the cumulative histogram s generated by Monte Carlo simulations. The dotted lines show the upper and lower envelopes of a random distribution. The solid line shows the data. (From: Widder E.A., Johnsen S., 3D spatial point patterns of bioluminescent plankton: a map of the minefield. Journal of Plankt on Research, 2000 22(3), 409-420, by permission of Oxford University Press.)
55 Results Hydrography Typical summer conditions were encountered on the W FS with the water column being well stratified with the halocline, pycnocline and thermocline found between 30 and 35 m (fig. 19.). The distribution of fluorescence and extracte d chlorophyll a concentrations were similar to the other measured environmental parameters, with l ow concentrations above 35 m and rapidly increasing concentrations below that depth (Fig.20) Figure 19. Vertical profiles of A: Temperature, B: Salinity and C: Sigma-t collected during the 24 hour sampling period.
56 Figure 20. Vertical profiles of A: fluorescence (Vo lts) and B: extracted chlorophyll collected during the 24 hour sampling period. Locations of bottle sa mples for extracted chlorophyll are noted by the white asterisks. Total SIPPER Particle Abundance and Size Distributi on A total of 1391227 particle images of >550 m m ESD were extracted from the 70 discretedepth SIPPER samples. The distribution of total pa rticle abundance was similar to the fluorescence and chlorophyll profiles, with conce ntrations highest in the subpycnocline layer and lowest near the surface. A maximum concentration o f 16608 particles m-3 was sampled at 45 m depth and a minimum of 2007 particles m-3 was collected at 3 m (fig. 21). Particle abundance increased in the upper water column during the late morning through early evening but never exceeded that of the subpycnocline layer. The major ity of these particles were small, with 37% of the total between 0.55 and 0.7mm ESD and 85% of the particles less than 1 mm ESD (Fig. 22) Selection of Image Classes and Classifier Performan ce A total of 48 separate particle and plankton image classes were identified in detectable numbers from the two SIPPER files for building the initial training library to automatically classify the SIPPER dataset (Table 7). This included an othe r and unknown class in which both unidentifiable particle images and identifiable but rare groups were placed. The training library also included two artifact classes: vertical lines that can appear when a particle temporarily adheres to the imaging window (scanline artifacts) and bubbles from towing too close to the surface, thereby sampling the shipÂ’s wake. In the i nitial sort, the number of individuals per
57 identified particle class ranged from 13 to over 20 00, with the majority of classes containing less than 100 image examples. Using this as the prelimin ary training library (with a maximum of 100 randomly selected images from the more abundant cla sses to reduce their influence on the classification task), the multi-class SVM classifie r was run on the entire SIPPER dataset. The resulting classifications were analyzed using a com mercially available thumbnail viewing software package Thumbs Plus. This involved a search of over 3300 separate image folders to determine which classes were to be used for further optimizat ion of the classifier training library. During this phase, the performance of the classifier was not de termined; rather we examined these results to estimate which image classes would be suitable for use in the final classifier. This phase was also used to further populate the training library with representative images for each class. Those image classes found in detectable numbers in over half the deployment samples were held onto for use in the training library for the next test of the classifier. Additionally, two image classes found at high abundances only amongst a narrow depth stratum and at nondetectable levels elsewhere, were also included in the reduced training library. Bubbles were found in 6 of the 10 SIPPER files closest to the su rface depth range, and most likely resulted from sampling the shipÂ’s wake. The cladoceran Penilia avirostris was found at the deepest sampled depth in high abundance but at non-detectable conce ntrations in shallower waters. Using the defined criteria of either being found in detectable numbers in over half the samples or being extremely abundant in just a few, the number of classes was reduced to 31. The number of individuals per training class was in creased up to 500 example images to ensure that the morphological and textural diversity of ce rtain groups were well represented. Discarded classes were either added to the other and unknown class or combined into a larger, more comprehensive class with shared morphological or te xtural traits. For example, a number of morphologically similar Â“shrimp-likeÂ” groups includ ing amphipods, decapod zoea, euphausiids, Lucifer sp., other shrimp and stomatopod larvae were combi ned into the larger group Crustaceaneumalacostraca that shared traits such as a large o paque carapace, obvious telsonic fan and relatively large size. Some of the discarded class es although rare, were observed exhibiting some interesting behaviors. Scyllarid lobster phyl losoma were often observed directly associated
58 with other plankton and suspended particle groups, especially fragile and gelatinous forms. Of the 62 phyllosomes detected in the initial search, 38 were imaged clinging to one or more hydromedusae, siphonophores, protoctists and marine snow using their pereopods. Figure 21. Distribution and abundance of all partic les (no bubbles and scanline artifacts) imaged by SIPPER greater than 600 m m ESD in size over the 24 hour sampling period.
59 Figure 22. Size distribution in equivalent spherical diameter of the total SIPPER particle images collected during this study (N=1391227)
60 Table 7. Image classes used during development of t he multiple-class SVM classifier used in this study The number of image classes was reduced during optimization of the clas sifier. Numbers in parentheses represent the image classes integrated into larger more generic image classes during optimization. Initial Training Library Reduced Training Library Final Training Library 1. Artifact lines 25. Gelatinous-ctenophore-cydippi d 1. Artifact lines 1. Artifact lines 2. Amphipod 26. Gelatinous-ctenophoreOcyropsis 2. Bubbles 2. Bubbles 3. Chaetognath 27. Gelatinous-doliolid 3. Chaetogna th 3. Chaetognath 4. CladoceranPenilia avirostris 28. Gelatinous-salp 4. CladoceranPenilia avirostris 4. CladoceranPenilia avirostris 5. CnidariaAglaura 29. Gelatinous-siphonophore-caly 5. CnidariaAglaura hemistoma 5. CnidariaAglaura 6. Cnidaria-leptomedusae 30. Gelatinous siphonophor eSphaero 6. Cnidaria-other (6) 6. Cnidaria-other 7. Cnidaria-other 31. Lancelet 7. Copepod-calanoid (9) 7. Copepod-calanoid 8. Copepod-calanoid 32. Larvacean 8. CopepodOithona 8. CopepodOithona 9. Copepod-calanoid Calocalanus 33. Lobster phyllosome 9. Copepod-other-eyed 9. Co pepod-otherMacrosetella 10. CopepodCopilia 34. Marine snow 10. Copepod-otherMacrosetella 10. Copepod-otherOncaea 11. Copepod-with eyes 35. Ostracod 11. Copepod-othe rOncaea 11. Crustacean-eumalacostraca 12. CopepodMacrosetella 36. Other-unknown 12. Crustacean eumalacostraca (2,15,40-42) 12. Echinoderm-plutei 13. CopepodOithona 37. Polychaete 13. Echinoderm-bipinnaria 13. Elong ate-phytoplankton 14. CopepodOncaea 38. PteropodCreseis 14. Echinoderm-plutei 14. Elongate-trichomes 15. Decapod-zoea 39. PteropodLimacina 15. Elongate-phytoplankton 15. Gelatinous-doliolid 16. Echinoderm-bipinnaria 40. Shrimp-Lucifer 16. El ongate-trichomes 16. Larvacean 17. Echinoderm-plutei 41. Shrimp-other 17. Fish 17. Marine snow 18. Egg clusters 42. Stomatopod 18. Gelatinous-cten ophore-cydippid 18. Ostracod 19. Elongate-diatom chains 43. ProtoctistAcanthametreon 19. Gelatinous-doliolid 19. Other-unknown (9,13,17-18,20-22,25-27) 20. Elongate-dinoflagellate colonies 44. Protoctist -radiolarian colony 20. Gelatinous-salp 20. Protoct ist-sarcodine 21. ElongateTrichodesmium colonies 45. Protoctist-species A 21. Gelatinous-s iphonophore (29,30) 21. Trichodesmium tuft and puff colonies 22. Fish 46. Protoctist-species B 22. Lancelet 23. Fish egg 47. ProtoctistThallasicola 23. Larvacean 24. Gelatinous-ctenophoreBeroe 48. Trichodesmium tuft and puff colonies 24. Marine snow 25. Polychaete 26. PteropodCreseis 27. PteropodLimacina 28. Ostracod 29. Other-unknown (16,23,24,26,33) 30. Protoctist-sarcodine (43-47) 31 Trichodesmium tuft and puff colonies
61 The classifier was then rerun using a split trainin g library, with the two shallow water samples using a 30 class training library that incl uded the bubble class, while the two deep water samples were run using a 30 class classifier that i ncluded the cladoceran class, so that the total number of classes used was 31. The results from the automatic classification of the test set were compared against the manually classified test set u sing a confusion matrix (table 8) created using PICES. Overall, there was a 71.8% agreement betwe en the results of the manual and automatic classification of the test set. Classification per formance was strongly associated with the relative abundance of the image class. Abundant groups compr ising 5% or more of the total had accuracies of between 61 and 83 % (mean=72.3%), wit h the lowest accuracies coming from marine snow and the other and unknown class, both o f which are comprised of a wide range of particle sizes and morphologies. Common groups mak ing up between 1 and 5 % of the total had a slightly lower overall recognition rate (mean=71. 5%) than the abundant groups and a slightly broader range in their individual classification pe rformance (54.0 to 78.0%). The remaining 16 image classes were considered rare, individually re presenting between 0.7 to .01 % of the total. The classification performance of these rare groups had the broadest range in classifier performance (10.0 to 94.7%) and the lowest aggregat e classification accuracy (61.6%). This led to instances where the computer count strongly dive rged from the human count for a given rare class. Whereas computer counts of the abundant and common groups were always within half or double that of the human counts (C/H ratio of 0.68 to 1.92), 9 of 16 rare groups had computer counts more than twice as high as the human count. While the overall classification accuracy was satis factory, the performance of the classifier with regards to the rare groups was less so. The proportion of incorrectly labeled images (Â“false positivesÂ”) is much higher for rare groups, leading to significant overestimation of their abundance (Solow et al., 2001, Davis et al., 2004). For example, while only 40 cydippid ctenophores were identified by the human expert in the test set, 554 were labeled ctenophore by the computer. Similarly, only 8 of the 158 images labeled fish or fish larvae were actually fish. This led to a second reduction to the number of cla sses used in the classifier. Ten groups representing less than 0.75% of the total were remo ved from training library and the classifier was
62 rebuilt. The subtracted groups training examples we re added to the other and unknown class training class. The classifier was then retrained w ith the 21 classes remaining and run again on the test set using the same split-training library described earlier. Classifier performance was recalculated using PICES with this smaller training library and accuracy was found to have improved to 73.8% (Table 9). The C/H ratio was also much narrower for the image classes, ranging from 0.69 to 2.62. F or the three classes with classification accuracies below 60%, two (cnidaria-other and crust acean-eumalacostraca) belonged to composite groups representing a broad range of shap e, color and size. The poor performance of the third group, the small robust poecilostomatoid copepod genus Oncaea, appeared to be due to a combination of small size and confusion with smal l calanoid copepods, ostracods and the other and unidentified class. The four groups with C/H va lues greater than 1.8 were the 4 least abundant classes in the test-set. The final classif ication was executed using three separate subsets of the remaining 21 classes in the training library. The classifier for 13 shallow water (<5 m) samples was built using all the of the training library except for the cladoceran class while the classifier for the 7 deepwater samples used all of the training library save for the bubble class. The remaining 50 SIPPER samples were classified usi ng a 19 class training library that did not include the bubbles and cladoceran class. Human expert classification error estimation For a machine classifier to perform well, it must b e trained with correctly labeled example images. Previous work has demonstrated that taxonom ists are imperfect in their labeling and classification performance (Culverhouse et al., 200 3). To quantify the potential problem of incorrect labeling by human experts, we compared a subset of the training data sorted by the primary operator of PICES with that of another plan kton biologist at the College of Marine Science at the University of South Florida. We ach ieved approximately 81% consensus (Table 10) in our labeling efforts with most of the error confined to Macrosetella gracilis (47% CA), echinoderm plutei (97% CA) and the other and uniden tified class (50% CA). This led to a very tight agreement in abundance estimates with ratios between the two human experts ranging between 48% and 152%. This is less than the variabi lity expected within taxonomic groups
63 between replicate net tows that can range between 2 5-300% (Wiebe and Holland, 1968; Pillar, 1984). Example Images The 21 image classes chosen for the final classifie rs were mostly comprised of distinct taxonomic categories that were recognizable to a hu man expert. Representative images for each class save the two artifact classes can be seen in figures 23-36. Four of the image classes were comprised of individual species: the trachymedusae Aglaura hemistoma the ostracod Euconchoecia chierchiae the cladoceran Penilia avirostris and the poecilostomatoid copepod Macrosetella gracilis while four others were comprised of single genera (the copepods Oithona and Oncaea, and two forms of Trichodesmium colonies. Trichodesmium sp. was split into two separate image classes because there were two disti nct groups of imaged morphologies: an elongate linear morphology, and a larger tuft and p uff morphology. The calanoid copepod, chaetognath, larvacean and echinoderm plutei classe s all consisted of an unknown number of relatively similar species. The doliolid class was made up of a number of life history stages with observable nurse, oozoid, trophozooid phorozoid a nd gonozoid forms but most were not identifiable to species. The cnidaria other, crusta cean-eumalacostraca, and protoctist classes were all comprised of numerous morphologically dive rse taxa, some of which were recognizable to genus or species but too uncommon or difficult t o identify to rate a class of their own. The elongate phytoplankton class consisted of dinoflage llate and diatom chains often identifiable to genera such as Chaetoceros and Proboscia for diatoms and Ceratium for dinoflagellates. The marine snow category was composed of a large variet y of detrital aggregates with individual aggregates consisting of recognizable components su ch as cast-off larvacean houses, fecal pellets, phytoplankton and protoctists as well as u nidentifiable debris. These marine snow particles spanned a wide range of sizes from less t han a millimeter in diameter to over 4 cm. Similarly the other and unidentified class was comp osed of both recognizable taxa (fish, polychaete, pteropod, salp, siphonophore, etc.) and unidentifiable particles spanning a wide size range.
64 1. Artifact lines 2. Bubbles 3. Chaetognath 4. Cladoceran-Penilia avirostris 5. Cnidaria-aglaura hemistoma 6. Cnidaria-other 7. Copepod-calanoid 8. Copepod-oithona 9. Copepod-other-eyed Table 8. Confusion matrix of 31 class SVM classifie r for SIPPER test set (N=109122 particle images). T H= Total human counts, TC= total computer counts and A= classification accuracy. 123456789101112131415161718192021222324252627282930 31 TH 1741 0001610000000930001013016000000782 2 0 2697 00125010271100001000010046185000030104079 3 00 3227 101536440101003959322163533011317626811164116 4 000 3740 1062023133438160104800001065296910019511005040 5 008917 6193 220857020112240832633691618290196127044228195 6 104043 120 010301102282121802642969532513446 7 00252684 7478 400771141363011126900405013422164102926779572 8 102161402499 5106 2137400341540000782843200001356176 9 000200250 89 47400000000000156000204158 10 000700930 274 300006000002206000001313 11 010160067167 800 20000000009337517200061151481 12 0001542150514 142 0003000102011121782412024302 13 000010100001 36 00000000440000000249 14 0000140102000000 145 000000056050000172224 15 14013347002000000 745 142900030410900100101093 16 00203312029190163230022 1642 00057254660715213700892545 17 1010030100000010 17 10300210000000040 18 1060321800000000602 185 210216012000030286 19 100011000000000000 7 00000000000111 20 001032521330001002019150 71 091204002003203 21 002500000000000060100 12 68002104000119 22 00162153762649591515411333103387271501594 10193 543404676523128612019912752 23 11114318038114339013254057666281361087011111040 5419 3236814419371642058870 24 0230231018304263251800010100078 5309 854000185536877 25 341843747876063963331636116565950769467117215122448 17038791324 19938 3513667220933230053 26 0000000000070000000003001 8 0100121 27 00300000000300020000050002 14 000231 28 0040000100020006000001940230 29 00171 29 0006002002230000000006235200 35 0169 30 0014304291110001162120917020225311360000 1124 181547 31 01920456853816415551037800401781151172622141312 2823 3601 TC1112280743985796758349697686155316761199660654528 210519905547712022219814243826581252362215811132436 916304039 109122 A0.950.660.780.740.760.270.780.830.560.880.540.470. 730.650.680.650.430.650.640.350.100.800.610.770.660 .380.450.410.510.730.78 0.72 C/H1.420.691.071.150.931.111.021.002.002.431.352.01 1.102.361.930.7813.852.701.821.091.661.120.931.180. 797.523.584.565.351.051.12 10. Copepod-otherMacrosetella 11. Copepod-otherOncaea 12. Crustacean-eumalocostraca 13. Echinoderm-bipinnaria 14. Echinoderm-plutei 15. Elongate-phytoplankton 16. Elongate-trichomes 17. Gelatinous-ctenophore 18. Gelatinous-doliolid 19. Gelatinous-salp 20. Gelatinous-siphonophore 21. Lancelet 22. Larvacean 23. Marine snow 24. Ostracod 25. Other and unknown 26. Other-fish 27. Other-polychaete 28. Other-pteropodLimacina 29. Other-pteropodCreseis 30. Protoctist-sarcodines 31. Trichodesmium tuft and puff colonies
65 1. Artifact lines 2. Bubbles 3. Chaetognath 4. CladoceranPenilia avirostris 5. Cnidaria-Aglaura hemistoma 6. Cnidaria-other 7. Copepod-calanoid 8. Copepod-oithona 9. Copepod-othermacrosetella Table 9. Confusion matrix of 21 class SVM classifier for SIP PER test set (N=109122 particle images). TH= Total human counts and TC= total computer counts. TH= Total human counts, TC= total computer counts and A= classification accuracy. 123456789101112131415161718192021 TH 1 727 030132000007202403100 782 2 0 3301 00125001810000015017554408 4079 3 00 3380 647185110003547303131141104114 4117 4 001 3909 2059992148171120103589596177 5040 5 007112 6339 29295320121802295183302094223 8195 6 1028277 317 3430405234427188116 689 7 00142659 7830 3399520325925045123581302355 9572 8 102186401529 5072 11670334608330336126 6176 9 0001100135 262 410050320601 313 10 010220012209 857 7001053427166118 1639 11 0011665130114 149 00202391227024 302 12 0000160821000 138 0005705222 224 13 40114172020000 740 1392840013210 1093 14 0024102312101617322017 1709 1240510138045 2545 15 0080312200000030 191 21501320 287 16 00145132683849581252916266818 10610 4944168319144 12754 17 501452253646534881840505126330771092 5417 29565160152 8870 18 023024231930522412000085 5508 820644 6877 19 290148472861566181836280825301136774492147184790914 47 20281 201325 30419 20 001428928140101581001626621338 1118 21 1547 21 021012471588353951204401721171511012 2858 3602 TC1028347546235849767099410376599764720105534861865 202875214715825381322421516113843109122 A0.930.810.820.780.770.460.820.820.840.520.490.620. 680.670.670.830.610.800.670.720.790.74 C/H1.310.822.214.171.1241.441.080.972.071.231.832.17 1.710.802.621.150.931.180.801.041.07 10. Copepod-other-Oncaea 11. Crustacean-eumalacostracan 12. Echinoderm-plutei 13. Elongate-phytoplankton 14. Elongate-trichomes 15. Gelatinous-doliolid 16. Larvacean 17. Marine snow 18. Ostracod 19. Other-unknown 20. Protoctist-sarcodine 21. Trichodesmium tuft and puff colonies
66 1. Artifact lines 2. Bubbles 3. Chaetognath 4. CladoceranPenilia avirostris 5. Cnidaria-Aglaura hemistoma 6. Cnidaria-other 7. Copepod-calanoid 8. Copepod-oithona 9. Copepod-othermacrosetella Table 10. Confusion matrix of classification perfor mance between the primary human classifier of SIPPE R images (TH1) and a secondary human expert (TH2) in sorting a subset (N=2178 images) of the SIPPER test set. A= classification accuracy. 123456789101112131415161718192021 TH 1 1 99 00000000000000000000 99 2 0 98 0000000000000001000 99 3 00 80 00000000010150100200 108 4 000 83 00000000000024100 90 5 0000 101 80100025041000905 154 6 00003 101 000000000010111 108 7 000100 109 50100000001000 117 8 0000005 75 1024010021700 98 9 00000064 38 1000500002700 81 10 020000000 56 00000009000 67 11 0000001000 68 00300200010 84 12 00000000000 96 000010020 99 13 000000000000 100 30010400 108 14 0000000000002 86 0000200 90 15 02000000000000 66 0201810 89 16 0023002000000011 76 40900 116 17 0000000000000010 118 0060 125 18 013050000000000000 62 100 81 19 01350000008402202253 67 02 133 20 00000000000250000000 72 0 97 21 000000000000012000011 121 135 TH2991281088910411112185396674150114132697915881149831 392178 A1.000.990.740.920.660.940.930.770.470.840.810.970. 930.960.740.660.940.770.500.740.900.81 TH 2 /TH 1 1.001.291.000.990.681.031.030.870.470.990.881.521.0 61.470.780.681.261.001.120.861.03 10. Other-unknown 11. Protoctist-sarcodine 12. Trichodesmium tuft and puff colonies 13. Copepod-other-Oncaea 14. Crustacean-eumalacostracan 15. Echinoderm-plutei 16. Elongate-phytoplankton 17. Elongate-trichomes 18. Gelatinous-doliolid 19. Larvacean 20. Marine snow 21. Ostracod
67 Figure 23. SIPPER images of A: the cladoceran Penilia avirostris and B: The ostracod Euconchoecia chierchiae Figure 24. SIPPER images of chaetognaths. Several behaviors are visible in these images: beginning clockwise from the top, reproduction, par asitism, defecation, cannibalism and predation.
68 Figure 25. SIPPER images of the trachymedusa Aglaura hemistoma Figure 26 SIPPER images of hydromedusae and narcomedusae mak ing up the other cnidarian class
69 Figure 27. SIPPER images of various calanoid copepo d species. Figure 28. SIPPER images of A: the cyclopoid copep od genus Oithona B: The poecilostomatoid copepod species Macrosetella gracilis and C: the poecilostomatoid copepod genus Oncaea
70 Figure 29. SIPPER images of eumalocostracan crustac eans Figure 30. SIPPER images of A: echinoderm plutei an d B: Elongate phytoplankton colonies.
71 Figure 31. SIPPER images of doliolids. Figure 32. SIPPER images of larvaceans.
72 Figure 33. SIPPER images of marine snow Figure 34. SIPPER images of the other and unknown p article class.
73 Figure 35. SIPPER images of various forms of sarcod ine protoctists Figure 36. SIPPER images of A: Elongate trichomes a nd linear colonies of Trichodesmium and B: Tuft and Puff colonies of Trichodesmium
74 Taxonomic composition of imaged dataset Close to one-quarter (23%) of the SIPPER images, co mprising 320,116 individual images belonged to the other and unidentified class (figur e 37), making it the single most abundant class in the dataset. The 13 groups making up the identif ied zooplankton totaled just over 52% of the total imaged particle assemblage. Larvaceans were t he most numerous of the identified zooplankton classes, comprising just over 13% of th e total extracted particle images. Four other zooplankton groups (calanoid copepods, ostracods, t he hydromedusae Aglaura hemistoma and Oithona sp.) comprised between 6-10% of the total. Chaetog naths comprised slightly less than 5% of the total, while the other 7 zooplankton imag e groups contributed between 2.1 and 0.75% to the total assemblage. The colonial cyanobacteria Trichodesmium was made up of two separate image classes (round and linear colonies) that together made up just over 8% of the total. Amongst protoctists, sarcodines contributed 3.5% to the total, while elongate dinoflagellate and diatom chains contributed another 2.8%. Marine snow was the third most abundant image group, comprising slightly less than 12% of the tot al extracted image assemblage. Figure 37. Composition of the SIPPER dataset as det ermined through the three multiple class SVMs run on the dataset.
75 Abundance and distribution from SIPPER-2 Other and unidentified class The other and unidentified class had a very broad size distribution and was comprised of particles from less than 0.6 mm to greater than 5 mm ESD. However, Small particles dominated the size distribution with 84% of the ima ged particles measuring 0.8 mm E.S.D. or less (figure 38). A substantial proportion of these par ticles were most likely small forms of the plankton classes chosen for classification but unab le to be identified due to lack of resolvable features and/or presenting an ambiguous profile whe n imaged by the SIPPER-2. Their distribution mostly trended with the total particle distribution with highest concentrations in the sub-pycnocline layer (figure 39). Abundances in the upper 30 m were mostly evenly distributed, ranging between 400-1600 m-3 while below the pycnocline they were more dynamic, with abundances between 2000-6000 m-3. More unknown and unidentified particles were foun d below the pycnocline in the first half of the sampling pe riod (midnight to noon) than in the latter half of the day. Proportionally, this class made up between 15-37% of the imaged particles in a sample with no apparent trend with depth.
76 Figure 38. Size distribution in ESD for the non-zoo plankton image classes. Figure 39. Distribution and abundance of the other and unidentified particles image class collected during this study. Classification accurac y of PICES for this group was 67%.
77 Zooplankton Total imaged zooplankton made up between 28-70% of the imaged particle assemblage for the 70 SIPPER samples and comprised over 52% (7 29513 zooplankton images) of the total SIPPER image dataset collected for this study. Cons equently, the distribution of zooplankton generally mirrored that of the total imaged particl e distribution, with maximum abundances found below the pycnocline and much lower abundances abov e (Figure 40). Most of the individual zooplankton classes observed this trend as well. Ab undances ranged between ~1000-2000 m-3 in the upper 30 m to greater than 6000 m-3 below 40 m. Zooplankton made up more of the total imaged particle assemblage with increasing depth, c omprising on average 46% of the images in the 10 m depth stratum and 58% below 40 m. A patch of elevated zooplankton abundance was observed between 15-30 m during the late afternoon and early evening. Figure 40. Distribution and abundance of the 13 SIP PER imaged zooplankton classes chosen for the final classification. Aggregate classification accuracy of PICES for these 13 groups was 74 %. Larvaceans were the dominant zooplankton group, com prising just over 14% of the particle assemblage as determined using the 21-clas s SVM. They were abundant at all depths, comprising between 10-40% of the imaged zooplankton assemblage, averaging just under 25%.
78 Larvacean abundance was highest below the pycnoclin e and generally decreased towards the surface (figure 41). Larvacean abundances ranged fr om 284 m-3 at 3 m depth to 2700 m-3 at 43 m. Larvacean abundance was elevated in the mid-wat er patch detected in the late afternoon and early evening. The majority of sampled larvaceans w ere imaged separate from any visible house structure. Approximately 15% of the larvacean image s contained the organism along with some part of the inner filter and or debris collected on the exterior of the house structure. The incidenc e of these images increased with depth. The majority of larvaceans were less than 1mm ESD in size, with only 20% of the classified larvaceans me asuring greater than that (figure 42). Much of those larger sizes can probably be explained by mea suring the imaged house and filters along with the larvacean. Larvaceans were not identified to species in the net samples. Figure 41. Distribution and abundance of larvacean images collected during this study. Classification accuracy for this group was 83%. Calanoid copepods were the next most abundant zoopl ankton group imaged by SIPPER. They displayed a similar trend as larvaceans with h ighest abundances below the pycnocline and decreasing concentrations toward the surface (figur e 43). They comprised between 6-32% of the imaged zooplankton at any one depth, averaging just over 17% of the assemblage. Abundances ranged from 129 m-3 at 6 m depth to 1956 m-3 at 42 m. A number of calanoid copepod genera
79 were recognizable from the SIPPER images and verifi ed from net samples including Calocalanus Centropages Candacia Eucalanus Euchaeta Mecynocera and Temora but the majority were small copepods with few distinguishing characterist ics and were most likely members of the Paracalanidae and Clausocalanus sp. and copepodite stages of larger copepods. Over 86 % of the imaged calanoid copepods had ESDs of 1 mm or le ss (fig 44) but some were measured as large as 3.5 mm. There was no apparent trend in cop epod size with depth. Figure 42. Size distribution in ESD for the non-cru stacean zooplankton image classes. Ostracods were the third most abundant zooplankton group, comprising just over 8% of the total imaged particle assemblage and from <1 to 40% of the zooplankton assemblage at the sampled depths. Ostracod abundance was highest betw een 30-40 m and was strongly associated with the pycnocline (figure 45). Ostraco ds were virtually absent in the upper 10 meters during the day but increased in the surface waters throughout the evening while decreasing at depth suggesting that part of the population may ve rtically migrate. The majority of classified
80 ostracods were small, with 95% of the imaged ostrac ods sized between 0.6 and 0.9 mm ESD. Net collected ostracods were exclusively comprised of the species Euconchoecia chierchiae. Figure 43. Distribution and abundance of calanoid copepod imag es collected during this study. Classification accuracy for this group was 82%. Small thimble-shaped trachymedusae, consisting main ly of Aglaura hemistoma were the next most abundant zooplankton group imaged by SIPP ER and comprised greater than 7% of the total imaged particle assemblage. These small hydr omedusae were common at all depths with concentrations ranging from 156-1316 m-3, with the greatest concentrations found below 40 m (fig. 46). Many imaged Aglaura were observed with their tentacles contracted tigh tly against the bell suggesting a possible behavioral reaction to b eing sampled, but a significant minority was observed with their tentacles fully extended in a f eeding posture in which they wag the very distal end of their tentacles (Colin et al, 2005, see exam ple image in figure 25). However, very few imaged specimens had any visible captured prey. Thi s behavior contributed to the broad size range of this class, as witnessed by ESD measuremen ts since tentacle area was included in the measurement of ESD. Manual analysis of the Aglaura images revealed a wide range of bell
81 diameters from approximately 0.75 mm to 2.5 mm, wit h very few larger individuals. No identifiable Aglaura specimens were collected in the nets. Figure 44. Size distribution in ESD for the crustac ean zooplankton image classes. The distribution of the cyclopoid copepod Oithona generally followed the trend of the other abundant zooplankton groups with highest abun dances encountered below 40 m (figure 47) High abundances however, were also found at interme diate depths during the latter half of the day. Abundances ranged from greater than 1300 ind. m-3 at 45 m to less than 50 ind. m-3 at 3 m depth. Oithona were small, with 85% of the imaged organisms less than 0.8 mm ESD. Net analysis indicated most of the small Oithona were i dentified as O. colcarva while the larger specimens were O. plumifera
82 Figure 45 Distribution and abundance of ostracod images colle cted during this study. Classification accuracy for this group was 80%. Chaetognaths made up between 4 and 11% of the zoopl ankton assemblage at any one depth. Abundances were fairly uniform (50-200 m-3) between 3-30 m deep while highest values were encountered below 40 m depth where abundances ranged between 494-726 ind. m-3 (figure 48). Imaged chaetognaths spanned a wide size as ind icated by their ESD size distribution (fig. 42). Average chaetognath size was slightly greater than 1.5 mm ESD, which was equivalent to approximately 4.8 mm in length. Measured chaetognat h lengths ranged from less than 1.5 mm in to over 25 mm, but the majority (~90 %) of the indi viduals were between 2.5 Â– 10 mm. Several apparent behaviors were visible in chaetognath imag es including cannibalism, fecal pellet production, parasitism and predation (see figure 24 ). Larvaceans were the most commonly imaged apparent prey item in chaetognath feeding im ages (>70 occurrences). Plutei from different echinoderm groups were the mo st abundant meroplankton group imaged by SIPPER (2% of the total extracted images) and were common but rarely abundant in the water column during this study (fig. 49). Half of the SIPPER samples had concentrations of 50 individuals m-3 or less and only 6 had concentrations greater than 300 individuals m-3. Highest concentrations were found in the middle of the wate r column between 15 and 35 m depth and
83 during the late afternoon and early evening. Both echinopluteus and ophiopluteus forms were imaged by SIPPER. Classified plutei had an average ESD of 1.05 mm and ranged in size between 0.6 and 2.5 mm ESD. The observed distributi on and the large size range was most likely influenced by the large number of sarcodine protoct ists that were misclassified as echinoderm plutei (see table 9). Figure 46. Distribution and abundance of Aglaura hemistoma images collected during this study. Classification accuracy for this group was 77%. A diverse group of hydromedusae and narcomedusae ma de up the other cnidaria class (fig. 26). No one species or morphology was abundan t enough to warrant their own class but as a whole the group comprised between 0.2 to 9% of the imaged zooplankton assemblage at any one depth. Highest abundances were encountered below 30 m but there was also relatively high concentrations observed in the afternoon and early evening between 5-30 m (fig. 50). This group had the widest and most broad size range with 50% o f the imaged animals having ESDs greater than 2.5 mm. Bell diameters ranged from 1 mm to ove r 2 cm. Including tentacle length, some of the larger medusae had overall lengths greater than 10 cm. At least 19 separate hydroand narcomedusan species were observed during this stud y. This group had the lowest classification
84 accuracy at 46% and also was strongly confused with both Aglaura hemistoma and the other and unknown group suggesting that this distribution sho uld be viewed with caution. Figure 47. Distribution and abundance of Oithona sp. images collected during this study. Classification accuracy for this group was 82%. Figure 48. Distribution and abundance of chaetognat h images collected during this study. Classification accuracy for this group was 82%.
85 The poecilostomatoid copepod Oncaea was a relatively minor but consistent contributor to the SIPPER imaged zooplankton assemblage contrib uting between 2-4% to the total at all depths. Oncaea is a robust and compact copepod with short antenna making it relatively difficult to classify and leading to a classification accurac y of just 52%. The majority (90%) of the imaged Oncaea were less than 0.8 mm ESD. Oncaea distribution trended with most of the other zooplankton groups, with highest abundances below t he pycnocline (figure 51). Imaged Oncaea were comprised mostly of the species O. venusta and O. mediterranea both of which were common in the net samples and within the correct si ze range. Cladocerans were an abundant but narrowly distribut ed taxa in the SIPPER samples. The cladoceran images were comprised solely of Penilia avirostris and they were found only in detectable numbers at the deepest sampled depth fro m each deployment at abundances between 328-1258 ind. m-3. Cladoceran images were small with an average ESD of 0.8 mm ESD. Even though they were only counted in 7 of 70 SIPPER samples, cladocerans made up almost 1.5% of the total extracted image particle a ssemblage. They contributed between 4-16% to the total zooplankton assemblage in those 7 samp les. The eumalacostracan crustacean class was relatively rare and patchily distributed throughout the water column (fig. 52). Concentratio ns exceeding 100 ind. m-3 were encountered only 5 times out of 70 samples. All 5 instances occ urred between 30 and 45 m depth. However, at all depths there were instances of relatively lo w occurrences (<25 individuals m-3). The size distribution of the eumalacostracan classified imag es was very broad, with 78% of the images having a size greater than 1 mm ESD and 31% were la rger than 1.5 mm ESD. Amphipods and calyptopis and furcilia stages of euphausiids, zoea l and megalopa stages of decapods comprised the majority of sub 1.5 mm ESD images while Stomato pod larvae, juvenile Lucifer faxoni and porcellanid crab zoea dominated the larger size ran ge. As with the cnidaria other class, this morphologically diverse class had one of the lowest classification accuracies at 49%. Doliolids were distributed throughout the water col umn with highest concentrations encountered between 30 and 45 m (fig. 53). The majo rity of imaged doliolids were of gonozooid and phorozooid morphological stages and less numero us oozoid and old nurse stages. Only a
86 few trophozooid stages were imaged. Based on the lo cation of the endostyle on the oozoid images between muscle band 3 and 5 (Godeaux, 1998), most of the doliolids were most likely Dolioletta sp. While the classification accuracy was relative ly high at 67%, the C/H ratio was the highest of the 21 classified groups at 2.62 meaning there was significant confusion with other groups. This led to the computer classification of more Aglaura hemistoma images, for example as doliolids than actual imaged doliolids (table 9) The poecilostomatoid copepod Macrosetella gracilis was the least abundant zooplankton group included in the final classifier. This specie s comprised between less than 1 to slightly fewer than 5% of the total imaged zooplankton assemblage, averaging 1.4% for the 70 SIPPER samples. Highest abundances were recorded below the pycnocline throughout most of the day except for during the afternoon and early evening, when lower abundances were found in deep water and abundances were elevated near the surface (fig. 54). This group had the highest classification accuracy of any zooplankton class in the final classifier at 83.7% although there was a problem with it being confused with other copepod groups that contributed to having a C/H ratio higher than 2.
87 Figure 49. Distribution and abundance of echinoderm plutei including both echinopluteus and ophiopluteus images collected during this study. Classification accuracy for this group was 62%. Figure 50 Distribution and abundance of unidentified hydromed usae and narcomedusae making up the other cnidarian class. Classification accura cy of this group was 46%.
88 Figure 51. Distribution and abundance of the poecilostomatoid copepod genus Oncaea images collected during this study. Classification accurac y of this group was 52%. Figure 52. Distribution and abundance of eumalocostracan crust acean images collected during this study. Classification accuracy of this group w as 49%.
89 Figure 53. Distribution and abundance of doliolid i mages collected during this study. Classification accuracy of this group was 67%. Figure 54 Distribution and abundance of the poecilostomatoid copepod Macrosetella gracilis images collected during this study. Classification accuracy of this group was 49%.
90 Trichodesmium Both morphologies of the diazotrophic colonial cyan obacteria Trichodesmium sp. were abundant throughout the water column. These two mo rphologies did not appear to represent separate species of Trichodesmium even though there are two species commonly collect ed on the WFS. Trichodesmium erythraeum and T. thiebaudii can be common on the WFS, but in this instance most of the images appeared to be T. thiebaudii (Judy OÂ’Neil, pers. comm.). Abundances of both groups followed the same distrib ution pattern, so their distribution and abundance data were combined and are presented in f igure 55. The combined classification accuracy for the two groups was 74%. Trichodesmium distribution contained a possibly strong diel signal, with highest abundances encountered in the morning and early afternoon in the upper 30 m with an abrupt decrease in surface abundance a nd increased abundance below 30 m during the evening. This signal was not evident in either the fluorescence or extracted chlorophyll -a profiles collected at the same time. Tuft and puff shaped colonies were larger (average size 1.32 mm ESD) than the elongated linea r colonies (0.85 mm ESD). Individual trichomes could be observed in both colony morpholo gies. Protoctists Foraminifera, acantharians and solitary radiolarian s made up the sarcodine protoctist class that comprised just over 3% of the total imag ed particle assemblage. Colonial spumellarian Radiolaria were noted in the samples especially nea r the surface, but were too rare and morphologically distinct from the rest of the sarco dine protoctist class to be included in the classifier. Most of the imaged sarcodines were rad ially symmetrical with varying lengths of extendible pseudopods (figure 35). While present t hroughout the water column in appreciable numbers, sarcodines formed a dense band between 3040 m depth at concentrations up to 1000 ind. m-3 (fig. 56). This band was comprised mainly of large radiolarians resembling Thalassicolla nucleata or a similar species while the other sampled depth s were comprised of a smaller and more diverse protoctist assemblage. Most imaged sa rcodines had exoskeletons less than 1 mm in diameter, but many had numerous pseudopodia exte nding far past their body. These larger forms were often associated with marine snow and un identifiable small particles that appeared to
91 be stuck onto the pseudopodia that may be indicativ e of feeding by these forms. This contributed to the broad size range apparent in figure 23 and t he average size of 1.21 mm ESD for sarcodines. Large elongate colonies of diatoms and dinoflagella tes comprising slightly greater than 2% of the imaged particle assemblage made up the el ongate phytoplankton class. Most of these colonies were small, with 95% of the imaged populat ion less than 1 mm ESD. The colonies tended to be narrow and long with 95% of those imag ed being less than 4 mm in total length. Highest concentrations were found at or below the p ycnocline for most of the day except for the afternoon and early evening when the distribution r eversed and highest concentrations were found near the surface and lower concentrations wer e found below the pycnocline (fig. 57). As with the Trichodesmium distribution, this trend was also not evident in e ither the fluorescence or extracted chlorophylla profiles. Figure 55 Distribution and abundance of Trichodesmium colony images collected during this study. Classification accuracy for this group was 7 4%.
92 Marine snow Marine snow comprised slightly less than 10% of the total imaged particle assemblage. This class was comprised of a wide range of materia l and morphologies ranging from identifiable fecal pellets, shed larvacean houses, diatom frustr ules, and protoctist tests to unidentifiable strands and detrital aggregates. Marine snow images were often labeled as larvacean images (12% false positive rate in the test set) due to th e presence of the morphologically similar house structure in many of the larvacean images used in t he training library. Marine snow distribution roughly followed the zooplankton distribution, with greatest concentrations found below the pycnocline and increased numbers in the upper 30 m during the afternoon and early evening suggesting the marine snow formation is mostly auto chthonous (fig. 58). However, the proportion of marine snow to the total particle assemblage sho wed no relation to depth. Most marine snow particles were small with 84% of the imaged particl es less than 1 mm ESD. While not a common occurrence, both Oncaea sp. and Macrosetella gracilis were imaged associated with marine snow particles. Small (<0.7 mm in length) opaque particl es were found on many larger marine snow particles that may have been smaller unidentified z ooplankton feeding or otherwise associating with these aggregates.
93 Figure 56. Distribution and abundance of sarcodine protoctist images collected during this study. Classification accuracy for this group was 72%. Figure 57. Distribution and abundance of the elonga te phytoplankton images collected during this study. Classification accuracy for this group was 6 8%.
94 Figure 58. Distribution and abundance of marine sno w images collected during this study. Classification accuracy for this group was 61%. Net Sample Analysis and Comparison with Concurrent SIPPER Data One hundred of the smallest individual images from each of the thirteen zooplankton classes manually classified in the test set were me asured to determine the appropriate minimum size of net collected zooplankton to compare agains t the SIPPER sampling effort. Total length was measured for the four copepod classes, chaetogn aths, cladocerans, doliolids, eumalacostracan crustaceans, echinoderm plutei and ostracods; head length for larvaceans, and bell diameter for the hydromedusae (Table 11). The counts for both Aglaura hemistoma and other cnidarians were combined in the net counts because identification to species for small hydromedusae and narcomedusae was often not possibl e. The number of net zooplankton above the minimum SIPPER size threshold for each class wa s then calculated by taking the measured size distribution for each zooplankton class and mu ltiplying the proportion of individuals greater than that size by the total number of zooplankton i n that class. This was done for all 12 zooplankton groups for the 19 net samples.
95 Table 11. Average minimum size of net collected zoo plankton used for comparison of sampling effort between plankton nets and the SIPPER. Standa rd deviation in parentheses. Zooplankton class Average Minimum net plankton size Chaetognath 1.4 mm (.12) Total Length (TL) Cladoceran 0.78 mm (0.047) TL CnidariaAglaura and other 0.63 mm (0.08) Bell Diameter (B.D.) Copepod-calanoid 0.75 mm (0.02) TL CopepodMacrosetella 1.02 mm (0.02) TL CopepodOithona 0.82 mm (0.04) TL CopepodOncaea 0.79 mm (0.03) TL Echinoderm plutei 0.76 mm (0.08) TL Eumalacostracan crustaceans 0.96 mm (0.12) TL Gelatinous-doliolid 0.80 mm (0.02) TL Larvacean 0.32 mm (.01) Head Length (H.L.) Ostracod 0.68 mm (.01) T.L. Net and SIPPER total zooplankton counts were then c ompared using paired t-tests. The SIPPER imaged between 9-15% more similar-sized zoop lankton than the nets, but the differences were only significant for the nighttime comparison (Table 12). These differences were caused by a combination of the SIPPER classifying l ess small crustacean zooplankton and the nets collecting less fragile zooplankton. The nets collected significantly more similar-sized calanoid Oithona and Oncaea species of copepods, cladocerans, and ostracods al though the Oithona and ostracod abundances were only significantly di fferent during the nighttime comparison. Nets counts for these groups were betwe en 19-212% greater than the concurrent SIPPER estimates. The copepod Macrosetella gracilis chaetognaths, and the eumalacostracan crustaceans were all sampled similarly by the two c ollection methods. The cnidarian hydromedusae and narcomedusae, echinoderm plutei, d oliolids and larvaceans were all significantly overrepresented in the SIPPER samples compared to the concurrently sampling nets. The SIPPER abundance estimates for these fra gile zooplankton groups were 85-1700% greater than the concurrent net estimates.
96 Fine Scale Distribution of Plankton and Particle Cl asses Tables 13-16 show the results from the Monte Carlo simulations and the resulting U2 statistics and Z-scores for the four SIPPER samples in the test set. The two daytime samples and the shallow nighttime imaged particle assemblages ( minus the two artifact classes) had distributions that were significantly different fro m complete spatial randomness. Compared against the results from the random simulations, mo st of the differences with the real data occurred at nearest neighbor differences less than 1 cm. On average, 37% of the observed particles had NND of 1 cm or less while only 32% of the randomly distributed simulated particles had NND of 1 cm or less, indicating an aggregated p article distribution. In contrast, the LloydÂ’s index of patchiness for all 4 particle assemblages measured between 1.01 and 1.14 indicating no difference with a random distribution at the meter scale. At the meter scale, abundances ranged from 0.5-5 the average abundance along the samplin g transect. NND analysis of the individual classes in the SIPPE R dataset indicated that approximately one quarter of the image classes had non-random distributions based on the ANND measurements. However, there was not 100% agre ement in detecting these non-random distributions between the two methods used. While t he U2 statistic predicted significant departures from spatial randomness in 24 of 76 clas ses, the Z-score predicted that only 17 classes had non-random distributions. The two metri cs both predicted non-random distributions for 14 of the image classes. All of the image class es predicted by the U2 statistic to be nonrandom were found to be more aggregated than could be explained by chance while 13 of the 14 classes indicated by the Z-score to be non-randomly distributed were aggregated and one was found to be more uniformly distributed (the other a nd unidentified class in the shallow night sample). The LloydÂ’s index of patchiness values we re far more conservative, with only six nonartifact classes having index values greater than 1 .4 and only two having index values greater than 2. This suggested that the plankton classes im aged by SIPPER-2 did not form appreciable aggregations in the horizontal domain at the meter scale. Bubbles were an artifact class and were only sample d because the HRS sampled too close to the surface on occasion and sampled the sh ips wake. Inspection of the spatial
97 distribution of the bubbles indicated they occurred infrequently but at high abundances indicating clustering. We used this group as a test to determi ne the power of the spatial statistics in detecting significantly non-random particle distrib utions. While the observed U2 ranks for the bubbles were both higher than all 1000 simulations, the Z-score for the nighttime sample was only -0.74 indicating a random pattern. The LloydÂ’ s index of patchiness for bubbles in the two surface samples was 9.45 and 6.33 indicating signif icant patchiness at the meter scale. The patchy nature of the nighttime bubble distribution is apparent in figure 59 in the NND cumulative histogram and a plot of bubble abundance per meter traveled during the sampling transect. Bubbles were not included in the calculation of ANN D and LloydÂ’s index of patchiness for the entire particle assemblage to ensure their extreme clustering did not bias the results. Marine snow was the only other group observed to be non-randomly distributed in all of the samples analyzed in which they occurred. Marine snow particles had U2 rankings greater than 100% of the simulated data for each of the four tow -transects analyzed. All of the distributions were aggregated, with the observed cumulative histo grams having between 5-11% more NND less than 5 cm than the simulated populations. The Z-score predicted similar results for all but the shallow nighttime marine snow distribution. The Z-s core for the deep daytime marine snow distribution of -13.1 was the lowest and most signi ficant Z-score recorded. The ANND measurements did not translate into strong meter-sc ale patchiness as the LloydÂ’s values ranged from 0.97 to 1.26. This ANND clustering could be du e to the coagulative properties of marine snow that tend to cluster other detritus around it (Kirboe et al., 1990), the distribution of the organisms responsible for the formation of the detr itus (Hansen et al. 1996), or it could be an artifact from the PICES image segmentation and extr action process. Large marine snow aggregates are often held together with transparent exopolymer particles or TEP (Alldredge et al., 1998) that might not be imaged by the 3-bit graysca le of the SIPPER-2. Thus, a single marine snow particle might be extracted as more than one a nd therefore bias the measurement of marine snow distribution towards clustering. Additi onally, these large aggregates are known to harbor large numbers of copepods, amphipods and oth er organisms (Steinberg et al., 1994; Koski et al., 2007) that may not be counted when th ese large particles are extracted.
98 Most of the other clustered particle distributions were found in the two deep water samples where approximately 75% of the clustered cl asses were found using either statistical metric. During the day, chaetognaths, Aglaura hemis toma, calanoid copepods, Oithona sp., echinoderm plutei, larvaceans, marine snow, ostraco ds and the other and unknown particle class were all found to be significantly aggregated based on their U2 ranking. Results from the Zscores were nearly identical, with significant clus tering found for each of the same groups except the echinoderm plutei class that was instead calcul ated to be randomly distributed. During the night, chaetognaths, cladocerans, Aglaura hemistoma Oithona, echinoderm plutei, elongate phytoplankton chains and marine snow were all found to be clustered based on their U2 rank. The Z-scores only indicated significant clustering for the distributions of chaetognaths, cladocerans, echinoderm plutei and marine snow while suggesting that the other and unidentified class was regularly distributed. While many individuals of th e imaged plankton groups were clustered, the degree of clustering along the sample transects wer e not very dynamic. LloydÂ’s index of patchiness only exceeded 1.5 for two groups of imag ed copepods sampled during the day, Macrosetella gracilis and Oithona Both the ANND clustering and meter-scale patchin ess of the imaged daytime Oithona assemblage is depicted in figure 50. Notice the mo re than 100 meter long zone of depressed Oithona abundance relative t o the rest of the transect that probably led to the relatively high LloydÂ’s index. For the other de epwater groups, there was not much meter-scale patchiness apparent. Distributions of individual particle classes were p redominantly randomly distributed in the two shallow water samples using the ANND measuremen ts. The other and unidentified particle class was the only particle class besides marine sn ow that was found to be clustered during the day while Trichodesmium tuft and puff colonies were the only other class f ound to be clustered during the nighttime samples using the U2 statistic. Results using the Z-score also showed few particle classes as being clustered, with only the copepod genera Oncaea and sarcodine protoctists with Z scores less than -2 during the d ay and none besides marine snow during the night. Only the other and unidentified class sample d during the daytime had a LloydÂ’s index of patchiness value greater than 2 suggesting most of the shallow water plankton groups were
99 randomly distributed at the meter-scale. A closer l ook at the meter-scale distribution of the other and unidentified class sampled during the daytime i ndicated that it was only a high aberrant number of particles encountered in one meter that t ilted the distributions towards a patchy distribution. This suggests the LloydÂ’s index of pa tchiness may be susceptible to overestimates of patchiness. The ANND calculations of the PICES classified test set were also investigated. PICES correctly identified significant clustering in 20 o f the 24 classes in which it was found in the manually classified test set using U2 statistics. However, 4 other classes that were not considered significantly clustered in the manual classified da taset were determined to be clustered using the PICES U2 results. Similarly, PICES correctly identified clu stering in 16 of the 21 classes that were found to be significantly clustered manually using Z-scores. Three image classes not considered clustered using Z-scores from the manually classifi ed test set were scored as significantly clustered when classified by PICES. LloydÂ’s indice s of patchiness were not calculated for the PICES classified dataset.
100 Table 12. Comparison of SIPPER zooplankton abundanc e estimates versus the abundance estimates of similarly sized zooplankton collected by concurrently sampling 162 m m plankton nets. Differences in bold indicate significant diff erences between the two sampling strategies (paired t-test, p <0.05).
101 Table 13. Results of statistical analysis of spatia l patterns in the shallow daytime test set sample. Numbers in bold indicate departures from a random d istribution for that class for either the U2 statistic or the Z-score. indicates the PICES cla ssification failed to detect significant deviations from randomness while indicates that the PICES cl assification detected a significant departure from randomness for that class not detected in the manually classified set. Day Shallow Abundance (# m-3) Median NND (cm) U2 statistic rank Zscore LloydÂ’s Index All particles (minus artifacts) 3711 2.2 >100 % of 1000 simulations -3.08 1.14 Bubbles 1193 1.0 > 100 % -30.36 9.45 Chaetognath 77 48.4 > 19.8 % -0.27 1.21 Aglaura hemistoma 282 15.3 > 3.20 % -0.99 1.04 Other hydromedusae 46 96.0 > 14.8 % -0.92 0.78 Calanoid copepods 383 11.5 > 51.4 % -0.77 1.09 Macrosetella gracilis 3 1717.2 > 1.02 % 0.87 N.A. Oithona sp. 60 75.9 > 17.6 % 0.12 0.90 Oncaea sp. 18 104.6 > 58.8 -2.00* N.A. Doliolids 12 222.5 > 92.5 % -1.22 N.A. Echinoderm plutei 11 237.7 > 92.7 % -1.39 N.A. Elongate phytoplankton chains 74 51.2 > 7.5 % -0.97 0.91 Eumalacostracan crustaceans 12 274.6 > 93.0 % -0.31 NA Larvaceans 337 13.2 > 64.7 % -0.89 1.00 Marine snow 141 31.4 > 100 %* -0.70 0.97 Ostracod 4 1285.3 > 0.06 % 0.04 N.A. Other and unidentified 759 7.0 > 100 % -0.53 2.32 Sarcodine protoctists 118 30.8 > 81.7 % 2.88* 1.04 Trichodesmium colonies-elongate 583 8.5 > 80.0 % -0.57 1.07 Trichodesmium colonies-tuft and puffs 791 7.1 > 83.6 % -1.17 1.05
102 Table 14. Results of statistical analysis of spatia l patterns in the deep daytime test set sample. Numbers in bold indicate departures from a random d istribution for that class for either the U2 statistic or the Z-score. indicates the PICES cla ssification failed to detect significant deviations from randomness while indicates that the PICES cl assification detected a significant departure from randomness for that class not detected in the manually classified set. Day Deep Abundance (# m-3) Median NND (cm) U2 statistic rank Zscore LloydÂ’s Index All particles (minus artifacts) 12779 1.4 >100 % of 1000 simulations -12.05 1.04 Chaetognath 447 8.6 > 95.8 % -2.19 1.19 Cladoceran 252 13.6 > 48.0% -1.66 1.35 Aglaura hemistoma 1155 5.3 > 100 % -3.08 1.23 Other hydromedusae 62 54.0 > 16.5 % -0.73 1.33 Calanoid copepods 1607 4.6 > 96.4 % -2.27 1.11 Macrosetella gracilis 52 43.0 > 60.6 % 0.56 2.12 Oithona sp. 637 6.2 > 100 % -3.34 1.71 Oncaea sp. 275 13.3 > 47.8 % -1.11 1.19 Doliolids 26 125.9 > 80.0 % -1.60 N.A. Echinoderm plutei 12 256.8 > 96.5 %* -0.65 N.A. Elongate phytoplankton chains 129 27.3 > 11.8 % -0.20 1.06 Eumalacostracan crustaceans 43 74.9 > 76.6 % -1.14 1.10 Larvaceans 2552 3.5 > 100 % -4.96 1.16 Marine snow 1610 3.9 > 100 % -13.10 1.26 Ostracods 973 5.5 < 100 % -4.23 1.32 Other and unidentified particles 2715 3.5 > 100 % -2.02 1.12 Sarcodine protoctists 91 40.9 > 44.8 % 0.40 1.41 Trichodesmium colonies-elongate 78 41.7 > 94.4 % 0.02 0.95 Trichodesmium colonies-tuft and puffs 64 56.5 > 91.2 % -0.70 0.97
103 Table 15. Results of statistical analysis of spatia l patterns in the shallow nighttime test set sample. Numbers in bold indicate departures from a random distribution for that class for either the U2 statistic or the Z-score. indicates the PICES cl assification failed to detect significant deviations from randomness while indicates that t he PICES classification detected a significant departure from randomness for that class not detect ed in the manually classified set. Night Shallow Abundance (# m-3) Median NND (cm) U2 statistic rank Zscore LloydÂ’s Index All particles (minus artifacts) 2760 3.0 96.1 % of 1000 simulations -3.32 1.01 Bubbles 167 2.7 > 100 % -0.74 6.33 Chaetognath 39 74.0 > 37.7 % -0.82 N.A. Aglaura hemistoma 214 18.1 > 52.2 % 0.15 1.00 Other Hydromedusae 35 118.2 > 14.2 % 0.99 N.A. Calanoid copepods 381 10.5 > 58.1 % 0.39 1.00 Macrosetella gracilis 4 814.2 > 58.3 % -1.10 N.A. Oithona sp. 75 57.1 > 92.0 % 0.65 1.01 Oncaea sp. 34 139.5 > 81.8 % 0.34 N.A. Doliolids 10 346.1 > 9.1 % 0.41 N.A. Echinoderm plutei 18 294.3 > 23.7 % 0.75 N.A. Elongate phytoplankton chains 18 194.7 > 72.8 % -0.11 N.A. Eumalacostracan crustaceans 23 164.0 > 24.9 % -0.13 N.A. Larvacean 523 9.3 > 72.1 % 0.34 1.02 Marine Snow 503 8.9 > 100 % -2.24* 1.00 Ostracod 73 39.2 > 52.9 % -0.50 1.36 Other and unidentified 405 10.2 > 62 % -1.43 1.04 Sarcodine protoctists 104 27.4 > 0.9 % -1.30 1.18 Trichodesmium colonies-elongate 87 41.6 > 68.9 % -0.04 1.19 Trichodesmium colonies-tuft and puffs 215 18.1 > 98.5 % 0.00 1.02
104 Table 16. Results of statistical analysis of spatia l patterns in the deep nighttime test set sample. Numbers in bold indicate departures from a random d istribution for that class for either the U2 statistic or the Z-score. indicates the PICES cla ssification failed to detect significant deviations from randomness while indicates that the PICES cl assification detected a significant departure from randomness for that class not detected in the manually classified set. Night Deep Abundance (# m-3) Median NND (cm) U2 statistic rank Zscore LloydÂ’s Index All particles (minus artifacts) 10112 1.7 >82.9 % of 1000 simulations -1.04 1.01 Chaetognath 774 8.6 > 96.9 % -3.17 1.06 Cladoceran 1366 13.6 > 96.2 % -4.54 1.12 Aglaura hemistoma 1032 5.3 > 96.4* % 0.15 1.04 Other hydromedusae 82 54.0 > 16.5 % 0.67 1.08 Calanoid copepods 780 4.6 > 88.0% -1.21 1.04 Macrosetella gracilis 44 43.0 > 61.4 % -0.42 1.12 Oithona sp. 1231 6.2 > 100 % -1.29 1.10 Oncaea sp. 209 13.3 > 38.6 % 0.25 1.01 Doliolids 46 125.9 > 74.8 % 0.14 1.38 Echinoderm plutei 31 256.8 > 99.6* % -2.03* N.A. Elongate phytoplankton chains 138 27.3 > 97.5 % 4.07 1.16 Eumalacostracan crustaceans 20 74.9 > 94.2 % 0.12 N.A. Larvaceans 798 3.5 > 16.4 % -0.85 1.03 Marine snow 660 3.9 > 100 % -3.38 1.08 Ostracod 1190 5.5 > 73.5% 0.14 1.04 Other and unidentified 1304 3.5 > 77.0 % 2.56* 1.02 Sarcodine Protoctists 190 40.9 > 91.7 % -0.46 0.98 Trichodesmium colonies-elongate 97 41.7 > 34.7 % -0.88 1.10 Trichodesmium colonies-tuft and puffs 121 56.5 > 89 % -0.89 0.96
105 Figure 59. A: ANND cumulative histogram of the obse rved shallow daytime bubble class compared to the 1000 random distributions. B: The m eter-scale distribution of observed bubbles. Figure 60. A: ANND cumulative histogram of the obs erved deep daytime Oithona class compared to the 1000 random distributions. B: The m eter-scale distribution of observed Oithona sp. of copepods.
106 Discussion This SIPPER-2 dataset is the most extensive datase t tested yet with PICES. Earlier work with SIPPER images detailed the initia l beginnings of the classifier development (Luo et al., 2004) and methods to optimize the construct ion of a training library (Luo et al., 2005). These experiments utilized between 5 and 6 image cl asses and less than 10,000 total images, whereas this dataset had a manually classified test -set of 109122 images and ultimately 21 separate image classes. The resulting classifier wa s then used to classify nearly one and half million unidentified particle images collected duri ng a single day of sampling. To put this in perspective, the magnitude of this classification e ffort can be compared to previously published classification efforts by the VPR, considered to be the preeminent imaging system available. The largest described VPR dataset consisted of over 200 ,000 particle images collected during 240 hours of VPR sampling in the Japan East Sea and cla ssified into 5-8 separate plankton groups (Ashjian et. al. 2005) although the VPR can collect between 104-106 images per day when sampling continuously (Hu et al., 2005). Classifier Performance and Comparison with Other Fi eld Deployed Classifiers Compared to most other plankton imaging sensors, th e SIPPER-2 samples a relatively large volume of water in a short period of time and allows for representatively sampling even rare (<1% of assemblage) organisms and more robust compa risons with net systems. This can be demonstrated by comparing the water volume sampled during a 5 minute tow at 1-1.5 ms-1. A 1m2 net tow would sample 200 m-3 while the high-resolution camera on a VPR would onl y sample 0.005 m-3 (Broughton and Lough, 2006) and the SIPPER-2 would sample 2.13 m-3. Therefore the VPR is only sampling .0025 % the volume of that of a net sampling the same plank ton population whereas the SIPPER is sampling 1%. This limits the VPR to accurately quantify organisms that only occur at concentrations greater than ~500 ind. m-3 (Davis et al., 1992) while SIPPER can quantify organisms that occur at concent rations less than 10 ind. m-3 and make up less than 1% of the total assemblage. This can be i mportant in subtropical environments like the Gulf of Mexico where the zooplankton populations ar e diverse (Hopkins 1981; Ortner et al., 1989;
107 Sutton et al., 2001) and many groups occur at detec table limits. To adequately describe this system, PICES had to incorporate far more plankton classes than that previously attempted for in-situ collections of plankton and particle images Even with this modification, our classification accuracy of 74% compared well with that of other cl assification and imaging systems. For example, Hu et al. 2005, reported a classification accuracy of 72% using a 7 class SVM on VPR data collected in Georges Bank. Identifying preserv ed net samples with the plankton scanner Zooscan, researchers using a discriminant vector fo rest classification algorithm achieved 75% accuracy while classifying the net samples into 29 separate groups (Grosjean et al., 2004). Limitations of the SIPPER Imaging System While the performance of PICES was similar to that from other plankton classifiers, the diversity of the particle assemblage encountered on the WFS revealed many limitations of the PICES automated classification system, the SIPPER-2 and imaging systems in general. The deeper waters of the WFS were populated by multiple crustacean taxa that were superficially similar at the resolution of SIPPER-2. Oncaea sp., small calanoid copepods, cladocerans and ostracods all were encountered in high concentratio ns below the pycnocline. There was much confusion between some of these groups. This could be alleviated by increasing the optical resolution of the SIPPER to allow better discrimina tion of these groups but at the expense of reducing the field of view. A doubling of the SIPPE R optical resolution to 35 m would reduce the SIPPER sampling area by 75%. This would affect its ability to sample larger and rarer particle classes. The relatively large size of the sampling area also presented problems as we were able to often image larvaceans within their mucous house The larvacean was thus often difficult to separate from the surrounding material of the house especially the pre-filter, in the extraction phase. This made confusion with the marine snow cla ss a particular problem, with 15% of larvaceans imaged with some associated debris settl ed on the house or pre-filter leading to between 5-12% of larvaceans and marine snow classif ied incorrectly. Similarly it was difficult to tell the difference between single trichomes and ve ry small linear shaped Trichodesmium colonies from elongate diatom chains. This has been noted as well in the VPR (Davis and McGillicuddy, 2006). The capability exists to incor porate a color line scan camera into the
108 SIPPER that might assist in better discriminating t hese groups by allowing us to separate groups by color and adding a new complement of color featu res to the PICES classifier. Additionally, as we begin building more SIPPER units, multiple syste ms could be deployed with different optical resolutions, focal depth and sample mouth area to m ore adequately cover the entire net-plankton size spectrum of given area, akin to net sampling w ith a variety of net mesh sizes to better capture the full size range of selected taxa (Galli enne and Robins, 2001; Hopcroft 2001) Human Classification Error Previous work documenting the development of the pl ankton classification system of the VPR operated under the assumption that the human ex pert classifying the training library was a Â‘perfect classifierÂ’ (Hu and Davis, 2005). However, itÂ’s since been demonstrated that human experts are not reliable classifiers of complex ima ge datasets as they have short-term memory limits, are affected by recency effects where new c lassifications are biased towards recently used labels and positivity bias, where classification of specimens are biased towards what should be expected in a sample (Culverhouse et al., 2006). I n a study by Culverhouse et al. (2003), 16 marine ecologists and harmful algal bloom specialis ts were asked to classify images of 6 species of dinoflagellates common to their area of study. T hese images were initially labeled by two separate taxonomists familiar with this plankton gr oup using the 2-expert protocol (Culverhouse et al., 1996). The 16 scientists achieved between 6 7-82% self-consistency and 43% consensus between experts for this complex labeling task (Cul verhouse et al., 2003). While a test on this scale was not possible with our dataset, I did have a second marine ecologist classify a subset of over 2000 SIPPER images from the test-set. We achie ved an 81% consensus in labeling these images with most of the dispute occurring amongst t hree groups, the copepod Macrosetella gracilis echinoderm plutei and the other and unidentified class. Most of the labeled groups were made up of rather broad taxonomic groupings, probab ly making the consensus between human experts easier than would be expected with a classi fier attempting to label into more specific taxa (species, sex or life stage). Therefore, as machine classifiers and in-situ imaging systems become more capable, these internal checks of the h uman experts responsible for training the classifier should become more common and rigorous. This holds true as well for plankton net
109 datasets from large research projects and time-seri es that are analyzed by more than one taxonomist where such error could propagate through the data if it was not accounted for. Comparison with Plankton Net Data Comparisons of manually determined abundance estima tes between plankton imaging systems and nets sampling in the same area have typ ically demonstrated that imaging systems can sample the crustacean and pteropod fraction as well as nets (Benfield et al., 1996; Remsen et al., 2004; Broughton and Lough, 2006) while outp erforming nets in collecting fragile and gelatinous taxa (Benfield et al., 1996; Villareal e t al., 1999; Ashjian et al., 1999; Remsen et al. 2004). This is especially true when measures are t aken to compare only that fraction where the sampling efficiencies of the two methods overlap (B roughton and Lough, 2006). In an earlier study with black and white SIPPER images, Remsen et al. 2004 compared abundances of SIPPER imaged plankton classes greater than 250 m ESD with concurrently collected 162 m m net samples. The majority of the smaller SIPPER ima ges were unidentifiable due to the lack of resolvable features, leading to the inability to cl assify 67% of the total SIPPER dataset. Although the differences were not significant, plankton net abundances for copepods were nearly twice as high as the SIPPER estimates and were probably caus ed by the inability to identify the smaller copepods collected by the nets. This leads to imagi ng systems having a capability to detect and image smaller organisms than an investigator can po sitively identify due to lack of resolvable features. Attempting to correct for this, Broughton and Lough, 2006 modified a VPR and MOCNESS dataset due to organism size, abundance, an d fragility so that only similarly sampled taxa would be compared. These measures were only pa rtly successful: while they found that the two systems were in close agreement on the contribu tion of different copepod taxa to the total copepod assemblage, the VPR produced standardized c opepod abundances that were twice as high as the nets. They determined this could be due to in incorrect calculation of the VPR field of view and consequent sample volume. This is not a pr oblem for the SIPPER-2 where the field of view has been determined in the lab and is bounded by the sample tube. To be capable of
110 resolving the smaller zooplankton collected by a 16 2 m m net, its field of view and corresponding sampling tube would need to be significantly reduce d. Ecological Implications for the WFS Previous studies of the zooplankton assemblage on t he WFS and surrounding waters have described the system as crustacean dominated ( Hopkins et al., 1981; Sutton et al., 2001), mostly by small sub-millimeter forms of Oncaea Oithona the Paracalanidae, and ostracods with increasing abundances below the pycnocline and abov e the bottom. The larvacean Oikopleura dioica was often sampled in significant numbers as well ( Dagg, 1995; Sutton et al., 2001). Taking into the account the sampling differences between n ets and imaging systems, the results from this research confirmed these findings with the PIC ES classification of similarly sampled data (chaetognaths, cladocerans, copepods, eumalacostrac an crustaceans, and ostracods) normally within half or double of each other and very much w ithin the range of differences found between replicate net samples (Wiebe and Holland, 1968). In the instances where net counts for some of the crustacean groups were significantly higher tha n the concurrent SIPPER-2 estimates, some of that difference may be explained by orientation effects where an imaged crustacean larger than the extracted size threshold is imaged at an o rientation that makes it appear smaller than the threshold and is subsequently not extracted and or by incorrect flow speed measurements through the sampling tube affecting the calculated size of the SIPPER-2 images. While the SIPPER-2 sampled crustacean distributions were similar to earlier studies, the abundance of the fragile and gelatinous zooplankton groups were significantly different. Gelatinous and fragile taxa made up almost half of the imaged zooplankton during this study with larvaceans comprising between 16-48% of the fragile community at any one depth. These numbers suggest that their ecological roles in the WFS may be much greater than previously realized. Prior studies of the WFS had demonstrated that larvaceans were an important contributor to the zooplankton assemblage but to a lesser degree than this study (Sutton et al., 2001). Larvaceans are more effective phytoplankton grazers than crustaceans on the WFS (Dagg 1995; Sutton et al., 2001), with grazing rate s between 1.5-18 that of the abundant crustacean taxa found there (Sutton et al., 2001). Therefore, accurate abundance estimates for
111 this group can have a significant effect of the est imated grazing impact of the zooplankton community on WFS phytoplankton standing stock. The occurrence of the abundant hydromedusae A. hemistoma had not been documented before on the WFS but has been noted as the dominan t hydromedusae in the southwestern Gulf of Mexico on the Campeche Bank (Segura-Puertes and Ordonez-Lopez, 1994) and is the dominant hydromedusae found on and off Caribbean re ef systems (Suarez-Morales et al., 1999; Gasca et al., 2003). Abundances of up to 179 m-3 were reported by Segura-Puertes and Ordonez-Lopez (1994) using a 0.5 m bongo net. Durin g this study, A. hemistoma abundances averaged 379 m-3 and exceeded 1300 m-3 below the pycnocline. These numbers approach the maximum recorded concentration of a single taxa of hydromedusae (Colin et al., 2005). This species has been shown to feed on a wide range of p rey ranging from photosynthetic protoctists to copepods. Ingestion rates of phytoplankton and o ther protoctists are unknown, so their impact as possible grazers of primary production has not b een investigated. Their numbers have been shown to increase during seasonal phytoplankton blo oms (Costello and Mathieu, 1995) but it is unknown whether that is in response to phytoplankto n or their grazers. Small hydromedusae could play in an important role in oligotrophic sys tems as remineralizers of nutrients as their fecal pellets do not sink and therefore are recycled with in the water column (Colin et al, 2005). Large sarcodine protoctists have long been known to be under-sampled by nets and bottles even before the advent of in-situ imaging s ystems due to their fragility, patchiness and difficulty in preservation (Michaels, 1988). Diver surveys have indicated that large acantharians, foraminiferans and both solitary and colonial radio larians can be important components of epipelagic plankton communities and and that they c ontribute significantly to both carbon flux and primary productivity (Michaels, 1988; Michaels et a l., 1995). Dennet et al. (2002) documented the distribution and abundance of colonial spumellarian radiolarians using the VPR in the tropical Pacific and estimated their abundance between 1-3 o rders of magnitude greater than previous estimates using traditional sampling methods. They estimated that colonial radiolarians alone could contribute up to 9% of the primary production in the upper 150 m. We observed large protoctist abundances between 100-800 m-3 during this study, suggesting that they may play
112 important roles as both predators and prey for meso zooplankton and be a significant contributor to primary production on the WFS. Marine snow was the third most abundant particle cl ass imaged during this study. Marine snow can be a food source for zooplankton and larva l fish (Lampitt, 1992) and provides a substrate on which zooplankton can reside (Steinber g et al., 1994; Green and Dagg, 1997). Marine snow was the most abundant particle class im aged by the VPR on Georges Bank (Ashjian et al., 2001) and found in greatest concentrations near the bottom and along mixing fronts between water masses and can indicate areas of dyna mic biological activity. Marine snow abundance was highly correlated with zooplankton ab undance in this study. Larvaceans were the dominant zooplankton class imaged and have been shown to be a major generator of marine snow. Sato et al. (1996) measured O. dioica house p roduction rates of 8-19 houses a day under experimental conditions while Hansen et al. 1996 me asured twice as many discarded houses as larvaceans in East Sound, Orcas Island, Washington. Larvacean communit y daily house production can be between 150-1100% of the larvacea n biomass (Hopcroft and Roff, 1998; Sato et al., 2001) so the imaged larvacean population sa mpled during this study could be the primary contributor to the observed marine snow distributio n. Discarded larvacean houses were observed in the SIPPER-2 marine snow images but they were no t a substantial component. However, larvacean houses undergo drastic morphological chan ge after they are discarded so it may be difficult to identify them via the SIPPER-2 images (Alldredge, 1976; Koski et al., 2007). Understanding the distribution and abundance of the colonial cyanobacteria Trichodesmium is critical for understanding nutrient dynamics in oligotrophic systems. Trichodesmium is the dominant nitrogen fixer in subtropical and tropical ocean waters (Falcon et al.,2004) and has been implicated as a precursor to blooms of the red-tide forming dinoflagellate Karenia brevis on the WFS (Walsh and Steidinger, 20 01). The fragile nature of Trichodesmium colonies make it difficult to accurately measure it s abundance using invasive sampling methods such as nets and bottles (Chang 2000). In-situ imag ing methods can more effectively sample Trichodesmium at low concentrations (Remsen et al., 2004) and ca n provide greater spatial resolution to assess its distribution across a broa d range of spatial scales (Davis and
113 McGillicuddy, 2006). Trichodesmium sampled by the SIPPER-2 exhibited peak abundances in the upper 30 m from the morning through early after noon after which surface water abundances decreased dramatically while deeper stocks increase d. This is consistent with the buoyancy regulating behavior of Trichodesmium in which they accumulate carbohydrate ballast dur ing the day as they photosynthesize, causing them to sink, and ascend towards the surface as they lose the ballast to respiratory consumption during the e vening (Villareal and Carpenter, 2000). This behavior is especially useful to Trichodesmium on the WFS where available stocks of phosphate are found at depth so that this reverse pattern exp oses them to light during the day to photosynthesize and to nutrients at night (Walsh et al., 2006). Possibility of Diel Vertical Migration of the WFS P lankton Assemblage Sampling only took place during a single 24 hour pe riod so it was not possible to determine the magnitude of diel vertical migration behavior in any of the imaged particle groups. Checkley et al. (1992) found that most zooplankton groups in the upper 30 m on the northwest Texas shelf underwent some degree of diel vertical migration, especially in the lower half of the water column. Ostracods and Oithona both showed so me indication that part of their population migrates up from the subpycnocline layer during the evening consistent with those findings, while Trichodesmium and to a lesser degree the elongate phytoplankton class exhibited reverse vertical migration towards the surface during the d ay. The elongate phytoplankton class was made of both diatoms and dinoflagellate colonies th at have the capability to vertically migrate (Villareal et al., 1999; Whittington et al., 2000). The vertical migration behavior of both Trichodesmium and the elongate phytoplankton had no discernible effect on the profiles of fluorescence or extracted chlorophyll. The importan ce of diel vertical migration of different plankton classes on the WFS could not be establishe d in this study, especially since the smaller crustacean zooplankton that are important grazers o n the WFS (Sutton et al., 2001) were not effectively sampled by SIPPER-2. The vertical distr ibution behavior of plankton on the WFS should be an area of further study as such behavior can concentrate plankton at specific depth strata that may be missed with conventional samplin g methods that may then lead to underestimates of rate processes such as production and grazing (Cowles et al, 1998).
114 Observations of Plankton Behavior Other observable behaviors were evident from this d ataset but were not explicitly explored. Although not included in the final classi fier, scyllarid lobster phyllosoma were enumerated during the building of the initial train ing library. Over half were observed directly associated with one or more hydromedusae, ctenophor e, siphonophore, protoctist and marine snow. While phyllosoma associations with hydromedus ae, ctenophores (Thomas, 1963; Barnett et al., 1986) and siphonophores (Ates et al., 2007) have been reported, our observations of associations with marine snow aggregates and large sarcodine protoctists appear to be firsts. While predation events were noted in many of the zo oplankton groups, they were most visibly present in the chaetognaths. While all of the chaet ognath images were not scanned for this behavior, several dozen were noted with prey items in their mouth. The majority of the identifiable prey items were of larvaceans, although cladocerans copepod and ostracods were also observed as prey items. Larvaceans are the one of t he primary prey item of chaetognaths (Feigenbaum, 1982; Kimmerer, 1984) so these observa tions are not surprising. Such observations however, could be useful in determinin g rate processes such as instantaneous predation, reproduction and encounter rates and so on, especially as imaging sensors continue to improve. Additionally, many of these behaviors may influence the accuracy of a classifier as well. For example, a pair of mating copepods or chaetogna ths will look far different to a classifier than the images of single organisms most likely in the t raining library. Fine Scale Distribution of SIPPER Particle Classes Until recently, direct observations of the fine-sca le distribution of plankton were difficult to collect and limited in scope (Cassie et al. 1963; O wen 1989). These observations demonstrated that plankton can be found in dense patches at abun dances many times above background concentrations and that these patches are important loci for enhanced production, feeding and reproductive opportunities (Lasker, 1974; Mullins a nd Brooks, 1976; Folt and Burns, 1999). The development of in-situ imaging systems now make it possible to observe individual organisms and measure the distance between them and examine t he spatial behavior of individual taxa (Davis et al;, 1992; Malkiel et al., 1999; Widder a nd Johnsen, 2000; Malkiel et al., 2006). These
115 observations can shed light on the mechanisms by wh ich plankton aggregations persist (Haury and Yamakazi, 1995), the degree of spatial overlap between predator and prey (DeRobertis, 2002) and zooplankton associations with marine snow (Malkiel et al., 2006). The field of view of the SIPPER-2 and other in-situ imaging systems is s imilar to that of larval fish (Kiorboe and Visser, 1999) and therefore observations from these sensors are useful models to determine the prey field individuals may encounter as they develo p Using SIPPER-2, we found that approximately 25% of the observed plankton and particle classes were non-randomly distributed relative to e ach other and formed observable finescale aggregations. Most of the imaged groups ANND were s imilar to those expected under spatial randomness. While we could not determine the threedimensional nearest NND using the SIPPER-2, we still feel our ANND findings were vali d, especially as edge effects should be leading us to underestimate the degree of clusterin g for those distributions considered as such. Most of the clustered zooplankton groups had median ANND outside the range of their reported perception distances (Feigenbaum and Reeve, 1977; K iorboe and Visser, 1999; Haury and Yamakazi, 1995) suggesting that most of these aggre gations were not behaviorally controlled. This fine-scale clustering did not translate into a ny appreciable meter-scale clustering along the horizontal sampling transects as determined with Ll oydÂ’s index of patchiness. Therefore for most of these groups there was similar or greater variab ility in the vertical domain. For grazers of these groups or conspecifics seeking mates, it might be a shorter distance vertically to encounter higher concentrations of selected plankton than it is to e ncounter the same abundance horizontally. This is especially true if there are vertical aggregatio ns in thin layers where abundances can be found many orders of magnitude higher than the water colu mn average (Cowles et al., 1998). These findings mostly agree with those of DeRobertis (200 2) who used acoustics and high resolution digital video to study the NND of euphausiids, amph ipods and fish and found that NND distances were mostly randomly distributed and that horizonta l meter-scale clustering was rarely observed. Additionally, PICES was able to correctly predict t he finescale distribution of most of the abundant particle classes observed by the SIPPER-2. This wil l allow for the increasingly large image
116 datasets being collected by in-situ imagers to be a nalyzed by automated classification systems for not only abundance but finescale distribution b ehavior as well (Ashjian et al., 2005a). Conclusions In summary, we found that the distribution of PICES classified SIPPER-2 images of zooplankton taxa were similar to that from previous studies for those groups that are representatively sampled by nets. For more fragile and gelatinous zooplankton taxa, we found that earlier studies significantly underestimated t he abundance of hydromedusae, larvaceans and sarcodine protoctists and that these taxa could pla y a significant role in the trophodynamics of the WFS. Marine snow was found in appreciable quantitie s throughout the water column and at high densities below the pycnocline and was closely asso ciated with zooplankton abundance. Trichodesmium was found in non-bloom concentrations but observations suggested it performs a buoyancy regulated diel vertical migration as noted by others elsewhere (Villareal and Carpenter, 2003). Using SIPPER_2 data we also found that an ap preciable proportion of the observed plankton taxa and particles form small scale aggreg ations but that these fine-scale patches do not create appreciable patchiness at the meter-scale in the horizontal domain. We found that PICES was capable of detecting these fine-scale distribut ion patterns, especially for the more abundant taxa where the false-positive detection rate of the classifier was less significant. Continued improvements of both classification methods and insitu imaging technology will allow for greater application of these methods to accurately describe planktic systems and processes in the water column while rapidly improving the turn-around time in getting the results analyzed and disseminated to the public.
117 References Akiba, T., Kakui, Y., 2000. Design and testing of a n underwater microscope and image processing system for the study of zooplankton dist ribution. IEEE Journal of Oceanic Engineering 25 (1), 97-104. Alldredge A.L. (1976) Discarded appendicularian hou ses as sources of food, surface habitats, and particulate organic matter in planktonic enviro nments. Limnology and Oceanography 21, 14-23 Alldredge A.L., Passow U., Hasddock H.D. (1998) The characteristics and transparent exopolymer particle (TEP) content of marine snow fo rmed from thecate dinoflagellates. Journal of Plankton Research 20(3), 393-406. Ashjian, C.J., Davis, C.S., Gallager, S.M., Alatalo P., 2001. Distribution of plankton, particles, and hydrographic features across Georges Bank descr ibed using the Video Plankton Recorder. Deep-Sea Research Part II 48 (1-3), 245-2 82. Ashjian C.J., Davis C., Gallager S.M., Alatalo P. ( 2005) Characterization of the zooplankton community, size composition, and distribution in re lation to hydrography in the Japan/East Sea. Deep Sea Research II 52(11-13), 136 3-1392 Austin, H.M., Jones, J.I., 1974. Seasonal variation of physical oceanographic parameters on the Florida Middle Ground and their relation to zooplan kton on the West Florida Shelf. Florida Scientist 37 (1), 5-16. Baumgartner, M.F., 2003. Comparisons of Calanus finmarchicus fifth copepodite abundance estimates from nets and an optical plankton counter Journal of Plankton Research, 25 (7), 855-868. Beaulieu, S.E., Mullin, M.M., Tang, V.T., Pyne, S.M ., King, A.L., Twining, B.S., 1999. Using an optical plankton counter to determine the size dist ributions of preserved zooplankton samples. Journal of Plankton Research 21 (10), 1939 -1956.
118 Benfield, M. C., Davis, C. S., Wiebe, P. H., Gallag er, S. M., Lough, R.G., Copley, N. J. 1996. Video plankton recorder estimates of copepod, ptero pod and larvacean distributions from a stratified region of Georges Bank with comparativ e measurements from a MOCNESS sampler. Deep-Sea Research Part II 43 (7-8), 1925-1 945. Benfield M.C., Grosjean P., Culverhouse P.F., Irigu oien X., Sieracki M., Lopez-Urrutia A., Dam H.G., Hu Q., Davis C.S., Hansen A., Pilskaln C.H., Riseman E., Schultz H., Utgoff P.E., Gorsky G. (2007) RAPID research on automated plankt on identification. Oceanography 20(2), 12-26. Biggs, D.C., Ressler P.H. 2001. Distribution and ab undance of phytoplankton, zooplankton, icthyoplankton and micronekton in the deepwater Gul f of Mexico. Gulf of Mexico Science 19, 7-29. Broughton E.A., Lough R.G. (2006) A direct comparis on of MOCNESS and Video Plankton Recorder zooplankton abundance estimates: Possible applications for augmenting net sampling with video systems. Deep-Sea Research II 5 3 (2006) 2789Â–2807 53, 2789Â– 2807. Capone, D.G., Zehr D.G., Paerl H.W., Bergmann B., C arpenter E.J. 1997. Trichodesmium a globally significant marine cyanobacterium. Science 276, 1221-1229. Cassie, R.M., 1963. Microdistribution of plankton. Oceanography and Marine Biology. 1, 223252. Chang J. 2000. Precision of different methods used for estimating the abundance of the nitrogenfixing marine cyanobacterium, Trichodesmium Ehrenberg. Journal of Experimental Marine Biology & Ecology 245, 215-224. Checkley D.M., Uye S., Dagg M.J., Mullin M.M., Omor i M., Onbe T., Zhu M.-y. (1992) Diel variations pf the zooplankton and its environmnet a t neritic stations in the inland sea of Japan and the north-west Gulf of Mexico. Journal of Plankton Research 14(1), 1-40.
119 Colin S.P., Costello J.H., Graham W.M., Higgins III J.H. (2005) Omnivory by the small cosmopolitan hydromedusa Aglaura hemistoma. Limnolo gy and Oceanography 50(4), 1264-1268 Costello J.H., Mathieu H.W. (1995) Seasonal abundan ce of medusae in Eel Pond, Massachusetts, USA during 1990-1991. Journal of Pla nkton Research 17, 199-204. Cowles T.J., Desiderio R.A., Carr M.-E. (1998) Smal l-scale planktonic structure: persistence and trophic consequences. Oceanography 11(1), 4-9. Culverhouse P.F., Simpson R.G., Ellis R., Lindley J .A., Williams R., Parisini T., Reguera B., Bravo I., Zoppoli R., Earnshaw G., McCall H., Smith G. (1996) Automatic classification of field-collected dinoflagellates by artificial neura l network. Marine Ecology Progress Series 139, 281-287. Culverhouse P.F., Williams R., Reguera B., Herry V. Gonzalez-Gil S. (2003) Do experts make mistakes? A comparison of human and machine indenti fication of dinoflagellates. Marine Ecology Progress Series 247, 17-25. Culverhouse P.F., Williams R., Benfield M.C., Flood P.R., Sell A.F., Mazzocchi M.G., Buttino I., Sieracki M. (2006) Automatic image analysis of plan kton: future perspectives. 312 297309, Cummings J.J. 1983. Habitat dimensions of calanoid copepods in the western Gulf of Mexico. Journal of Marine Research 41,163-188. Dagg M.J. (1995) Copepod grazing and the fate of ph ytoplankton in the northern Gulf of Mexico. Continental Shelf Research 15(11/12), 1303-1317. Davis, C.S., Gallagher, S.M., Berman, M.S., Haury, L.R., Strickler, J.R., 1992. The Video Plankton Recorder (VPR): Design and initial results Archiv fr Hydrobiologie Advances in Limnology 36, 67-81. Davis C.S., Hu Q., Gallager S.M., Tang X., Ashjian C.J. (2004) Real-time observation of taxaspecific plankton distributions: an optical samplin g method. Marine Ecology Progress Series 284, 77-96
120 Davis C.S., McGillicuddy Jr. D.J. (2006) Transatlan tic Abundance of the N2-Fixing Colonial Cyanobacterium Trichodesmium Science 312, 1517-1520. Dennett, M. R., Caron, D. A., Michaels, A. F., Gall ager, S. M., Davis, C. S., 2002. Video plankton recorder reveals high abundances of colonial Radiol aria in surface waters of the central North Pacific. Journal of Plankton Research 24 (8), 797-805. De Robertis A. (2002) Small-scale spatial distribu tion of the euphausiid Euphausia pacifica and overlap with planktivorous fishes. Journal of Plank ton Research 24(11), 1207-1220. Falcon L.I., Carpenter E.J., Cipriano F., Bergman B ., Capone D.C. (2004) N2 Fixation by Unicellular Bacterioplankton from the Atlantic and Pacific Oceans: Phylogeny and In Situ Rates. Applied and Environmental Microbiology 70(2) 765-770. Feigenbaum, D., Reeve, M.R. (1977) Prey Detection in the Chaetognatha: Response to a Vibrating Probe and Experimental Determination of A ttack Distance in Large Aquaria. Limnology and Oceanography 22 (6), 1052-1058. Feigenbaum D. (1982) Feeding by the chaetognath, Sagitta elengans at low temperatures in Vineyard Sound, Massachusetts. Limnology and Oceano graphy 27(4), 699-706 Folt, C.L., Burns, C.L. (1999) Biological Drivers o f zooplankton patchiness. Trends in Ecology and Evolution 14 (8), 300-305. Foote, K.G., 2000. Optical Methods. In: Harris, R.P ., Wiebe, P.H., Lenz, J., Skjoldal, H. Â–R., Huntley, M. (Editors), ICES Zooplankton Methodology Manual. Academic Press, San Diego, pp. 259-295. Gallager, S.M., Davis, C.S., Epstein, A.W., Solow, A., Beardsley, R.C., 1996. High-resolution observations of plankton spatial distributions corr elated with hydrography in the Great South Channel, Georges Bank. Deep-Sea Research Part II 43 (7-8), 1627-1663. Gallienne, C.P., Robins, D.B., 2001. Is Oithona the most important copepod in the world's oceans? Journal of Plankton Research 23 (12), 14211432. Gallienne, C.P., Robins, D.B., Woodd-Walker, R.S., 2001. Abundance, distribution and size structure of zooplankton along a 20o west meridional transect of the northeast Atlantic Ocean in July. Deep Sea Research II 48 (4-5), 925-9 49.
121 Gasca R., Segura-Puertes L., Suarez-Morales E. (200 3) A survey of the medusan (Cnidaria) community of Banco Chinchorro, Western Caribbean Se a. Bulletin of Marine Science 73(1), 37-46. Gorsky G., Aldorf C., Kage M., Picheral M., Garcia Y., Favole J. (1992) Vertical distribution of suspended aggregates determined by a new underwater video profiler. Annales de l'Institut ocanographique 68, 275-280.. Grant, S., Ward, P., Murphy, E., Bone, D., Abbott, S., 2000. Field comparison of an LHPR net sampling system and an Optical Plankton Counter (OP C) in the Southern Ocean. Journal of Plankton Research 22 (4), 619-638. Green E.P., Dagg M.J. (1997) Mesozooplankton associ ations with medium to large marine snow aggregates in the northern Gulf of Mexico. Journal of Plankton Research 19(4), 435-447. Grosjean P., Picheral M., Warembourg C., Gorsky G. (2004) Enumeration, measurement, and identification of net zooplankton samples using the ZOOSCAN digital imaging system. ICES Journal of Marine Science 61, 518-525. Halliday, N.C., Coombs, S.H., Smith, 2001. A compar ison of LHPR and OPC data from vertical distribution sampling of zooplankton in a Norwegian fjord. Sarsia 86(2), 87-99. Haury, L.R., McGowan, J.A., Wiebe, P.H., 1977. Patt erns and processes in the time-space scales of plankton distributions. In: Steele, J.H. (ed), S patial Patterns in plankton communities. Plenum Press, New York, pp. 277-326. Herman, A.W., 1988. Simultaneous measurement of zoo plankton and light attenuance with a new optical plankton counter. Continental Shelf Researc h 8 (2), 205-221. Herman, A.W., 1992. Design and calibration of a new optical plankton counter capable of sizing small zooplankton. Deep Sea Research I 39 (3-4), 39 5-415. Hopcroft, R., Roff, J., Chavez, F., 2001. Size para digms in copepod communities: a reexamination. Hydrobiologia 453 (1-3), 133-141. Hopkins, T.L., 1981. The vertical distribution of z ooplankton in the eastern Gulf of Mexico. DeepSea Research 29 (9A), 1069-1083.
122 Hopkins, T.L., Sutton, T.T., Lancraft, T.M., 1996. The trophic structure and predation impact of a low latitude midwater fish assemblage. Progress in Oceanography 38, 205-239. Iwamoto, S., Checkley, D.M. and Trivendi, M.M., 200 1. REFLICS: Real-time flow imaging and classification system. Machine Vision and Applicati ons 13, 1-13. Jeffries, H.P., Berman, M.S., Poularikas, A.D., Kat sinis, C., Melas, I., Sherman, K., Bivins, L., 1984. Automated sizing, counting and identification of zooplankton by pattern recognition. Marine Biology 378 (3). 329-334. Karl, D.L., R; Tupas, L; Dore, J; Christian, J; Heb el, D, 1997. The role of nitrogen fixation in biogeochemical cycling in the subtropical North Pac ific Ocean. Nature 388 (6642), 533538. Kimmerer W.J. (1984) Selective predation and its im pact on prey of Sagitta enflata (Chaetognatha). Marine Ecology Progress Series 15, 55-62. Kirboe T., Andersen K.P., Dam H.G. (1990) Coagulat ion efficiency and aggregate formation in marine phytoplankton. Marine Biology 107(2), 235-24 5. Kirboe T., Visser, A.W., (1999) Predator and prey perception in copepods due to hydromechanical signals. Marine Ecology Progress Se ries 179, 179-185. Koski M., Mller E.F., Maar M., Visser A.W. (2007) The fate of discarded appendicularian houses: degradation by the copepod, Microsetella norvegica and other agents. Journal of Plankton Research 29(7), 641-654. Labat, J.M., P; Dallot, S; Errhif, A; Razouls, S; S abini, S, 2002. Mesoscale distribution of zooplankton in the Sub-Antarctic Frontal system in the Indian part of the Southern Ocean: a comparison between optical plankton counter and n et sampling. Deep-Sea Research Part I 49 (4), 735-749. Lampitt R.S. (1992) The contribution of deep-sea ma croplankton to organic remineralization: Results from sediment trap and zooplankton studies over the Madeira Abyssal Plain. Deep Sea Research I 39(2), 221-233. Lasker, R. (1975) Field criteria of first-feeding a nchovy larvae: the relation between inshore chlorophyll maximum layers and successful first fee ding. Fisheries Bulletin 73, 453-462.
123 Lenes, J. M., Darrow, B. P., Cattrall, C., Heil, C. A., Callahan, M., Prospero, J. M., Bates, D. E., Fanning, K. A., Walsh, J. J., 2001. Iron Fertilizat ion and the Trichodesmium response on the West Florida Shelf. Limnology and Oceanography 46 (6), 1261-1277. Lenz, J., 2000. Introduction. In: Harris, R.P., Wie be, P.H., Lenz, J., Skjoldal, H.Â–R., Huntley, M. (eds.) ICES Zooplankton Methodology Manual. Academi c Press, San Diego, 1-32. Luo T., Kramer K., Goldgof D., Hall L., Samson S., Remsen A., Hopkins T. (2004) Recognizing plankton images from the Shadow Image Particle Prof iling Evaluation Recorder. IEEE Transactions on Systems, Man, and Cybernetics Part B. 34(4), 1753-1762. Luo T., Kramer K., Goldgof D., Hall L., Samson S., Remsen A., Hopkins T. (2005) Active Learning to Recognize Multiple Types of Plankton. J ournal of Machine Learning Research 6, 589Â–613. Mackas D.L., Denman K.L., Abbot M.R. (1985) Plankto n patchiness: Biology in the physical vernacular. Bulletin of Marine Science 37(2), 652-6 74. Malkiel E., Alquaddoomi O., Katz J. (1999) Measurem ents of plankton distribution in the ocean using submersible holography. Measurement Science a nd Technology 10, 1142Â–1152. Malkiel E., Abras J.N., Widder E., Katz J. (2006) O n the spatial distribution and nearest neighbor distance between particles in the water column dete rmined from in situ holographic measurements. Journal of Plankton Research 28(2), 1 49-170. Maul, G.A., Vukovich F.M., 1993 The relationship be tween variations in the Gulf of Mexico Loop Current and Straits of Florida volume transport. Jo urnal of Physical Oceanography 23, 785-796. Michaels A.F. (1988) Vertical distribution and abun dance of Acantharia and their symbionts. Marine Biology 97(4), 559-569. Michaels, A.F., Caron, D.A., Swanber, N.R., Howse, F.A., Michaels, C.M. (1995) Planktonic sarcodines (Acantharia, Radiolaria, Foraminifera) i n surface waters near Bermuda: abundance, biomass and vertical flux. Journal of Pl ankton Research 17(1), 131-163. Mullin M.M., Brooks E.R. (1976) Some consequences o f distributional heterogeneity of phytoplankton and zooplankton. Limnology and Oceano graphy 21(6), 784-796.
124 Mller-Karger, F.E., 2000. The spring 1998 Gulf of Mexico (NEGOM) cold water event: Remote sensing evidence for upwelling and for eastward adv ection of Mississippi Water (or : How an errant Loop Current anticyclone took the NEGOM f or a spin). Gulf of Mexico Science 19, 55-67. Mller-Karger, F.E., Walsh, J.J., Evans, R.H. and M eyers, M.B., 1991. On the seasonal phytoplankton concentration and sea surface tempera ture cycles of the Gulf of Mexico as determined by satellites. Journal of Geophysical Re search 96 (C7), 12645-12665. Nishikawa, J., Terazaki, M., 1996. Tissue shrinkage of two gelatinous zooplankton, Thalia democratica and Dolioletta gegenbauri (Tunicata:Thaliacea) in preservative. Bulletin of Plankton Society of Japan, 43 (1), 1-7. Norrbin, M.F., Davis, C.S., Gallagher, S.M., 1996. Differences in fine-scale structure and composition of zooplankton between mixed and strati fied regions of Georges Bank. Deep Sea Research Part II, 43 (7-8), 1905-1924. Omori, M., 1978. Some factors affecting dry weight, organic weight and concentration of carbon and nitrogen in freshly prepared and preserved zoop lankton. Internationale Revue der Gesamten Hydrobiologie, 63, 261-269. Omori, M.H., Hamner, W.M., 1982. Patchy distributio n of zooplankton: Behavior, population assessment and sampling problems. Marine Biology, 7 2 (2), 193-200. Ortner, P.B., Hill, L.C. and Cummings, S.R., 1989. Zooplankton community structure and copepod species composition in the northern Gulf of Mexico. Continental Shelf Research 9 (4), 387-402. Ortner, P. B., Lee, T. N., Milne, P. J., Zika, R. G ., Clarke, E., Podesta, G. P., Swart, P. K., Tester P. A., Atkinson, L. P., Johnson, W. R., 1995. Missi ssippi River flood waters that reached the Gulf Stream. Journal of Geophysical Research 10 0 (C7), 13595-13601. Owen R.W. (1989) Microscale and finescale variation s of small plankton in coastal and pelagic environments. Journal of Marine Research 47, 197-24 0. Pillar, S.C. 1984. A comparison of the performance of four zooplankton samplers. South African Journal of Marine Science 2 1-18.
125 Pilskaln C.H., Villareal T.A., Dennett M., Darkange lo-Woodd C., Meadows G. (2005) High concentrations of marine snow and diatom algal mats in the North Pacific Subtropical Gyre: Implications for carbon and nitrogen cycles i n the oligotrophic ocean. Deep-Sea Research I 52, 2315-2332 Postel, L., Fock, H., Hagen, W., 2000. Biomass and abundance. In: Harris, R.P., Wiebe, P.H., Lenz, J., Skjoldal, H.Â–R., Huntley, M. (eds.), ICES Zooplankton Methodology Manual. Academic Press, San Diego, pp. 83-192. Remsen A., Wilcox T., Hopkins T., LeBlanc L., Sutto n T. (1996) A high frequency chirp sonar to be deployed on the High Resolution Sampler capable of sizing small zooplankton Oceans 96, Ft. Lauderdale, pp 1480-1484. Remsen A., Samson S., Hopkins T. (2004) What you se e is not what you catch: a comparison of concurrently collected net, Optical Plankton Counte r, and Shadowed Image Particle Profiling Evaluation Recorder data from the northea st Gulf of Mexico. Deep Sea Research I 51(1), 129-151 Rolke M., Lenz. J. (1984) Size structure analysis o f zooplankton samples by means of an automated image analyzing system. 6:637Â–645. Journ al of Plankton Research 6, 637645. Sameoto, D., Cochrane, N., Herman, A.W., 1993. Conv ergence of acoustic, optical, and net catch estimates of euphausiid abundance: Use of artificia l light to reduce net avoidance. Canadian Journal of Fishery and Aquatic Science 50, 334-346. Samson, S., Hopkins, T., Remsen, A., Langebrake, L. Sutton, T., Patten, J., 2001. A system for high resolution zooplankton imaging. IEEE Journal o f Oceanic Engineering 26 (4), 671676. Sato R., Tanaka Y., Ishimaru T. (2001) House Produc tion by Oikopleura dioica (Tunicata, Appendicularia) Under Laboratory Conditions. Journa l of Plankton Research 23(4), 415423.
126 Schulze PC, Williamson CE, Sprules WG (1992) Conclu ding remarks: A comparison of new devices for studying zooplankton in situ. Archiv f r Hydrobiologie Advances in Limnology 36, 135-140. Segura-Puertes L., Ordonez-Lopez U. (1994) Analisis de la Comunidad de Medusas (Cnidaria) de la Region Oriental del Banco de Campeche y el Ca ribe Mexicano. Caribbean Journal of Science 30(1-2), 104-115. Sieracki M., Gifford D., Gallager S.M., Davis C.S. (1998) Observations on a dense patch of the diatom, Chaetoceros socialis on the southern flank of Georges Bank: distributi on, colony structure and grazing losses. Oceanography 11, 30-3 5. Skjoldal, H.-R., Wiebe, P.H., Foote, K.G., 2000. Sa mpling and experimental design. In: Harris, R.P., Wiebe, P.H., Lenz, J., Skjoldal, H.Â–R., Huntl ey, M. (eds.), ICES Zooplankton Methodology Manual. Academic Press, San Diego, pp. 33-54. Sprules WG, Bergstrom B, Cyr H, Hargreaves BR, Kilh am SS, MacIsaac HJ, Matsushita K, Stemberger RS, Williams R (1992) Non-video optical instruments for studying zooplankton distribution and abundance. Archiv fr Hydrobiologie Advances in Limnology 36, 45-58. Sprules, W.G., Jin, E.H., Herman, A.W. and Stockwel l, J.D., 1998. Calibration of an optical plankton counter for use in fresh water. Limnology and Oceanography 43 (4), 726-733. Steinberg D.K., Silver M.W., Pilskaln C.H., Coale S .L., Paduan J.B. (1994) Midwater zooplankton communities on pelagic detritus (giant larvacean ho uses) in Monterey Bay, California. Limnology & Oceanography 39(7), 1606-1620 Suarez-Morales E., Segura-Puertes L., Gasca R. (199 9) Medusan (Cnidaria) assemblages off the Caribbean coast of Mexico. Journal of Coastal Resea rch 15, 140-147 Sutton, T.T., Hopkins, T.L., Remsen, A., Burghart, S., 2001. Multisensor sampling of pelagic ecosystem variables in a coastal environment to est imate zooplankton grazing impact. Continental Shelf Research 21 (1), 69-87. Tang, X., Stewart, W.K., Vincent, L., Huang, H., Ma rra, M., Gallager, S.M., Davis, C.S., 1998. Automatic plankton image recognition. Artificial In telligence Review 12, 177-199.
127 Vidal, V. M. V., Vidal, F. V., Hernndez, A. F., Me za, E., Zambrano, L., 1994. Winter Water Mass Distributions in the Western Gulf of Mexico Affecte d by a Colliding Anticyclonic Ring. Journal of Oceanography 50 (5), 559-588. Villareal T.A., Pilskaln C., Brzezinski M., Lipschu ltz F., Dennett M., Gardner B.G. (1999) Upward transport of oceanic nitrate by migrating diatom ma ts. Nature 397, 423-425. Villareal, T.A. and E.J. Carpenter. 2003. Buoyancy regulation and the potential for vertical migration in the oceanic cyanobacterium Trichodesmium. Microbial Ecology 45:1-10. Walsh J.J., and K. A. Steidinger (2001), J. Geoph ys. Res., 106, (2001) Saharan dust and Florida red tides: The cyanophyte connection. Journal of Ge ophysical Research -Oceans 106, 11,597 Â–511,612 Walsh J.J., J.K. Jolliff B.P. Darrow J.M. Lene s S.P. Milroy A.W. Remsen D.A. Dieterle K.L. Carder F.R. Chen G.A. Vargo R.H. Weisber g K.A. Fanning F.E. Muller-Karger, E. Shin K.A. Steidinger C.A. Heil C.R. Tomas, J.S. Prospero, T.N. Lee G.J. Kirkpatrick, A.E. Whitledge D.A. Stockwell, T.A. Villareal A.E. Jochens, and P.S. Bontempi (2006) Red tides in the Gulf of Mexico: wh ere, when, and why. Journal of Geophysical Research -Oceans 111(C11003), 1-46. Warren, J.D., Stanton, T.K., Benfield, M.C., Wiebe, P.H., Chu, D., Sutor, M., 2001. In-situ measurements of acoustic target strengths of gas-be aring siphonophores. ICES Journal of Marine Science 58 (4), 740-749. Whittington J., Sherman B., Green D., Oliver R.L. ( 2000) Growth of Ceratium hirundinella in a subtropical Australian reservoir: the role of verti cal migration. Journal of Plankton Research 22(6), 1025-1045 Widder E.A., Johnsen S. (2000) 3D spatial point pat terns of bioluminescent plankton: a map of the minefield. Journal of Plankton Research 22(3), 409-420. Wiebe P.H., Holland W.R. (1968) Plankton Patchiness : Effects On Repeated Net Tows. Limnology and Oceanography 13, 315-321.
128 Wiebe, P.H., Benfield, M.C., 2003. From the Hensen net toward four-dimensional biological oceanography. Progress in Oceanography 56 (1), 7-13 6. Wieland, K., Petersen, D. and Schnack, D., 1997. Es timates of zooplankton abundance and size distribution with the Optical Plankton Counter (OPC ). Archiv fur Fischereiund Meeresforschung 45 (3), 271-280. Woodd-Walker, R.S., Gallienne, C.P. and Robins, D.B ., 2000. A test model for optical plankton counter (OPC) coincidence and a comparison of OPC-d erived and conventional measures of plankton abundance. Journal of Plankton Research 22 (3), 473-483. Yamazaki H., Mackas D., Denman K. (2002) Coupling s mall scale physical processes with biology. In: Robinson AR, McCarthy JJ, Rothschild B J (eds) The Sea: Biological-Physical interaction in the Ocean. John Wiley and Sons, pp 5 1-112. Zar, J.H., 1984. Biostatistical Analysis. Prentice Hall, Englewood Cliffs, 1-718. Zhang, X., Roman, M., Sanford, A., Lascara, C. and Burgett, R., 2000. Can an optical plankton counter produce reasonable estimates of zooplankton abundance and biovolume in water with high detritus? Journal of Plankton Research 22 (1), 137-150. Zhou, M. and Tande, K., 2002. Optical Plankton Coun ter Workshop. GLOBEC Report 17, University of Tromso, Tromso.
About the Author Andrew Remsen graduated from the University of Wisc onsin in 1991 with a focus on limnology. He then worked for the Ecosystems Center at the Marine Biological Laboratory in Woods Hole Massachusetts until he was admitted to t he University of South Florida College of Marine Science graduate program in 1993. Mr. Remsen was admitted to Ph.D. candidacy in 2000. HeÂ’s been a principal investigator of the SI PPER project since 2001 and a full time employee of the Center for Ocean Technology since 2 002. HeÂ’s been chief scientist on 10 research cruises and has participated in over 20 mo re, spending more than half a year at sea. HeÂ’s presented research at over 15 domestic and int ernational conferences and workshops. HeÂ’s the author of two and co-author of 7 publications. He is married to a chemical oceanographer and has two daughters.