Remote Sens. 201 4 6 330 351 ; doi:10.3390/ rs60 1 0 330 remote sensing ISSN 2072 4292 www.mdpi.com/journal/remotesensing Article Karst Depression Detection Using ASTER ALOS / PRISM and SRTM Derived Digital Elevation Models in the Bambu Group, Brazil Osmar Ab lio de Carvalho J nior 1 *, Renato Fontes Guimar es 1 David R. Montgomery 2 Alan R. Gillespie 2 Roberto Arnaldo Trancoso Gomes 1 der de Souza Martins 3 and Nilton Correia Silva 4 1 Departamento de Geografia Campus Universit rio Darcy Ribeiro Universidade de Bras lia (UnB) Asa Norte, Bras lia DF 70910 900 Brazil; E Mails: renatofg@unb. br (R.F.G.); firstname.lastname@example.org (R.A.T.G.) 2 Department of Earth and Space Sciences University of Washington Seattle, WA 98195 USA; E Mails: email@example.com (D.R.M.); firstname.lastname@example.org (A.R.G.) 3 EMBRAPA Cerrados Planaltina, DF 73310 970 Brazil; E Mail: email@example.com 4 Faculdade de Engenharias do Gama Universidade de Bras lia (UnB), Gama DF 72444 240 Brazil; E Mail: firstname.lastname@example.org Author to whom correspondence should be addressed; E Mail: email@example.com ; Tel.: + 55 613 367 5001 Received: 10 November 2013; in revised form: 1 2 December 2013 / Accepted: 1 8 December 2013 / Published: 27 December 2013 Abstract: Remote sensing has been used in karst studies to identify limestone terrain, describe exokarst features, analyze karst depression s and detect geological structures important to karst development. The aim of this work is to investigate the use of ASTER SRTM and ALOS/PRISM derived digital elevation models (DEMs) to detect and quantify natural karst depressions along the S o Francis co River near Barreiras city, northeast Brazil. The study area is a karst landscape characterized by karst depressions ( dolines ) closed depressions in limestone, many of which contain standing water connected with the ground water table. The base of dolin es is typically sealed with an impermeable clay layer covered by standing water or herbaceous vegetation. We identify dolines by combining the extraction of sink depth from DEMs, morphometric analysis using GIS, and visual interpretation. Our methodology i s a semi automatic approach involving several steps: (a) DEM acquisition ; (b) sink depth calculation using the difference between the raw DEM and the corresponding DEM with sinks filled ; and (c) elimination of false ly identified karst OPEN ACCESS
Remote Sens. 2014 6 331 depressions using mo rphometric attributes. The advantages and limitations of the applied methodology using different DEMs are examined by comparison with a sinkhole map generated from traditional geomorphological investigations based on visual interpretation of the high resolution remote sensing image s and field surveys. The threshold values of the depth, area size and circularity index appropriate for distinguishing doline s w ere identified from the maximum overal l accuracy obtained by comparison with a true doline map Our results indicate that the best performance of the proposed methodology for meso scale karst feature detection was using ALOS/PRISM data with a threshold depth > 2 m ; area s > 13 125 m 2 and circu larity index es > 0.3 ( overall accuracy of 0.53 ). The overall correct identification of around half of the true dolines suggests the potential to substantially improve doline identification using higher resolution LiDAR generated DEMs. Keywords: Karst; limestone; DEM analysis; GIS; remote sensing; Brazil 1. Introduction Karst depressions cause damage both in rural areas through the loss of arable land and in urban areas due to damage to buildings, roads, and water supply systems [1,2]. Problems caused b y karst depressions have motivated many studies on their identification and spatial distribution using remote sensing data [3,4]. Historical changes and variations in the number and shape of karst depression s can be obtained from comparative studies of mul titemporal images [5,6]. Remote sensing also allows inferences about subsurface karst structures (endokarst). The regularity of the patterns and surface alignments of karst features often are associated with joint patterns, faulting and folding. Conduits i n karst groundwater are formed from rock dissolution along planes or discontinuities where the flow has characteristics similar to the water surface [7,8]. Endokarst environments are typically characterized by open conduits with low capacity for storage an d rapid groundwater flow. This intimate relationship between surface water and groundwater defines a system of interconnected caves and superficial features. Due to such relationships, the locations of karst aquifers or preferential flowpaths for groundwat er have been inferred by the positions of fracture sets or doline alignments apparent on aerial photographs and satellite images [9 11]. In addition, remote sensing data are used as inputs in GIS models for detecting and monitoring areas vulnerable to gro undwater pollution in karst terrain [12,13]. Digital terrain data are widely used to describe surface features and quantify topographic characteristics [14 17] and morphometric attributes derived from Digital Elevation Model s (DEM s ) have been used for automatic detection of elemental forms associated with landforms [18 20]. Several studies employ terrain attributes to characterize and describe karst processes [21 23]. In karst depression detection studies a promising terrain attribute is the sink dep th derived from the depression filling algorithm [24 27]. Such algorithms are an integral component of spatially distributed hydrological models that delineate watersheds, drainage networks and overland flowpaths [28 32]. Other methods to automate karst d epression recognition include convolution or filtering with kernel windows using focal functions  and the active contour method [34,35]
Remote Sens. 2014 6 332 an algorithm that delineates sinkhole boundaries with a compactness test and by fitting a local bi quadratic su rface to the points surrounding the potential sinkhole locations. However, tests made with ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) and SRTM (Shuttle Radar Topography Mission) [25 26] DEM data have shown that these data are no t sufficient for karst depression detection, making it necessary to combine the above approach with other methods to automate and improve the process of depression mapping. For example, Guimar es et al.  combined this approach with digital classification of spectral images and Siart et al.  used an iron oxide ratio and the vegetation infrared / red ratio from Quickbird imagery The increasing availability of high resolution DEMs and satellite images promises improved detection of karst de pressions through the combination of DEM analysis and remote sensing. The present paper aims to develop a semi automatic method for doline identification in central Brazil using three different DEMs: (a) the ASTER Global Digital Elevation Model (ASTER GDEM ) Version 1 from the USA s National Aeronautics and Space Administration (NASA) and Japan s Ministry of Economy, Trade and Industry (METI); (b) the SRTM Version 4.1 DEM compiled by the Consultative Group on International Agricultural Research Consortium for Spatial Information (CGIAR CSI); and (c) the DEM made from the high resolution satellite sensor Advanced Land Observing Satellite/Panchromatic Remote sensing Instrument for Stereo Mapping (ALOS/PRISM) The approach comb ines a threshold sink depth and morphometric analysis in order to refine the identification of karst depressions. Here, our focus is limited to automated DEM based classification and we do not address the strong potential for improved identification of dolines by combining DEM analysis with image classification [22,23]. We compare performance among sensors to evaluate the efficiency of each data type for use in automated sinkhole mapping and use traditional geomorphological methods of field surveys and visual image interpretation to assess the advantages and limitations of the automated technique. 2. Study Area Brazil has 425,000 600,000 km 2 of limestone rocks in different biomes  and knowledge of karst areas has been reinforced by speleological studies and investigations of biological, paleoenvironmental, paleontological, and archaeological attributes. Karmann & S nchez s  classified speleological provinces in Brazil based on common geological histo ry, stratigraphic associations with carbonate and pelitic sequences, and thickness and extension of carbonate rocks. We analyzed a small area of the Bambu Speleological Province in Central Brazil. This province is underlain by rocks of the Bambu Group [3 7], a Neoproterozoic sedimentary sequence that records at least two transgressive regressive cycles in the epicontinental basin and possibly deposition in a foreland basin along the west side of the S o Francisco Craton during the Brasiliano orogeny [38 41 ]. The study area is located in Bahia State, northeastern Brazil (Figure 1) where the tropical environment favors karst formation due to the growth of vegetation and biochemical activity increasing water acidity and promoting the development of vertical fl ow and sinkhole (doline) development. The study area has a high density of dolines associated with the exposure of karstified limestone. The stagnant surface water and shallow groundwater mostly reside in the sinks, forming lakes. Herbaceous vegetation dom inates in topographic depressions because trees have low tolerance to shallow
Remote Sens. 2014 6 333 groundwater. Thus, dolines commonly are floored by open water or herbaceous vegetation (Figure 2). The vegetation cover in the study area therefore has little potential influence on the results from different sensor types unlike other areas with forest Figure 1. Study area loca t ion. Figure 2. Karst depression covered by herbaceous vegetation and stagnant surface water. 3. Material and Methods We compared three different DEMs ASTER GDEM, ALOS/PRISM DEM and SRTM DEM all processed in four steps the first of which was DEM acquisition and evaluation of the best DEM for our purposes. We next identified closed depressions or sinks in the DEMs and digitally filled them by inte rpolation from neighboring elevations outside the depression polygons. We calculated sink depth s as the difference between the original and processed DEM s. The final step was the elimination of false detections using morphometric attributes and visual interpretation overlaying the depression vectors on the high resolution remotely sens ed image.
Remote Sens. 2014 6 334 3.1. Digital Elevation Model We analyzed the potential of DEMs for mapping closed depressions using data from different sensors with various data acquisition methods and spatial resolution: ASTER GDEM (30 m), ALOS/PRISM DEM (5 m) and SRTM DEM (90 m) The ASTER GDEM was generated using ASTER Level 1A bands 3N (nadir viewing) and 3B (backward viewing) images from the Visible/Near Infrared (VNIR) sensor. The VNIR subsystem consists of nadir and rear viewing telescopes looking 0 and 27.7 backwards that allow the generation of stereoscopic data with a time lag around one minute [ 42 ] The Band 3 stereo pair is acquired in the spectral range of 0.78 0.86 m with a base to height ratio of 0.6. In 2006, LP DAAC implemented software based on an automated stereo correlation method that utilizes the ephemeris and attitude data derived from both the ASTER instrument and the Terra spacecraft platform. This method generates a relative DEM without any ground control points (GCPs). The ASTER GDEM is an image product with a horizontal resolution of 1 arc second (30 m) referenced to the UTM coordin ate system, and referenced to Earth s geoid using the EGM96 geopotential model. This product is generated from a stereo pair of images using the SilcAST software and covers the earth s surface between 83 N and 83 S, encompassing 99 percent of Earth s landm ass. The SRTM flown on Space Shuttle Endeavour in February 2 000 carried in the cargo bay two synthetic aperture radars, a C band system (5.6 cm; C RADAR) and an X band system (3.1 cm; X RADAR) [ 43 ]. Radar data are less sensitive to cloud cover than optic al data. Topographic data were acquired from a single flight covering 80% of Earth's land surface in just 11 days, between the latitudes 60 N and 57 S. The flyover produced three dimensional models with spatial resolutions of 1 arc sec (30 m) and 3 arc sec (90 m) using WGS84 horizontal datum and vertical datum WGS84/EGM96. Vertical accuracy was on the order of 5 m [ 44 ]. The continuous data acquisition ( i.e ., day and night regardless of clouds, which are transparent to the RADAR) ensured homogeneous data throughout the globe, making the SRTM DEM an important tool for studies of the land surface [ 44 46 ]. SRTM DEM data have been widely used for geomorph o logica l studies [ 47,48 ]. The ALOS was launched on 24 January 2006 by the Japan Aerospace Exploration Agency (JAXA) with PRISM on board, which acquires images with spatial resolution of 2.5 m. PRISM produces triplet images that achieve along track stereoscopy by three independent cameras for viewing nadir, forward and backward where the images are acquired in the same orbit at almost the same time [41 49 ]. The nadir looking radiometer can provide coverage 70 km wide, and the forward looking and backward looking ra diometers each provide coverage 35 km wide [ 50 ]. Several studies assess the absolute vertical accuracy ( relationship between DEM elevation and true elevation relative to an established vertical datum) of the ALOS/PRISM DEM. Gruen et al. [ 51 ] compared the DEM accuracy of ALOS/PRISM data with other satellite or ground control points and found th at its accuracy is similar to that obtained by SPOT 5, IKONOS and QuickBird data. Kocaman and Gruen [ 52 ] also found similar results for IKONOS, but rep orted a lower accuracy compared to SPOT 5 results. Maruya and Ohyama [ 49 ] used ground control points derived from the 1:25,000 map ping to analyze elevation accuracy and found a 6.2 m mean error and a 4.8 m RMSE. Saunier et al. [ 53 ] verified that the accura cy in height for the ALOS/ PRISM DEM is about 1 m using either five or nine ground control points. In Brazil ALOS/PRISM DEM was tested in an area with steep slopes and compar ison of the
Remote Sens. 2014 6 335 results to ground control points (GCP), indicated an accuracy comparabl e to 1:25,000 scale mapping [ 54 ]. Although the studies described above address the absolute vertical accuracy, for this study the relative vertical accuracy is most important and is obtained from a vertical difference between two points, i.e. a measure of the point to point vertical accuracy within a specific dataset. T he absolute vertical error is greater than the relative vertical error, establishing a n upper limit for empirical evaluation of relative vertical accuracy The stereo DEM extrac tion from ALOS/PRISM data was done using the commercial software, PCI Geomatica orthoengine. Artifacts in the data were assessed by visual inspection and specific algorithms [55,56] that reveal errors on the image border and in areas with cloud cover. Imag e border errors were eliminated from the image during resizing by simply removing the noisy strip. The cloud cover is a limitation of this type of data and anomalous values are easily identified by digital processing from a threshold value and thus can be identified and discarded. The scenes used in this study have few and isolated clouds. 3.2. Sink Depth Image The methodology we used to determine the sink depth of closed depressions involved two DEM operations. The first step used the Fillsink algorithm from the ArcMap software package [ 25 ] that identifies the point or set of adjacent points surrounded by neighbors with higher elevation and rises to the lowest value on the depressions boundary. This procedure then fills all depressions in the D EM, including both those generated from data errors (spurious artifacts) and those that record real topographic features, such as karst depressions (dolines). The second step was to extract the sink depths in these areas by differencing the maps between th e sink filled ( depressionless DEM) and original DEM ( Figure 3 ). The difference image ( Figure 3C ) highlights the different depressions, including the karst enclosed depressions. A b inary image is generated from the sink depth image where the depressed are as have value 1, while all other areas have value 0. This binary image is then converted to vector format. The minimum area of depressions corresponds to the spatial resolution of the sensor However, the polygons show both natural features as well as pits from surface imperfections. Thus, the vectors need to be checked in order to eliminate the errors. Thus, the key issue is to establish criteria to separate the dolines from the spurious artifacts and other types of depression ( e.g., reservoirs or quarries ). In this paper, the delimitation of the spurious depressions is derived from threshold values of morphometric attributes, specifically depth size, and shape. Evaluation of appropriate threshold values to represent the boundary between dolines and the su rrounding landscape ( no dolines ) was obtained by comparing maps of identified dolines with previous mapping of dolines from field validation and interpretation of higher spatial resolution imagery (ALOS PRISM and Google Earth images). The karst features investigated in the study area are easily identified by visual interpretation, as they are characterized by natural moist grassy vegetation where the water table approaches the surface for part of the year, leading to a striking difference in visual appearance from the surrounding vegetation
Remote Sens. 2014 6 336 Figure 3. Methodological procedures to determine the terrain attribute of sink depth. D igital elevation model ( A ) (DEM) Fillsink minus ( B ) original DEM results in the ( C ) sink depth distribution. We used a r ange of different empirical threshold values for the minimum sink depth to identify the best threshold value from the maximum accuracy index between manual and automated classification. In assessing classification and change detection techniques using remot ely sens ed images, the threshold values for the delimitation of classes are commonly identified from overall accuracy and a Kappa index [ 57 59 ]. In this paper we applied the overall accuracy (OA) Performance in identifying true sinkholes is assessed through the intersection of reference and classified polygons Usually, the polygons obtained by the two methods show distinctions in the dimensions and shapes ; but conventional accuracy analysis typically consider s only the number of the overlapping polygons Overall accuracy is calculated by summing the number of polygons classified correctly (True Positive TP) and dividing by the total number of polygons: OA = TP/(TP + FP + FN) where the number of polygons misclassified is determined by summing the number of False Negatives (FN) ( i.e ., no doline identified where one is actually present) and False Positives (FP) ( i.e doline indicated where none exists). This accuracy analysis does no t directly assess true negative. Inevitably, using the proposed method a mapped doline will be represented by more than one polygon. Redundant data (R) must be considered in computing the accuracy index in order to avoid the overestimation. We used the following equation:
Remote Sens. 2014 6 337 Therefore comparison between doline classifications with previous maps allows the determination of threshold values for delimiting karst depressions in similar regions and reveals the uncertainties o f the method. 3.3. Morphometric Analysis Morphometric analysis may be used to improve the accuracy of predicting doline occurrence and eliminate false dolines Many of the false dolines can be easily eliminated as being incompatible with known characteristics of karst depressions in the study area ( Figure 4 ). We used the following morphometric attributes to automate delineation of doline polygons: area, perimeter and c ircularity index (C I) The area and perimeter of the polygons are automatically added to topological vector data structure. We defined a C I based on area and perimeter values using the following equation: C I = 4 A/P 2 where A is the area of a polygon and P is its perimeter. A circular shape is represented by value of 1.0, i.e. the maximum value. In contrast, elongated shapes are represented by lower values. The more circular polygons indicate locations of karstic negative relief elements. I n addition to the C I values, average and maximum depth were extracted from the DEM for each polygon. Figure 4. Hypothetical dolines enlargement from threshold depths equal to 1 m and shallow ( i.e > 0 m). Visual inspection revealed that erroneous polygons generally had a small area, low circularity and were shallow suggesting that undesirable polygons can be eliminated using threshold value criteria for these morphometric attributes. We superimposed vector s on Google Earth images to identify polygons that represent correctly karst depressions and thereby evaluate the range of morphometric values suitable for use as threshold criteria. The dolines obtained by morphometric attributes were compared with a refe rence map using traditional geomorphological procedures such as field surveys and visual interpretation of high resolution remote sensing image from ALOS/PRISM and Google Earth. In this comparison are evaluated quantities of correct predictions (TP), Type I errors (FP) and Type II errors (FN).
Remote Sens. 2014 6 338 3.4. Enlargement of Doline Polygons A greater average depth threshold decreases the number of false dolines. However these deeper thresholds do n o t give the real plan view area of the dolines. In order to correct f or this error, a new mask with a threshold valu e > 0 m was built and applied to the true dolines previously selected eliminating the noisy polygons In doing this, the polygons generated from the depth data are replaced by polygons generated using the shallow depth threshold This procedure allows having a real area of the doline to be compared with the areas calculated in visual interpretation. However the exact polygon representation is inherently imprecise because the landform itself is hard to defin e precisely, i.e. the closer you look, the less well defined the edge becomes Figure 4 illustrates the hypothetical area of an idealized doline using different thresholds, and how the area decreases with doline depth 4. Results 4.1. Results of the Visual Interpretation The reference map was built from the visual interpretation of ALOS/PRISM image (2.5 m); Google Earth images, and detailed field validation. This mapping identified 249 dolines ( Figure 5), with the largest sinkholes located on interflu ves and smaller sinkholes located closer to the river networks. The comparison between these mapped dolines with those identified through automated classification enables evaluating different threshold values for delimiting karst depressions in similar reg ions and reveals the uncertainties of the method. F igure 5 Reference map from the visual interpretation of ALOS/PRISM and Google Earth images. 4.2. Results of the Sink Depth Image The distribution and depths of closed topographic depressions detected u sing ASTER GDEM SRTM and ALOS data reflected the different spatial resolutions and patterns of noise. For each threshold depth a binary mask image was generated and then represented with polygons. The sink depth images were classified using threshold values ranging from 1 m. The best threshold for
Remote Sens. 2014 6 339 each DEM is defined from the overall accuracy compared to the entire reference map. Therefore a set of classified images made with different thresholds was te sted to obtain the optimal value. Our analysis shows that in our study area, the ASTER GDEM is susceptible to noise leading to inaccurate results and visual anomalies and artifacts that represent barriers to its effective utilization for doline detection ( Figure 6 ). Beyond residual cloud anomalies, the ASTER GDEM has a variety of pervasive artifacts that appear as stripes and other geometric shapes. A significant number of holes in ASTER GDEM exceed tens of meters and do not correspond to natural dolines ( Figure 6 ). Other studies conducted in karst areas by Guimar es et al.  and Siart et al.  described difficulties in using the ASTER model in flat areas and in areas with sloping terrain or relief Previous assessments of ASTER GDEM accuracy highligh t anomalies that prevent its immediate use for a wide range of applications [ 60,61 ]. Thus, ASTER GDEM data are disregarded for further analysis due to their poor performance in karst depression detection Figure 6. ( A ) Depth image and ( B ) its derived binary mask made from ASTER GDEM data using threshold value of 1 m. Figure 7 shows the depth image made from SRTM DEM threshold values of 1, 2 and 3 m, and a sequence of derived binary masks. A ccuracy analyses fo r different threshold values are presented in Table 1 However, Table 1 presents only the main depth threshold values in order to demonstrate the convergence to a maximum of overall accuracy; where from it either higher or lower values decrease the accuracy. Using a 1 m threshold depth value resu lted in the detection of 1 246 polygons, where the majority is no dolines (false positives). In this map 256 true dolines are identified (more than the 249 mapped manually) of which 25 are redundant ( i.e more than one predicted doline demarcates the same natural doline). The increased depth sink threshold provides a reduction in the number of predicted dolines and false positives and an increased in false negatives. The sink depth threshold that had the best overall accuracy was the de pth of 3 m (0.4 9 ), where the errors of omission and commission are approximately equivalent.
Remote Sens. 2014 6 340 Figure 7. Depth image from ( A ) SRTM DEM and binary mask images made with sink depth thresholds of ( B ) 1 m, ( C ) 2 m and ( D ) 3 m. Table 1. Dolines predicted from SRTM DEM data using different threshold values for sink depth. The comparison is made with 249 true dolines obtained by visual interpretation. Depth (m) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negati ves Overall Accuracy 1 1 246 256 25 990 18 0.19 2 521 240 35 281 44 0.36 3 297 204 38 93 83 0.49 4 205 152 20 53 117 0.44 5 125 111 9 14 147 0.39 Figure 8 shows the depth image from the ALOS/PRISM DEM and binary masks made with the same three threshold values. Table 2 shows the threshold values for each sink depth using ALOS/PRISM DEM and its corresponding overall accuracy. The best fit results are obtained for deeper than 3 m, where the overall accuracy is 0.43. Thus threshold value for the ALOS/PRISM DEM is similar to that for the SRTM DEM and both perform well in identifying medium size landforms. ALOS/PRISM DEMs have limitations in the presence of clouds and SRTM DEMs have problems with fine scale distortions. In this paper, the image used had a small cloudy area that was disregarded. The radar based SRTM DEM does not have this problem.
Remote Sens. 2014 6 341 Figure 8. Depth image from ( A ) ALOS/PRISM DEM and binary mask image s made considering sink depth thresholds of ( B ) 1 m, ( C ) 2 m and ( C ) 3 m. Table 2. Dolines predicted using different threshold values for sink depth from ALOS/PRISM DEM data. The comparison is made with true dolines ( 249) obtained by visual interpretation. Depth (m ) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy 1 1 276 271 45 1 005 23 0.18 2 590 254 59 336 54 0.33 3 348 205 38 143 82 0.43 4 230 160 38 70 127 0.38 5 176 133 36 43 152 0.33 4.3. Results of the Morphometric A nalysis Additionally, in order to increase the overall accuracy we developed a morphometric analysis that considers area and CI attributes. At this stage we analyzed the polygons of sinkholes with depths greater than 3 m ( depth with highest accuracy among those tested with the SRTM DEM ) and 2 m (for reconciling high value of overall accuracy and high number of true positives). Fi gure 9 shows the morphometric attributes for the depression polygons deeper than 2 m (since it includes all polygons with depth values more than 3 m).
Remote Sens. 2014 6 342 Figure 9. Spatial distribution of dolines considering the following attributes: area from ( A ) SRTM DEM and ( B ) ALOS/PRISM DEM and circularity indexes from ( C ) SRTM DEM and ( D ) ALOS/PRISM DEM Initially, the threshold analysis was done for the area and then for CI In this text, tables of morphometric attributes show only the main results highlighting the convergence to an optimal threshold value with maximum overall accuracy For area size analysis from SRTM DEM, the best overall accuracy (0.52) was obtained from the following conditions: depth > 2 m and area > 16,200 m 2 (Table 3). This combination resulted in values of overall accuracy greater than obtained for depth > 3 m (0.49) ( Table 1 ) or supplemented by the restriction of area > 8,100 m 2 (0.49) ( Table 4 ). The area attribute eliminated mainly small spurious artifacts (False Positives and False Negatives). The threshold analysis for the CI allowed a little improvement in the already obtained overall accuracy. Assuming a CI > 3 resulted in an overall accuracy of 0.5 3 ( Table 5 ). This procedure eliminated large and long polygons corresponding mainly to fluvial features Thus the best overall accuracy for the doline delimitation using SRTM DEM has the following constraints : depth > 2 m ; area > 16,200 m 2 and CI > 0.3.
Remote Sens. 2014 6 343 Table 3. Dolines predicted using the area attribute derived from the doline depth > 2 m (SRTM DEM). Comparison is made with the actual dolines ( 249) mapped by visual interpretation. Area (m 2 ) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >8,100 293 189 7 104 67 0.52 >16,200 229 164 1 65 86 0.52 >24,300 179 143 0 36 106 0.50 >32,400 158 132 0 26 117 0.48 Table 4. Dolines predicted using the area attribute calculated from the doline depth > 3 m (SRTM DEM). The comparison is made with actual dolines ( 249) mapped by visual interpretation. Area (m 2 ) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >8,100 190 150 7 40 106 0.49 >16,200 157 131 1 26 119 0.47 > 24,300 123 110 1 13 140 0.42 >32,400 108 96 0 12 153 0.37 Table 5. Predicted dolines using SRTM DEM and circularity index attribute for dolines depth > 2 m and area > 16,200. The comparison is made with true dolines ( 249) obtained by visual interpretation. CI Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >0.3 224 164 1 60 86 0.53 >0.35 221 162 1 59 88 0.52 >0.4 216 160 1 56 90 0.52 ALOS/PRISM DEM demonstrated that the best overall accuracy was obtained by the following combination: area > 16,200 m 2 and depth > 2 m (0.53) (Table s 6 and 7 ). This overall accuracy was better than simply restricting depth to > 2 m (0.33) or depth to > 3 m (0.4 3 ) ( Table 2 ). The best overall accuracy from the ALOS/PRISM DEM was slightly higher than that for the SRTM DEM Table 6. Predicted dolines using area attribute from dolines depth > 2 m (ALOS/PRISM DEM). The comparison is made with true dolines (249) obtained by visual interpretation. Area (m 2 ) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >8,125 263 176 8 87 81 0.50 >10,625 237 170 6 67 85 0.52 >13,125 211 162 4 49 91 0.53 > 15,625 195 153 0 42 96 0.53 >18,125 182 146 0 36 103 0.51
Remote Sens. 2014 6 344 Table 7. Predicted dolines using area attribute from dolines depth > 3 m (ALOS/PRISM DEM). The comparison is made with true dolines ( 249) obtained by visual interpretation. Area (m 2 ) Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >625 316 190 26 126 85 0.44 >3,125 225 155 8 70 102 0.46 >5,625 179 138 3 41 114 0.47 >8,125 157 125 0 32 124 0.44 >10,625 144 116 0 28 133 0.42 As for the SRTM DEM the threshold analysis of CI in the ALOS/PRISM DEM allowed a slight improvement in the overall accuracy (0.53) ( Table 8 ). At this morphometric attribute, both the DEMs showed the same threshold value (0.3). Thus, the best configuration for ALOS/PRISM DEM considers the following constrains: depth > 2 m; area > 13 ,125 m 2 and IC > 0.3. Table 8. Predicted dolines using ALOS/PRISM DEM and circularity index attribute for dolines depth > 2 m and area > 13,125 m 2 The comparison is made with true dolines ( 249) obtained by visual interpretation. CI Numbers of Predicted Dolines True Positives Redundant Data False Positives False Negatives Overall Accuracy >0.3 207 161 4 46 92 0.53 >0.35 201 157 4 44 96 0.52 >0.4 197 153 3 44 99 0.51 Figure 10. Comparison between delimited dolines from depth > 1 (blue polygon) and using the morphometric analysis (red polygon) that provides a significant decrease of sinkholes. ( A ) SRTM DEM and ( B ) ALOS/PRISM DEM. Once appropriate threshold values are defined from a reference area, they can be applied to other similar locations. Morphometric analysis ide ntified little more than half of the actual dolines, which may reduce the need for visual interpretation. Figure 10 shows all polygons > 1 m initially obtained
Remote Sens. 2014 6 345 with sink depth images and the polygons obtained after the restrictions by morphometric attribut es for DEMs from SRTM and ALOS/PRISM. The differences in spatial resolution of the DEMs generate singular shapes for the depressions. Due to the high resolution of the ALOS/PRISM image, the depressions are better defined in them, whereas in the SRTM DEM the depression outlines are less detailed. Despite shape differences the SRTM DEM indicates the great majority of the points described by ALOS/PRISM DEM 4.5. Results of the Adjustment of the Size of Dolines Larger doline polygons with better overall acc uracy from SRTM and ALOS data are associated with a threshold depth of 2 m. This procedure increased the area s of sinkholes, making them more consistent with visual interpretation. Figure 11 shows in detail the effect caused for some sinkholes. Figure 11. The sinkholes obtained from the morphometric attributes have increased their sizes considering a shallow depth from the ( A ) SRTM DEM and ( B ) ALOS/PRISM DEM 4.6. Spatial Distribution of Dolines An analysis of the spatial distribution of morphometric attributes shows a correlation with structures and geological settings. The deepest depressions are connected to the main drainage network. In particular, the largest dolines show a NE SE trend in the interfluve region following the preferred direction of regional geological structures and the drainage network. With regard to morphological manifestations, two landscape types are distinguished in the study region: the upland karst and the incised river valleys. The plateau karst shows inter relationships be tween surface and sub surface drainage, developing karstic processes. The upland surfaces exhibit a parallel drainage pattern, with dolines concentration along the interfluves. The dolines are all on the undissected upper (older) landscape surface, indicat ing that dolines increase in number through time, as observed in other tropical areas . The lower elevation dissected surface is a region of well developed fluvial features, where few dolines occur on river terraces and low elevation, erosion surfaces.
Remote Sens. 2014 6 346 5. Conclusions Semi automated landform classification using DEMs provides several advantages: fast acquisition of data over large areas at low cost, analysis of inaccessible zones, reduction of human errors by eliminating manual classification steps, read y comparison of results derived from different datasets, and the reduction in processing time. Our methodology for mapping karst depressions combines morphometric analysis with a minimum threshold value for the sink depth to identify karst depressions. How ever, some depressions selected by sink depth are actually data errors introduced in generating the surface and should be interpreted with prudence We suggest a method for the separation of real from incorrectly identified topographic features. The identification of the optimal threshold values for assessing morphometric attributes was made from the highest overall accuracy, considering a reference map built from visual interpretation. These threshold values are limited to areas with similar envi ronmental conditions. Thus studies in new locations must first define the best thresholds using a test area. In this work, the ASTER GDEM had severe limitations in the detection of karst features. However, this is not a general criticism of the ASTER GDEM, which may perform better in other areas or other problems. The GDEM data generation process consists of scene selection, scene division, same path mosaic generation, stacking, sticking, and filtering for both elevation data and water body data . Despi te this processing some images may contain imperfections, which are being fixed in another version of the ASTER GDEM product. Therefore, the doline detection considering other areas or new products may get better results. SRTM DEM and ALOS/PRISM DEM identi fy more than half of the true dolines with an overall accuracy of approximately 0.5 The ALOS/PRISM DEM has potentially serious limitations in the presence of clouds, whereas the SRTM DEM is insensitive to cloud cover. However, the coarser resolution of th e SRTM DEM limits its utility in delineating fine scale karst features. Practically all the larger sinkholes detected by ALOS were also identified by the SRTM. Therefore SRTM and ALOS data are reasonably well suited for use in large scale mapping of karst features in Central Brazil. The algorithm results could be improved by considering three distinct aspects: (a) the addition of other DEM attributes or polygon geometry; (b) the use of DEM with higher spatial resolution (e.g. airborne laser scanning ) ; and (c) the application of other complementary data to the DEM, such as multi spectral images. In detecting sinkholes other morphometric information can be used, providing an additional component for spatial analysis Siart et al.  used the following DEM attributes to assist doline detection: slope, aspect and curvature. However the results were difficult to interpret, due to the visual complexity of the image, and did not contribute significantly to doline detection. With respect to the polygon geomet ry the eccentricity of an ellipse remains to be tested , although for the present study area this should help little, because the dolines are circular. Thus, tests with other morphometric attributes may provide a little improvement in accuracy of the a lgorithm in the general application. The main improvements probably lie in the other two mentioned alternatives. The use of images with higher spatial resolution undoubtedly enables better detection of sinkholes. A particularly promising area for future st udy is the acquisition of the morphometric attributes from LiDAR (Light Detecting and Ranging) data, which has impressive accuracy for detecting sinkholes
Remote Sens. 2014 6 347 and estimating subsidence rates [65 68]. Another major advantage of utilizing laser scanning technolo gy is its canopy penetration ability. However, many areas do not have this type of data. Although beyond the scope of this paper, the use of classification of multi spectral images can significantly increase the detection accuracy of dolines. Remote sensin g data, whether airborne or satellite based, offer an alternative way for the doline detection considering karst dependent features (e.g. sedimentary infills and vegetation). Thus the depressions can locally modify environmental variables: soil moisture, water stagnation, soil erosion and sedimentation, and spatial distributions of the vegetation with relative presence of dominant species. In this approach, some works [25,26] demonstrates that the combination of DEMs and high resolution satellite imagery p roduce the best results in karst geomorphological research. Acknowledgments The authors are grateful for the support of the following institutions: Funda o de Empreendimentos Cient ficos e Tecnol gicos (FINATEC), for logistical support; and Conselho Nacional de Desenvolvimento Cient fico e Tecnol gico (CNPq) for s ponsoring research grants. Conflicts of Interest The authors declare no conflict of interest. References 1. Galve, J.P.; Guti rrez, F.; Lucha, P.; Guerrero, J.; Bonachea, J.; Remondo J.; Cendrero, A. Probabilistic sinkhole modelling for hazard assessment. Earth Surf. Proc. Landf. 2009 34 437 452. 2. Guti rrez, F.; Galve, J.P.; Guerrero, J.; Lucha, P.; Cendrero, A.; Remondo, J.; Bonachea, J.; Guti rrez, M.; S nchez J.A. The origin, typology, spatial distribution, and detrimental effects of the sinkholes developed in the alluvial evaporite karst of the Ebro River valley downstream Zaragoza city (NE Spain). Earth Surf. Proc. Landf. 2007 32 912 928. 3. Soriano, M.A.; S im n, J.L. Alluvial dolines in the central Ebro Basin Spain: A spatial and developmental hazard analysis. Geomorphology 1995 11 295 309. 4. White, E.L.; Gert, A.; White, W B. The influence of urbanization in sinkhole development in central Pennsylvania. En viron. Geol. Water Sci. 1986 8 91 97. 5. Soriano, M.A. Characteristics of the alluvial dolines developed because of gypsum dissolution in the central Ebro Basin. Z. Geomorphol. Suppl. 1992 85 59 72. 6. Casta eda, C.; Guti rrez, F.; Manunta, M.; Galve, J.P. D InSAR measurements of ground deformation by sinkholes, mining subsidence, and landslides, Ebro River, Spain. Earth Surf. Proc. Landf. 2009 34 1562 1574. 7. Ford, D.C.; Williams, P.W. Karst Geomorphology and Hydrology ; Unwin Hyman: London, UK 1989. 8. Trudgil l, S. Limestone Geomorphology ; Longman Group: New York, NY, USA, 1985. 9. Kresic, N. Remote sensing of tectonic fabric controlling groundwater flow in Dinaric Karst. Remote Sens. Environ. 1995 53 85 90.
Remote Sens. 2014 6 348 10. Florea L. Using state wide GIS data to identify the coincidence between sinkholes and geologic structure. J. Cav. Karst. Stud. 2005 67 120 124. 11. Closson, D.; Karaki, N.A. Salt karst and tectonics: Sinkholes development along tension cracks between parallel str ike slip faults, Dead Sea, Jordan. Earth Surf. Proc. Landf. 2009 34 1408 1421 12. Doerfliger, N.; Jeannin, P.Y.; Zwahlen, F. Water vulnerability assessment in karst environments: A new method of defining protection areas using a multi attribute approach and GIS tools (EPIK method). Environ. Geol. 1999 39 165 176. 13. Van Stempvoort, D.; Evert, L.; Wassenaar, L. Aquifer vulnerability index: A GIS compatible method for groundwater vulnerability mapping. Can. Water Resour. J. 1993 18 25 37. 14. Quinn, P.; Beven, K. ; Chevallier, P.; Planchon O. The prediction of hillslope flow paths for distributed hydrological modelling using digital terrain models. Hydrol. Proc. 1991 5 59 79. 15. Zhang, W.; Montgomery, D.R. Digital elevation model grid size, landscape representation and hydrologic simulations. Water Resour. Res. 1994 30 1019 1028. 16. Tucker, G.E.; Hancock, G.R. Modelling landscape evolution. Earth Surf. Proc. Landf. 2010 35 28 50. 17. Carvalho, J nior, O.A.; Guimar es, R.F.; Freitas, L.; Gomes Loebmann, D.; Gomes, R.A.; Martins, E.S.; Montgomery, D.R. Urbanization impacts upon catchment hydrology and gully development using mutli temporal digital elevation data analysis. Earth Surf. Proc. Landf. 2010 35 611 617. 18. Bolongaro Crevenna, A.; Torres Rodriguez, V.; Sorani, V.; Frame, D.; Arturo Ortiz, M. Geomorphometric analysis for characterizing landforms in Morelos State, Mexico. Geomorphology 2005 67 407 422. 19. Podobnikar, T. Detecting mountain peaks and delineating their shapes using digital elevation models remote sensin g and geographic information systems using autometric methodological procedures Remote Sens 2012 4 784 809. 20. Wood, J.D. The Geomorphologic Characterization of Digital Elevation Models PhD Thesis, University of Leicester, London, UK 1996. 21. Segura, F.S.; Pardo Pascual, J.E.; Rossell V.M.; Forn s, J.J.; Gelabert, B. Morphometric indices as indicators of tectonic, fluvial and karst processes in calcareous drainage basins, South Menorca Island, Spain. Earth Surf. Proc. Landf. 2007 32 1928 1946. 22. Lyew Ayee P.; Viles H.A.; Tucker, G.E. The use of GIS based digital morphometric techniques in the study of cockpit karst. Earth Surf. Proc. Landf. 2007 32 165 179. 23. Piccini, L.; Landelli, N. Tectonic uplift, sea level changes and Plio Pleistocene evolution of a coastal karst system: the Mount Saint Paul (Palawan, Philippines). Earth Surf. Proc. Landf. 2010 36 594 609. 24. Antonic, O.; Hatic, D.; Pernar, R. DEM based depth in sink as an environmental estimator. Ecol. Model. 2001 138 247 254. 25. Guimar es, R.F.; Carvalho Junior, O.A.; Souza Martins, E.M.; Carvalho, A.P.F.; Gomes, R.A.T. Detection of karst depression by ASTER image in the Bambui Group, Brazil. Proc. SPIE : 2005 ; 5983 pp. 328 339.
Remote Sens. 2014 6 349 26. Siart, C.; Bubenzer, O.; Eitel B. Combining digital elevation data (SRTM/ASTER), high resolution satellite imagery (Quickbird) and GIS for geomorphological mapping: A multi component case study on Mediterranean karst in Central Crete. Geomorphology 2009 112 ,106 121. 27. Pardo Ig zquiza, E.; Valsero, J.J.D.; Dowd, P.A. Automatic detection and delineation of karst terrain depressions and its application in geomorphological mapping and morphometric analysis. Acta. Carsologica. 2013 42 17 24 28. Marks, D.; Dozier, J.; Frew, J. Automated basin delineation from digital elevation data. Geo Processing 1984 2 299 311. 29. Jenson, S.K.; Domingue, J.O. Extracting topographic structure from digital elevation data for geographical information system analysis. Photogramm. Eng. Remote Sens. 1987 54 1593 1600. 30. Tianqi, A.; Takeuchi, K.; Ishidaira, H.; Yoshitani, J.; Fukami, K. Development and application of a new algorithm for automated pit removal for grid DEMs. Hydrol. Sci. J. 2003 48 985 997. 31. Soille, P.; Vogt, J.; Colombo, R. Carving and adaptive drain age enforcement of grid digital elevation models. Water Resour. Res. 2003 39 1366 1375. 32. Grimaldi, S.; Nardi, F.; Benedetto, F.; Istanbulluoglu, E.; Br s, R.L. A physically based method for removing pits in digital elevation models. Adv. Water Resour. 200 7 30 2151 2158. 33. Obu, J.; Podobnikar, T. Algorithm for karst depression recognition using digital terrain models. Geodetski Vestnik 2013 57 260 270. 34. Kass, M; Witkin, A; Terzopoulous, D. Snakes: Active contour models. Int. J. Comput. Vis. 1988 1 321 331. 35. Rahimi, M.; Alexander, E.C., Jr. Locating Sinkholes in LIDAR Coverage of a Glacio Fluvial Karst Winona County, MN. In Proceedings of the 13th Multidisciplinary Conference on Sinkholes and the Engineering and Environmental Impacts of Karst, Carlsbad, CA USA, 1 May 15 August 2013 ; pp. 469 480. 36. Karmann, I. Evolu o Din mica Atual do Sistema C rstico do Alto Vale do Ribeira de Iguape, Sudeste do Estado de S o Paulo; S o Paulo. Ph D Thesis, U niversidade de S o P au lo S o Paulo, Brazil, 1994. 37. Karmann, I.; S nchez, L.E. Distribui o das rochas carbon ticas e prov ncias espeleol gicas do Brasil. Espeleo Tema 1979 13 105 167. 38. Carvalho J nior, O.A.; Berbet Born, M.; Martins, E.S.; Guimar es, R.F.; Gomes, R.A.T. Ambientes C rsticos. In Geomorfologia: Conce itos e Tecnologias Atuais ; Florenzano, T.G., Ed.; Oficina de Textos: S o Paulo, Brazil, 2008; Volume 1, pp. 183 218. 39. Castro, P.T.A. Os Conglomerados da Borda SW do Cr ton do S o Francisco Junto a Por o S da Faixa Bras lia: Sedimentologia e Rela es Estratigr ficas com as Rochas do Grupo Bambu Ph D Thesis Universidade de Bras lia, Bras lia, Brazil, 1997. 40. Castro, P.T.A.; Dardenne, M.A. The sedimentology stratigraphy and tectonic context of the S o Francisco Supergroup at the southern boundary of th e S o Francisco craton, Brazil. Rev. Brasil. Geocincia. 2000 30 345 437. 41. Martins Neto, M.A.; Pedrosa Soares, A.C.; Lima, S.A.A. Tectono sedimentary evolution of sedimentary basins from Late Paleoproterozoic to Late Neoproterozoic in the S o Francisco Cr aton and Ara ua Fold belt, Eastern Brazil. Sediment. Geol. 2001 141/142 343 370.
Remote Sens. 2014 6 350 42. Abrams, M. The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER): Data products for the high spatial resolution imager on NASA s Terra platform. Int. J. Remote Sens. 2000 21 847 859. 43. Farr, T.G.; Rosen, P.A.; Caro, E.; Crippen, R.; Duren, R.; Hensley, S.; Kobrick, M.; Paller, M.; Rodriguez, E.; Roth, L.; et al. The Shuttle Radar Topography Mission. Rev. Geophys. 2007 doi: 10.1029/2005RG000183. 44. Smith, B.; Sandwell, D. Accuracy and resolution of Shuttle Radar Topography Mission data. Geophys. Res. Lett. 2003 30 1467 45. Van Zyl, J.J. The Shuttle Radar Topography Mission (SRTM): A breakthrough in remote sensing of topography. Acta. Astronaut. 2001 doi: 1 0.1016/S0094 5765(01)00020 0. 46. Rabus, B.; Eineder, M.; Roth, A.; Bamler, R. The shuttle radar topography mission A new class of digital elevation models acquired by spaceborne radar. Photogramm. Eng. Remote Sens. 2003 57 241 262. 47. Rossetti, D.F. Multiple r emote sensing techniques as a tool for reconstructing late Quaternary drainage in the Amazon lowland. Earth Surf. Proc. Landf. 2010 35 1234 1239. 48. Petersen, G.; Sutcliffe, J.V.; Fohrer, N. Morphological analysis of the Sudd region using land survey and re mote sensing data. Earth Surf. Proc. Landf. 2008 33 1709 1720. 49. Maruya, M.; Ohyama, H. Accurate DEM and Ortho Rectified Image Production from ALOS/PRISM. In Proceedings of the 2008 IEEE International Geoscience and Remote Sensing Symposium, Boston, WA, USA 7 11 July 2008 ; pp. 185 188. 50. Tadono, T.; Shimada, M.; Watanabe, M.; Rosenqvist, A. Overview of ALOS research and science program. Proc. SPIE 2004 doi: 10.1117/12.565485 51. Gruen, A.; Kocaman, S.; Wolff, K. Calibration and validation of EARLY ALOS/PRISM images J. Jpn. Soc. Photogramm. Remote Sens. 2007 46 24 38. 52. Kocaman, S.; Gruen, A. Orientation and Calibration of ALOS/PRISM Imagery In Proceedings of the 2007 ISPRS Hannover Workshop: High Resolution Earth Imaging for Geospatial Information, Hanover Germany, 29 May 21 June 2007; p. 6. 53. Saunier, S.; Santer, R.; Goryl, P.; Gruen, A.; Wolf, K.; Bouvet M.; Viallefont, F.; Chander, G.; Rodriguez, Y.; Mambimba, A. The Contribution of the European Space Agency to the ALOS PRISM/AVNIR 2 Commissioning Phase In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Barcelona, Spain, 1 July 2007 ; pp. 208 211. 54. Brazilian government Avalia o Planialtim trica de Dados ALOS/PRISM Estudo de Caso: Itagua RJ ; I nstituto B rasileiro de G eografia e E statstica : Rio de Janeiro, Brazil, 2009. 55. Futamura, N.; Takaku, J.; Suzuki, H.; Iijima, T.; Tadono, T.; Matsuoka, M.; Shimada, M.; Igarashi, T.; Shibasaki, R. High Resolution DEM Generation from ALOS PRISM Data Algorithm Development and Evaluat ion In Proceedings of the IGARSS 200 2 IEEE International Geoscience and Remote Sensing Symposium, Toronto Canada, 24 28 June 2002 ; Voulme 1, pp. 405 407 56. Geomatics P.C.I PCI Geomatica OrthoEngine User Guide, version 10.3 ; GEOMATIC: Richmond Hill, ON, Canada, 2010. 57. Carvalho J nior, O.A.; Guimar es, R.F.; Gillespie, A.R.; Silva, N.C.; Gomes, R.A.T. A new approach to change vector analysis using distance and similarity measures Remote Sens. 2011 3 2473 2493.
Remote Sens. 2014 6 351 58. Congalton, R.; Green, K. Assessing the Accu racy of Remotely Sensed Data: Principles and Practices ; CRC and Lewis Press: Boca Raton, FL, USA, 1999. 59. Fung, T.; Le Drew, E. Application of principal components analysis to change detection. Photogramm. Eng. Remote Sens. 1987 53 1649 1658. 60. Hirt, C.; Fil mer, M.S.; Featherstone, W.E. Comparison and validation of recent freely available ASTER GDEM ver1, SRTM ver4.1 and GEODATA DEM 9S ver3 digital elevation models over Australia. Aust. J. Earth Sci. 2010 57 337 347. 61. Reuter H.I.; Nelson, A.; Strobl, P.; Me hl, W.; Jarvis, A. A First Assessment of ASTER GDEM Tiles for Absolute Accuracy Relative Accuracy and Terrain Parameters In Proceedings of the IGARSS 2009 Geoscience and Remote Sensing Symposium, Cape Town, South Africa, 12 17 July 2009 ; Volume 5, pp. 240 243. 62. Day, M. Doline morphology and development in Barbados. Ann. Assoc. Am. Geogr. 1983 73 206 219. 63. Fujisada, H.; Urai, M.; Iwasaki, A. Technical methodology for A STER global DEM. IEEE Trans. Geosci. Remote Sens. 2012 50 3725 3736. 64. Doctor D.H.; Young, J.A. An Evaluation of Automated GIS Tools for Delineating Karst Sinkholes and Closed Depressions from 1 Meter LIDAR Derived Digital Elevation Data In Proceedings of the 13th Multidisciplinary Conference on Sinkholes and the Engineering and Environmental Impacts of Karst, Carlsbad, NM, USA, 1 May 15 August 2013 ; pp. 449 458. 65. Filin, S.; Baruch, A. Detection of sinkhole hazards using airborne laser scanning data Photog ramm. Eng. Remote Sens. 2010 76 577 587. 66. Filin, S.; Baruch, A.; Avni, Y.; Marco, S. Sinkhole characterization in the Dead Sea area using airborne laser scanning. Nat. Hazard 2011 58 1135 1154. 67. Shaw Faulkner, M.G.; Stafford, K.W.; Bryant, A.W. Delineation and Classification of Karst Depressions using LiDAR: Fort Hood Military Installation Texas. In Proceedings of the 13th Multidisciplinary Conference on Sinkholes and the Engineering and Environm ental Impacts of Karst, Carlsbad, NM, USA, 1 May 15 August 2013; 459 468. 68. Waltham, T.; Bell, F; Culshaw, M. Ground investigation in sinkhole terrains In Sinkholes and Subsidence ; Springer: Chichester, UK, 2005; p. 181 204 2013 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license ( http://creativecommons.org/licenses/by/3.0/ ).