Automated Generation of a Digital Elevation Model Over Steep Terrain in Antarctica From High-Resolution Satellite Imagery

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

Automated Generation of a Digital Elevation ModelOver Steep Terrain in Antarctica From

High-Resolution Satellite ImageryChangno Lee, Jaehong Oh, Changhee Hong, and Junhee Youn

Abstract—The automated generation of a digital elevationmodel over the Antarctic using stereo matching high-resolutionsatellite images is a challenging task. Moreover, the homogeneousradiometry in the icy environment and the strong geometric dis-similarity between stereo pairs over the steep terrain limit the useof area-based matching techniques. To overcome this issue, wepropose template matching with image transformation in orderto reduce the geometric dissimilarity. First, we generated epipolarresampled images to ensure the ease of estimation and handlingof image dissimilarities from various viewing directions. We thenutilized the normalized cross-correlation (NCC) and transformedthe image patches within the matching window along the sam-ple, line, and diagonal directions in order to improve the matchrates within the steep areas. Furthermore, we tested the proposedmethod using Antarctic IKONOS stereo images and found thatthe overall matching success rate improved from 93.5% to 97.0%for all image pixels. We then computed the success rates over anarea in which the NCC produced a low elevation point densityand observed a more significant improvement from 58.7% to79.26%. When compared to the manually generated elevation, themaximum vertical difference improved from 11.4 to 4.7 m. Withthese improvements, we can build a 1-m resolution elevation modelover the glaciated high relief terrain.

Index Terms—Antarctic, digital elevation model (DEM), epipo-lar, image patch, stereo matching.

I. INTRODUCTION

SCIENTIFIC studies over large glaciated regions, such asGreenland and Antarctica, often require accurate and up-

to-date digital elevation models (DEMs). DEMs are used toestimate ice velocity and thickness changes [1], as well asfor drainage basin area extraction [2] and to monitor sea-levelrise [3]. The elevation accuracy and the high spatial resolutionof DEM are important for research activities in the area sur-rounding the Jang Bogo station, which was recently built as

Manuscript received November 17, 2013; revised May 1, 2014 and June 19,2014; accepted June 29, 2014. This work was supported by a grant from astrategic research project (Development of Extreme Cold Region Site Investi-gation and Monitoring System) funded by the Korea Institute of ConstructionTechnology.

C. Lee is with the Seoul National University of Science and Technology,Seoul 139-743, Korea (e-mail: [email protected]).

J. Oh is with the Chonnam National University, Gwangju 500-757, Korea(e-mail: [email protected]).

C. Hong and J. Youn are with the Korea Institute of Construction Technology,Goyang-si 411-712, Korea (e-mail: [email protected]; [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2014.2335773

South Korea’s second Antarctic station because the elevationerrors in the topographic information directly propagate into theestimations.

Former studies based on radar and laser altimetry, as well asoptical satellites as presented in [4] and [5], generated DEMfrom hundreds of meters to hundreds of kilometers of pointspacing over the entire Antarctic. A 5-km resolution DEM wasderived for a steep terrain using European Remote SensingSatellite-1 (ERS-1) radar altimeter data and terrestrial data[4], [5]. The errors in the data conveyed a linear trend withsteepness; furthermore, the topographic information tended tohave poor vertical accuracy near the regions of the steep terrain(> 100 m). Geoscience Laser Altimeter System/ICE, Cloud,and land Elevation (GLAS/ICESat) laser altimeter profile datawere used for a 500-m resolution DEM [6]. This DEM ofAntarctica provided new surface elevation grids that covered alarger area and exhibited fewer slope-related effects comparedto radar altimetry. The accuracy of the topographic informationobtained from the GLAS/ICESat was within a few decimeterson the flat terrain [7], [8]; however, the major limitation of thedata was the coarse across-track spacing [9], such as 2.5 kmat latitude 80◦ S [10]. Later, a DEM with 1-km spacing wasgenerated by combining the radar and laser altimetry data inorder to increase the accuracy and extend the coverage between81.5 and 86◦ S [9]. The accuracy of this DEM was roughlytwice that of the aforementioned 5-km DEM when comparedto the airborne altimeter data. By 2014, the TerraSAR-X add-on for Digital Elevation Measurement (TanDEM-X) mission isscheduled to generate a complete global DEM with approxi-mately 12-m resolution and 10-m vertical accuracy, althoughit will not be available for Antarctica in 2014. However, radaraltimetry-based approaches are less efficient for steeper areassuch as ice caps and coastal regions [11], [12]. Hence, a globalDEM, known as Global Digital Elevation Model version 2(GDEM-2) from the Advanced Spaceborne Thermal Emissionand Reflection Radiometer (ASTER), was released in October2011. Elevation data obtained from high latitude areas, suchas Greenland and Antarctica, are poor due to difficult imagematching over featureless snow fields [13]. A 100-m DEMof the Antarctic Peninsula was generated by minimizing theartifacts of the ASTER GDEM [14], and it was reported thatthe DEM shows lower vertical accuracy on mountain peaks andsteep slopes compared to other terrains.

Other studies have focused on generating detailed localDEM. An airborne LiDAR system was used to generate

0196-2892 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 1. Steep terrain creates geometric dissimilarities in the stereo images.

high-resolution local Antarctic DEM with 2- and 4-m spac-ings for 18 sites around McMurdo Sound [15]. The NationalAeronautics and Space Administration also provides IceBridgeLand, Vegetation, and Ice Sensor (LVIS L2) geolocated surfaceelevation products acquired using the airborne LiDAR system.Although the operating airborne system in Antarctica is attrac-tive for its high vertical accuracy (e.g., less than 50 cm) andspatial resolution (e.g., a few meters), it is costly and needsmuch logistical infrastructure; it is therefore inadequate, partic-ularly over large regions. For this reason, spaceborne systemsgain more popularity. In-track stereoscopic data from SatellitePour l’Observation de la Terre-5 High Resolution Stereoscopic(SPOT-5 HRS) were used to generate a 40-m DEM for theAntarctic margins; moreover, the data conveyed root meansquare (rms) elevation errors of less than 10 m [11]. Similarly,SPOT-5 HRS data were used for the Canadian Arctic [8], whichis an 80% ice-covered region. Both studies reported large errorsof over 100 m on slopes that are steeper than 40◦. Few studieshave investigated the use of 1 m or submeter spatial resolutionsatellite images for generating DEM over glaciated regions,although one study tested DEM over polar outlet glaciers witha 0.5–10-m overall vertical error using Worldview 1 and 2 andthe ERDAS LPS eATE algorithm [16].

We propose a new stereo matching method of high-resolutionsatellite images for the generation of accurate DEM over theAntarctic, including steep slope terrains. Although the defini-tion of “high resolution” in the satellite imaging system is notfixed, in topographic mapping, the term implies that it startsapproximately at a ground sampling distance of 10 m [17].The major obstacle when generating accurate DEM with meter-level spacing from the stereoscopic satellite images over steepglaciated regions is that stereo image matching does not workwell in these areas. The homogeneous radiometry within theice and the strong geometric image distortions over the steepterrain limit the use of image matching techniques such asnormalized cross-correlation (NCC) [8], [13]. Furthermore, thehomogeneity within the ice can produce false positives; hence, amore serious problem is that the forward and backward viewingangles result in geometrically dissimilar topographic images,particularly over steeper areas such as ice caps and coastalregions (see Fig. 1). In the image in Fig. 1(a), the terrain facingthe bottom appears to be wide when viewed from the bottom. Incontrast, Fig. 1(b) shows the same terrain, but the image is muchnarrower when viewed from the top down. Steep slopes create

geometrically dissimilar images, and the dissimilarity differsdepending on the viewing angle and direction. This largedissimilarity significantly lowers the matching success ratebecause template-based image matching such as NCC worksbest on relatively flat or smooth surfaces.

To overcome this difficulty, we propose a new templatematching approach that transforms the image patches withinthe matching window. To this end, we first carried out epipolarimage resampling in order to accurately align the viewingangles of the stereo pair parallel to the image rows. Notethat epipolar resampling removes all y-parallax (image rowdifference between the stereo pair) and leaves only x-parallax,which is directly related to height or elevation [19]. We thenused the NCC by transforming the image patches within thematching window along the sample (scaling), line (skewing),and diagonal (scaling and skewing) directions in order to im-prove the match rates over the large geometric dissimilaritywithin the steep areas.

II. PROPOSED METHOD

A flowchart of the proposed method is presented in Fig. 2.Given the high-resolution stereoscopic satellite images with ra-tional polynomial coefficients (RPCs), the positional accuracyof the original RPCs was enhanced using the bias-correctionprocedure with ground control points [18]. The optimal pa-rameterization for the correction may vary depending on thesatellites; however, one pixel level of positional accuracy canbe achieved by introducing only the shift terms for IKONOS.Following the correction of the RPCs, epipolar image resam-pling was carried out. In the epipolar image space, the conjugatepoints in the stereo images are aligned in the same image linesuch that a faster point search is enabled. In addition, the terraindistortion can be easily handled due to the steep slopes becausethe major terrain reliefs are aligned parallel to the image row.To create the epipolar resampled images, a piecewise approachwas followed because either the RPCs or the rigorous sensormodel can be used.

We applied template-based stereo matching, NCC, to theGaussian resampled images using the proposed image patchtransformation scheme. First, NCC matching was carried outfor the image pair at the lowest image pyramid. At this stage,image transformation was not enabled because the resolutionwas too low to consider the local terrain. We then continuedNCC matching in the next image pyramid. If the best cor-relation was higher than the predefined threshold (e.g., 0.8),then the point pair was accepted for the terrain reconstruction.Otherwise, the terrain was analyzed using the lower scalestereo matching results. The image patch of the left image wasthen incrementally transformed while considering the terrainaspect. If the slope was along the sample direction in theepipolar image, the image patch was rescaled along the sampledirection. Next, we skewed the image patch if the slope wasalong the line direction; if the slopes were in the diagonaldirections, the image was resized along the sample direction,and the image skewing was then applied while considering thedirection. As a result, the left image patch was transformedincrementally to a predefined level, and these transformed

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 3

Fig. 2. Flowchart of the study.

image sets were used for the conjugate point search in theright epipolar image. If the best correlation was higher thanthe established threshold, the point pair was accepted forground reconstruction. We continued stereo matching up tothe largest pyramid image to the generated DEM of full res-olution of the input images, e.g., 1-m resolution for IKONOSimages.

A. Epipolar Image Resampling

Epipolar image resampling has three major advantages inthe proposed stereo image matching approach. First, the re-sampling process helps to improve the matching speed byreducing the search area and minimizing the y-parallax. Ac-curate epipolar image resampling was often carried out for amaximum line difference between the stereo pair of one pixellevel. Therefore, the search range can be set to only one pixel

larger than the matching window height. Second, the minimizedsearch area helps to limit the potential inclusion of the matchingoutliers. Third and most importantly, the resampling enablesthe easy estimation and handling of image dissimilarity fromthe diverse viewing directions because a major dissimilarityappeared along the left and right directions in the epipolarresampled image. The epipolar resampled images from Fig. 1are shown in Fig. 3. Note that the viewing angles of the stereopair are aligned along the horizontal direction, which is theflight pass of the satellite.

The epipolar geometry of a satellite pushbroom sensor differssomewhat from that of frame cameras. Previous studies (suchas [20] and [21]) have shown that the epipolar curve shape ofthe pushbroom sensor is hyperbola-like and that the epipolarcurve pairs do not exist for the entire scene. These propertieslimit the accurate epipolar image resampling of the satelliteimages. Thus, in a couple of studies, the method used has

4 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 3. Epipolar resampled images of Fig. 1.

relied on approximate models such as 2-D affine [22], [23];however, this method has shown rather poor results for highapparent-field-of-view sensors with a dynamic trajectory. Thus,a piecewise method was developed to accurately determineepipolar geometry and to resample the epipolar images [19].The piecewise method is based on iterative projections from animage to the ground and from the ground to the paired image.The projection can be carried out using rigorous physical sensormodels or RPCs without further approximations.

B. Template Matching

In-track high-resolution stereoscopic satellite images showpractically no spectral differences between the image pairs; fur-thermore, feature matching methods that are more sophisticatedand computationally expensive are usually not required. There-fore, NCC can be used for stereo image matching using (1).Epipolar image resampling and the pyramid image approachsignificantly reduce the search space along the line and sampledirections, respectively. In addition, the convolution theoremcan be utilized for a better matching speed [24]

ncc=

∑wi=1

∑wj=1

[(Lij−L)(Rij −R)

]√[∑w

i=1

∑wj=1(Lij−L)2

] [∑wi=1

∑wj=1(Rij−R)2

](1)

where L is the image patch in the left epipolar image, R is theimage patch of the search region in the right epipolar image,Lij and Rij are the elements of image patches L and R,respectively, at line i and sample j, L and R are the averagesof all intensity values in images L and R, respectively, and w isthe image patch size.

C. Matching by Transforming the Image Patch

Area-based matching methods, including NCC, inherentlyassume a smooth surface within a small matching window, suchas 7 × 7 pixels or a larger size. This assumption may not be truefor image patches over a steep terrain because different satelliteviewing angles may introduce different terrain relief patterns inthe matching windows.

These problematic relief distortions are explained in theepipolar resampled image, assuming that the terrain has a steepbut constant slope over the window (see Fig. 4). Fig. 4(a)

presents the case of a steep slope facing the left or right. Notethat the dotted arrow indicates the orthogonal direction to theslope, and the elevation along the line is assumed to be constant.The left-facing slope is wide in the left epipolar image, while itappears to be narrow in the right epipolar image. The conversewould be true for a right-facing slope. Due to such dissimilarity,template-based matching would not be successful. Fig. 4(b)shows the case of a steep slope facing up or down. In theright epipolar image, the higher terrain is displaced to the leftcompared to the lower terrain due to the viewing direction.Steep slopes facing the diagonal directions are depicted inFig. 4(c) and (d). The steep slope facing the upper right is widerin the right image, and the higher terrain in the lower left ismore displaced to the left [see Fig. 4(c)]. Conversely, the slopefacing the lower left appears to be narrower, and the upper rightterrain is more displaced to the left. This contrasts the imagein Fig. 4(d), where the slopes facing the upper left and lowerright appear to be narrow and wide, respectively. Furthermore,the high terrain also appears to be more displaced to the left.

To overcome these issues, we transformed the image patcheswithin the matching window. The benefit of this process was tominimize the geometric dissimilarity between the stereoscopicdata. For example, in Fig. 4(a), the simple scaling of the leftimage patch along the sample (horizontal) direction wouldreduce the difference. In the cases presented in Fig. 4(b)–(d),skewing the image patch would also be required. Thus, optimalimage transformation strongly depends on the local slope di-rection. Therefore, we incrementally transformed the left imagepatch along the sample (scaling), line (skewing), and diagonal(scaling and skewing) directions for the best image matching.

The aforementioned image transformation method requiresthe estimation of the slope pattern in order to select an optimalimage transformation. In the pyramid image approach, wecomputed the slope information from the lower scale imagematching. However, we did not need to compute the real groundelevation in this stage; the slope was analyzed in the epipolarimage space itself. Because the x-parallax from stereo matchingwas proportional to the ground height in the epipolar image, weused the x-parallax itself for the slope estimation.

D. Filtering and Refinement

The proposed method did not guarantee any outlier-freematching results; therefore, we refined those results. The firsttarget included the points that had a matching correlationhigher than the threshold but were not located correctly in theright image, i.e., false positives. We filtered out all the pointsthat showed abnormal x-parallaxes if the parallax was largerthan the median x-parallaxes of the neighboring points, e.g.,smoothing using a 3 × 3 median filter. Second, we continuedthe refinement of the conjugate point pairs with low cross-correlations, implying a high possibility of a mismatch. In thisstage, we can use a better slope estimation and a more mini-mized search space from the full-scale image matching results.Note that the lower scale image matching results were usedfor this information in the aforementioned pyramid approach.The refinement was carried out iteratively until no furtherimprovement was found.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 5

Fig. 4. Relief distortion of steep slopes in the epipolar resampled images. (a) Slope facing left or right. (b) Slope facing up or down. (c) Slope facing the upperright or lower left. (d) Slope facing the upper left or lower right.

Fig. 5. Test data location in Antarctica.

III. EXPERIMENT

A. Data

The test area was located in Terra Nova Bay, where JangBogo Station (the second Korean research base) was com-pleted in February 2014 (see Fig. 5). Terra Nova Bay is about64 km long and is located between Victoria Land and RossSea, Antarctica. The Italian and German research stations arealso located on Terra Nova Bay. The temperature rises abovefreezing during the summer (from November to January), andlarge rock and soil areas are clearly visible when the snowmelts. In-track IKONOS stereoscopic data were acquired onJanuary 16, 2011; the specifications are presented in Table I andFig. 6, portraying the IKONOS image. In the previous research,4-m multispectral information of this data was used to monitorvarious land cover types and to define the accurate detailed landcover boundaries [25]. Here, most surfaces are covered withice, and some bare surfaces can be observed at the lower rightcorner where Jang Bogo Station is built. The steep slopes areobserved at the center and the upper left of the image. Also,cliffs are located near Ross Sea at the bottom of the image. Theproduct level of the images was Geo, which is map projected.Note that, in Fig. 5, the upper direction of the IKONOS image is

TABLE ISPECIFICATIONS OF IKONOS TEST DATA

the lower right direction because the upper direction of Fig. 5 isoriented to 0◦ longitude while the IKONOS image is located inthe Universal Transverse Mercator (UTM) 58-Zone (longitude162–168◦). The image coverage is about 11 km × 11 km, andthe target terrain elevation range is approximately 0–1000 m.The convergence angle of the stereo pair was computed to beabout 30◦ from the azimuth and elevation angles. Based onthe preinstalled Global Navigation Satellite System (GNSS)base station (8 km away), four ground control points wereGNSS surveyed in real-time kinematic (RTK) for 10 min usingTrimble R8 at the German station, which is shown in the lowerright corner of Fig. 6(a), with the point locations highlighted inthe boxes in Fig. 6(b). Note that the photogrammetric groundcontrol points require at least less than a half pixel of thespatial resolution, e.g., less than 50 cm for the tested IKONOSimage. This positional accuracy is obtained with GNSS RTKsurveying, which shows the centimeter level of 3-D accuracy.The corners of the building were used for the controls inorder to easily identify the images [see Fig. 6(b)–(c)]. Thesecontrol points enabled a bias correction of IKONOS RPCs by

6 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 6. Test data. (a) IKONOS image in UTM58. (b) Ground control points.(c) GNSS measurements at German base.

Fig. 7. Epipolar image resampling: (a) Resampling each image and(b) anaglyph image that can be stereo-viewed using cyan–red color glasses.The cyan-colored area around the center and the upper right corner of theimage indicates the high elevation region that produced large x-parallaxes inthe stereoscopic data.

introducing only the shift term. After the bias correction, thepositional residuals of the control points dropped from aboutseven pixels to less than one-half of a pixel. In addition, wemanually observed the conjugate point pairs over the entireimage [small circles in Fig. 6(a)] in order to verify the stereomatching quality. A total of 58 points were taken in the digitalphotogrammetric workstation.

B. Epipolar Generation

Using the bias-compensated RPCs, epipolar image resam-pling was carried out using the piecewise approach. Consid-ering the elevation range of the terrain, the minimum andmaximum elevation for the piecewise approach was set at 0 and1000 m, respectively. Next, a second-order polynomial equationwas used for the image transformation. The transformationresiduals were 0.02–0.1 pixels, indicating that practically notransformation errors were introduced. Epipolar image resam-pling conveyed the maximum y-parallax of 1.1 pixels. Fig. 7(a)

Fig. 8. Successful matching points without patch transformation (correlationwas larger than 0.8 in the white region).

shows the epipolar image resampling for the left and rightimages, and Fig. 7(b) shows the anaglyph image. Note that weplaced the cyan and red colors to the left and right of the epipo-lar images, respectively, in order to generate an anaglyph image.Therefore, the cyan colored areas around the center and theupper right corner of the image indicate a high elevation region,which produced large x-parallaxes in the stereoscopic data.

C. Stereo Matching Without Image Patch Transformation

We generated an image pyramid for the hierarchical stereomatching. The image pyramid of the five stages was createdusing Gaussian image resampling such that the lowest imagehad a spatial resolution 16 times lower than the original image.For stereo matching, we used a matching window of 21 ×21 pixels in order to increase the possibility of including ter-rain primitives over the radiometrically homogeneous surface.Window sizes that are too large may show negative effectsbecause the window tends to include various x-parallaxes overthe rugged terrain, thus leading to the failure of stereo matching.

For a selected image patch in the left epipolar image, the con-jugate patch was searched in the right image. The smaller searchregion was preferred for faster computation and for mitigatingthe matching outliers. The line and sample sizes of the searchrange were established using the y-parallax error information ofthe data and the terrain elevation range, respectively. Therefore,the line size of the search region was set to one pixel largerthan that of the image patch. The sample size of the searchregion was then derived from the lower scale image match-ing result. Hence, the results from the lower scale matchingx-parallax estimates over the stereoscopic data, and the initialx-parallax of an image patch in the current matching stage couldbe derived using a linear interpolation. For the lowest scaleimage, the search range along the sample direction was set to−44 to +11 pixels in order to correspond to the elevation rangeof 0–1000 m, respectively. In Fig. 8, the white region indicatesa correlation larger than 0.8, while the black region showsa low correlation in the left epipolar image. Overall, stereomatching did not produce satisfactory results over the ruggedterrain region in the center and in the upper part of the image.Note that the lower left of this image shows the Ross Sea.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 7

Fig. 9. Improvement of matching correlation by transforming the imagepatch. The original image patches are shown in the left column, the transformedpatches are in the middle column, and the conjugate images are shown in theright column. (a) Scaling for the sample direction slope (correlation increasedfrom 0.31 to 0.94). (b) Skewing for the line direction slope (correlationincreased from 0.36 to 0.89). (c) Scaling and skewing for the slope facing theupper right (correlation increased from 0.55 to 0.83). (d) Scaling and skewingfor the slope facing the upper left (correlation increased from 0.56 to 0.95).

D. Stereo Matching With Image Patch Transformation

Stereo matching without a patch transformation did notproduce dense matching point pairs over the steep terrain.Therefore, we applied the image transformation method for theregion where stereo matching failed. In the lowest scale image,only stereo matching was carried out without an image patchtransformation because the relief was not significant at that spa-tial resolution. In the next pyramid stage, we first analyzed theslope aspect from the lower scale stereo matching results andclassified them into four major cases: sample, line, diagonal-1(upper right to lower left), and diagonal-2 (upper left to lowerright). We then applied the image transformation accordingly.

The transformation level was incrementally increased, andeach transformed image was used for NCC matching in orderto compute the correlations. Next, the location showing the bestcorrelation was accepted for the successful matching point inthe right image. Fig. 9(a)–(d) shows the sample image patchesusing this approach. In the three columns, the left column isthe original image patch from the left epipolar image, andthe middle image shows the transformed left image. The right

Fig. 10. Region where stereo matching without patch transformation showedpoor results [white regions in (a) and box regions in (b)]. (a) White regions arewhere the correlation was less than 0.8, but the proposed method gave a highercorrelation. (b) Box regions showing less than 70% success rate with matchingwindows of 21 × 21 pixels.

image of the search region is the right epipolar image. Fig. 9(a)shows the scaling of the image patch along the sample direction.Although stereo matching using the original image patch wasunsuccessful, with a correlation of 0.31, the scaled image inthe middle produced a correlation of 0.94 with the right imagefor a successful conjugate point location. Fig. 9(b) presentsdown-facing slope images. The matching without patch trans-formation indicated a correlation of 0.36; however, the skewingimproved the correlation to 0.89. Note that the upper sideimage was more displaced to the left. Fig. 9(c) and (d) showsthe diagonal direction slopes. Here, we can observe that theimage patches were skewed and then scaled along the sampledirection; moreover, the correlation for the probable conjugatepoint locations was increased from 0.55 to 0.83 and from 0.56to 0.95, respectively.

We then filtered and refined the matching results. The whiteareas in Fig. 10(a) depict the region where stereo matchingwithout patch transformation showed a correlation of less than0.8; however, the proposed method gave a higher correlation.Moreover, the points were dense over the steep terrains, wherethey are located in the center and the upper right corner of theimage. This result corresponds to the black region in Fig. 8.Furthermore, Fig. 11 shows the derived DEM with 1-m spacingfor the test area in Terra Nova Bay.

E. Comparison and Evaluation

Stereo matching without patch transformation produced100 833 114 matching points out of the 107 896 865 points at-tempted, while excluding those produced in the sea region, andthe overall success rate was 93.5%. The proposed transforma-tion method without refinement produced 102 402 187 points,and the success rate was 94.9%. Hence, we obtained a 1.4%increase in the number of matching points. We could also haveobtained 104 613 870 matching points when the refinement wasenabled, which corresponds to a success rate of 97.0% and a netgain of 3.5%.

We continued the comparison by narrowing the regionwhere stereo matching without patch transformation showedunsatisfactory results. Fig. 10(b) shows the sample regionswhere stereo matching without patch transformation had asuccess rate of less than 70%. The overall success rate of thematching without patch transformation was 58.7% over the

8 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 11. One-meter resolution DEM in Terra Nova Bay (UTM58, unit: meter).

Fig. 12. DEM samples from (left) stereo matching without patch transforma-tion and (right) this study.

region; however, when the image patch transformation wasutilized without and with refinement, the success rate increasedto 67.7% and 79.26%, respectively. This indicated that theproposed approach facilitated a net improvement of over 20%in areas that are difficult for stereo matching without patchtransformation.

Fig. 12 presents the DEM samples in order to show theimprovements. The left figures are those from matching withoutpatch transformation, which show topographic noises creatingfalse linear and rectangular features over the steep terrain be-cause the matching point density was sparse. The DEM sampleson the right were obtained from the proposed approach, whichprovided more detail over the steep slopes.

TABLE IIELEVATION DIFFERENCE FROM THE MANUALLY ACQUIRED ELEVATION

To overcome the lack of GNSS measurements over theentire image region for validating the elevation reliability of thegenerated DEM, we utilized the manually observed checkpointspreviously shown in Fig. 6(a). The checkpoints were acquiredby a skilled operator considering the terrain types and thendistributed over the entire image region. The aim of this studyis to develop an automated stereo matching algorithm that hasa performance close to image matching by human operators.Therefore, we compared the elevation of the automaticallygenerated DEM to the elevation of the manually acquiredcheckpoints; the resulting elevation differences are presented inTable II. Stereo matching without patch transformation showedan elevation difference of 2.2 m in rms and a maximum of11.4 m from the manually acquired elevation. However, theproposed method with refinement reduced the difference to1.3 m in rms and the maximum difference to 4.7 m. Fromthis result, it can be concluded that the proposed approach canautomatically generate DEM a few meters close to that obtainedby human operators; this is known as a 1-m level of verticalaccuracy, as presented in [16].

IV. CONCLUSION

To generate a 1-m resolution DEM over the steep terrainin the Antarctic from the in-track high-resolution satellite im-ages, we proposed a hierarchical template matching algorithmaugmented by transforming the template images. Using theepipolar image, we classified slopes into four cases: sample,line, diagonal-1 (upper right to lower left), and diagonal-2(upper left to lower right). Given the assumption that the slopewas steep but constant, the template image was transformedfor stereo matching. The slope information could be obtainedusing the slope analysis from the lower scale hierarchical stereomatching. We tested the proposed method using AntarcticIKONOS stereo images. The result showed that the overallmatching success rate improved from 93.5% to 97.0% for allimage pixels. For the area where stereo matching without patchtransformation produced sparse matching point density, thesuccess rates showed an improvement from 58.7% to 79.26%.This improvement will significantly reduce the workload ofmanual editing over the area. Finally, we generated a 1-mresolution DEM over the glaciated high relief terrain, whichshowed a maximum elevation error of 4.7 m. This was animprovement over the DEM, which had a maximum elevationerror of 11.4 m. Overall, major improvements were made overthe steep slope, as shown in Fig. 12.

A future study will include an absolute accuracy assessmentusing dense GPS control points. In addition, template-basededge matching with transformation for highly shaded regionswill be performed. Heterogeneous satellite image pairs will alsobe studied for greater economical DEM generation.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 9

REFERENCES

[1] R. R. Muskett et al., “Airborne and spaceborne DEM- and laser altimetry-derived surface elevation and volume changes of the Bering Glaciersystem, Alaska, USA, Yukon, Canada, 1972–2006,” J. Glaciol., vol. 55,no. 190, pp. 316–326, 2009.

[2] D. G. Vaughan, J. L. Bamber, M. Giovinetto, J. Russel, andA. P. R. Cooper, “Reassessment of net surface mass balance in Antarc-tica,” J. Climate, vol. 12, no. 4, pp. 933–946, Apr. 1999.

[3] E. Berthier, E. Schiefer, G. K. C. Clarke, B. Menounos, and F. Rémy,“Contribution of Alaskan glaciers to sea-level rise derived from satelliteimagery,” Nat. Geosci., vol. 3, no. 2, pp. 92–95, Feb. 2010.

[4] J. L. Bamber, “A digital elevation model of the Antarctic ice sheet de-rived from ERS-1 altimeter data and comparison with terrestrial measure-ments,” Ann. Glaciol., vol. 20, no. 1, pp. 48–54, Jun. 1995.

[5] J. L. Bamber and J. L. Gomez-Dans, “The accuracy of digital elevationmodels of the Antarctic continent,” Earth Planet. Sci. Lett., vol. 237,no. 3/4, pp. 516–523, Sep. 2005.

[6] J. P. DiMarzio, GLAS/ICESat 500 m Laser Altimetry Digital ElevationModel of Antarctica. Boulder, CO, USA: National Snow and Ice DataCenter, 2007.

[7] C. A. Shuman et al., “ICESat Antarctic elevation data: Preliminary pre-cision and accuracy assessment,” Geophys. Res. Lett., vol. 33, no. 7,pp. L07501-1–L07501-4, Apr. 2006.

[8] T. Toutin, “Digital elevation model generation over glacierized regions,”in Encyclopedia of Snow, Ice and Glaciers. New York, NY, USA:Springer-Verlag, 2011, pp. 202–213.

[9] J. L. Bamber, J. L. Gomez-Danz, and J. A. Griggs, “A new 1 km digitalelevation model of the Antarctic derived from combined satellite radar andlaser data—Part 1: Data and methods,” Cryosphere, vol. 3, no. 1, pp. 101–111, May 2009.

[10] H. Zwally et al., “ICESat’s laser measurements of polar ice, atmosphere,ocean, land,” J. Geodyn., vol. 34, no. 3/4, pp. 405–445, Oct./Nov. 2002.

[11] J. Korona, E. Berthier, M. Bernard, F. Remy, and E. Thouvenot, “Spirit.Spot 5 stereoscopic survey of polar ice: Reference images and topogra-phies during the fourth international polar year (2007–2009),” ISPRS J.Photogramm. Remote Sens., vol. 64, no. 2, pp. 204–212, Mar. 2009.

[12] T. Toutin, C. Schmitt, E. Berthier, and D. Clavet, “DEM generation overice fields in the Canadian Arctic with along-track SPOT5 HRS stereodata,” Can. J. Remote Sens., vol. 37, no. 4, pp. 429–438, Aug. 2011.

[13] M. Urai, T. Tachikawa, and H. Fujisada, “Data acquisition strategies forASTER global DEM generation,” ISPRS Ann. Photogramm. Remote Sens.Spatial Inf. Sci., vol. 1–4, pp. 199–202, Jul. 2012, 2012 XXII ISPRSCongress, Melbourne, Australia, 25 Aug.–01 Sep..

[14] A. J. Cook, T. Murray, A. Luckman, D. G. Vaughan, and N. E. Barrand,“A new 100-m digital elevation model of the Antarctic Peninsula derivedfrom ASTER global DEM: Methods and accuracy assessment,” EarthSyst. Sci. Data, vol. 4, no. 1, pp. 129–142, Jan. 2012.

[15] T. Shenk et al., “DEM generation from the Antarctic LiDAR data,” OhioState Univ., Columbus, OH, USA, 2004.

[16] C. C. Porter, P. J. Morin, I. M. Howat, S. Niebuhr, and B. E. Smith, “DEMextraction from high-resolution stereoscopic Worldview 1 & 2 imagery ofpolar outlet glaciers,” presented at the American Geophysical Union, FallMeeting, San Francisco, CA, USA, Dec. 5–9, 2011, Paper C41E-0461.

[17] K. Jacobsen, “High resolution imaging satellite systems,” presented at theEarsel Workshop Remote Sensing Use 3rd Dimension Remote SensingPurposes, Porto, Portugal, 2005.

[18] C. S. Fraser and H. B. Hanley, “Bias-compensated RPCs for sensor ori-entation of high-resolution satellite imagery,” Photogramm. Eng. RemoteSens., vol. 71, no. 8, pp. 909–915, Aug. 2005.

[19] J. Oh, W. H. Lee, C. K. Toth, D. A. Grejner-Brzezinska, and C. N. Lee,“A piecewise approach to epipolar resampling of pushbroom satelliteimages based on RPC,” Photogramm. Eng. Remote Sens., vol. 76, no. 12,pp. 1353–1363, Nov. 2010.

[20] R. Gupta and R. Hartley, “Linear pushbroom cameras,” IEEE Trans.Pattern Anal. Mach. Intell., vol. 19, no. 9, pp. 963–975, Sep. 1997.

[21] T. Kim, “A study on the epipolarity of linear pushbroom images,”Photogramm. Eng. Remote Sens., vol. 66, no. 8, pp. 961–966, Aug. 2000.

[22] T. Ono, “Epipolar resampling of high resolution satellite imagery,” inProc. Joint Workshop ISPRS WG I/1, I/3/IV/4 Sensors Mapping Space,Hannover, Germany, 1999, [CD-ROM].

[23] M. Morgan, “Epipolar resampling of linear array scanner scene,” Ph.D.dissertation, Univ. Calgary, Calgary, AB, Canada, 2004.

[24] J. Oh, C. K. Toth, and D. A. Grejner-Brzezinska, “Automatic geo-referencing of aerial images using high-resolution stereo satellite im-ages,” Photogramm. Eng. Remote Sens., vol. 77, no. 11, pp. 1157–1168,Nov. 2011.

[25] S. H. Kim and C. H. Hong, “Antarctic land-cover classifier using IKONOSand Hyperion data at Terra Nova Bay,” Int. J. Remote Sens., vol. 33,no. 22, pp. 7151–7164, Nov. 2012.

Changno Lee received the B.S. and M.S. degreesin civil engineering from the Seoul National Univer-sity, Seoul, Korea, in 1987 and 1989, respectively,and the Ph.D. degree from Purdue University, WestLafayette, IN, USA, in 1999.

He was a Senior Researcher at the Electronicsand Telecommunications Research Institute and cur-rently works as an Associate Professor in Civil Engi-neering at the Seoul National University of Scienceand Technology. His research interests are sensormodeling, robust estimation, and automated feature

extraction from imagery.

Jaehong Oh received the B.S. and M.S. degrees incivil engineering from the Seoul National University,Seoul, Korea, in 1999 and 2001, respectively, andthe Ph.D. degree from The Ohio State University,Columbus, OH, USA, in 2011.

He was a Senior Researcher at the Space Imag-ing Asia and Electronics and TelecommunicationsResearch Institute and currently works as an Assis-tant Professor in Civil Engineering at the ChonnamNational University, Gwangju, Korea. His researchinterests cover georeferencing techniques including

photogrammetric sensor modeling and stereo image processing.

Changhee Hong received the B.S. and M.S. degreesin geospatial information technology from Inha Uni-versity, Incheon, Korea, in 1997 and 1999, respec-tively, and the Ph.D. degree from Seoul NationalUniversity, Seoul, Korea, in 2006.

He works as a Senior Researcher at the Informa-tion and Communication Technology Convergenceand Integration Research Division in the Korea Insti-tute of Construction Technology, Goyang-si, Korea.His research interests cover ecology-based remotesensing and extreme cold region monitoring using

remote sensing.

Junhee Youn received the B.S. and M.S. degreesin civil engineering from Yonsei University, Seoul,Korea, in 1996 and 1998, respectively, and the Ph.D.degree from Purdue University, West Lafayette, IN,USA, in 2006.

He was a Senior Consultant at SAMSUNG SDSand currently works as a Senior Researcher atthe Korea Institute of Construction Technology,Goyang-si, Korea. His research interests cover dis-aster management with geospatial data, man-madeobject extraction from multiple sensors, and real-

time photogrammetry with mobile mapping system.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

Automated Generation of a Digital Elevation ModelOver Steep Terrain in Antarctica From

High-Resolution Satellite ImageryChangno Lee, Jaehong Oh, Changhee Hong, and Junhee Youn

Abstract—The automated generation of a digital elevationmodel over the Antarctic using stereo matching high-resolutionsatellite images is a challenging task. Moreover, the homogeneousradiometry in the icy environment and the strong geometric dis-similarity between stereo pairs over the steep terrain limit the useof area-based matching techniques. To overcome this issue, wepropose template matching with image transformation in orderto reduce the geometric dissimilarity. First, we generated epipolarresampled images to ensure the ease of estimation and handlingof image dissimilarities from various viewing directions. We thenutilized the normalized cross-correlation (NCC) and transformedthe image patches within the matching window along the sam-ple, line, and diagonal directions in order to improve the matchrates within the steep areas. Furthermore, we tested the proposedmethod using Antarctic IKONOS stereo images and found thatthe overall matching success rate improved from 93.5% to 97.0%for all image pixels. We then computed the success rates over anarea in which the NCC produced a low elevation point densityand observed a more significant improvement from 58.7% to79.26%. When compared to the manually generated elevation, themaximum vertical difference improved from 11.4 to 4.7 m. Withthese improvements, we can build a 1-m resolution elevation modelover the glaciated high relief terrain.

Index Terms—Antarctic, digital elevation model (DEM), epipo-lar, image patch, stereo matching.

I. INTRODUCTION

SCIENTIFIC studies over large glaciated regions, such asGreenland and Antarctica, often require accurate and up-

to-date digital elevation models (DEMs). DEMs are used toestimate ice velocity and thickness changes [1], as well asfor drainage basin area extraction [2] and to monitor sea-levelrise [3]. The elevation accuracy and the high spatial resolutionof DEM are important for research activities in the area sur-rounding the Jang Bogo station, which was recently built as

Manuscript received November 17, 2013; revised May 1, 2014 and June 19,2014; accepted June 29, 2014. This work was supported by a grant from astrategic research project (Development of Extreme Cold Region Site Investi-gation and Monitoring System) funded by the Korea Institute of ConstructionTechnology.

C. Lee is with the Seoul National University of Science and Technology,Seoul 139-743, Korea (e-mail: [email protected]).

J. Oh is with the Chonnam National University, Gwangju 500-757, Korea(e-mail: [email protected]).

C. Hong and J. Youn are with the Korea Institute of Construction Technology,Goyang-si 411-712, Korea (e-mail: [email protected]; [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2014.2335773

South Korea’s second Antarctic station because the elevationerrors in the topographic information directly propagate into theestimations.

Former studies based on radar and laser altimetry, as well asoptical satellites as presented in [4] and [5], generated DEMfrom hundreds of meters to hundreds of kilometers of pointspacing over the entire Antarctic. A 5-km resolution DEM wasderived for a steep terrain using European Remote SensingSatellite-1 (ERS-1) radar altimeter data and terrestrial data[4], [5]. The errors in the data conveyed a linear trend withsteepness; furthermore, the topographic information tended tohave poor vertical accuracy near the regions of the steep terrain(> 100 m). Geoscience Laser Altimeter System/ICE, Cloud,and land Elevation (GLAS/ICESat) laser altimeter profile datawere used for a 500-m resolution DEM [6]. This DEM ofAntarctica provided new surface elevation grids that covered alarger area and exhibited fewer slope-related effects comparedto radar altimetry. The accuracy of the topographic informationobtained from the GLAS/ICESat was within a few decimeterson the flat terrain [7], [8]; however, the major limitation of thedata was the coarse across-track spacing [9], such as 2.5 kmat latitude 80◦ S [10]. Later, a DEM with 1-km spacing wasgenerated by combining the radar and laser altimetry data inorder to increase the accuracy and extend the coverage between81.5 and 86◦ S [9]. The accuracy of this DEM was roughlytwice that of the aforementioned 5-km DEM when comparedto the airborne altimeter data. By 2014, the TerraSAR-X add-on for Digital Elevation Measurement (TanDEM-X) mission isscheduled to generate a complete global DEM with approxi-mately 12-m resolution and 10-m vertical accuracy, althoughit will not be available for Antarctica in 2014. However, radaraltimetry-based approaches are less efficient for steeper areassuch as ice caps and coastal regions [11], [12]. Hence, a globalDEM, known as Global Digital Elevation Model version 2(GDEM-2) from the Advanced Spaceborne Thermal Emissionand Reflection Radiometer (ASTER), was released in October2011. Elevation data obtained from high latitude areas, suchas Greenland and Antarctica, are poor due to difficult imagematching over featureless snow fields [13]. A 100-m DEMof the Antarctic Peninsula was generated by minimizing theartifacts of the ASTER GDEM [14], and it was reported thatthe DEM shows lower vertical accuracy on mountain peaks andsteep slopes compared to other terrains.

Other studies have focused on generating detailed localDEM. An airborne LiDAR system was used to generate

0196-2892 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 1. Steep terrain creates geometric dissimilarities in the stereo images.

high-resolution local Antarctic DEM with 2- and 4-m spac-ings for 18 sites around McMurdo Sound [15]. The NationalAeronautics and Space Administration also provides IceBridgeLand, Vegetation, and Ice Sensor (LVIS L2) geolocated surfaceelevation products acquired using the airborne LiDAR system.Although the operating airborne system in Antarctica is attrac-tive for its high vertical accuracy (e.g., less than 50 cm) andspatial resolution (e.g., a few meters), it is costly and needsmuch logistical infrastructure; it is therefore inadequate, partic-ularly over large regions. For this reason, spaceborne systemsgain more popularity. In-track stereoscopic data from SatellitePour l’Observation de la Terre-5 High Resolution Stereoscopic(SPOT-5 HRS) were used to generate a 40-m DEM for theAntarctic margins; moreover, the data conveyed root meansquare (rms) elevation errors of less than 10 m [11]. Similarly,SPOT-5 HRS data were used for the Canadian Arctic [8], whichis an 80% ice-covered region. Both studies reported large errorsof over 100 m on slopes that are steeper than 40◦. Few studieshave investigated the use of 1 m or submeter spatial resolutionsatellite images for generating DEM over glaciated regions,although one study tested DEM over polar outlet glaciers witha 0.5–10-m overall vertical error using Worldview 1 and 2 andthe ERDAS LPS eATE algorithm [16].

We propose a new stereo matching method of high-resolutionsatellite images for the generation of accurate DEM over theAntarctic, including steep slope terrains. Although the defini-tion of “high resolution” in the satellite imaging system is notfixed, in topographic mapping, the term implies that it startsapproximately at a ground sampling distance of 10 m [17].The major obstacle when generating accurate DEM with meter-level spacing from the stereoscopic satellite images over steepglaciated regions is that stereo image matching does not workwell in these areas. The homogeneous radiometry within theice and the strong geometric image distortions over the steepterrain limit the use of image matching techniques such asnormalized cross-correlation (NCC) [8], [13]. Furthermore, thehomogeneity within the ice can produce false positives; hence, amore serious problem is that the forward and backward viewingangles result in geometrically dissimilar topographic images,particularly over steeper areas such as ice caps and coastalregions (see Fig. 1). In the image in Fig. 1(a), the terrain facingthe bottom appears to be wide when viewed from the bottom. Incontrast, Fig. 1(b) shows the same terrain, but the image is muchnarrower when viewed from the top down. Steep slopes create

geometrically dissimilar images, and the dissimilarity differsdepending on the viewing angle and direction. This largedissimilarity significantly lowers the matching success ratebecause template-based image matching such as NCC worksbest on relatively flat or smooth surfaces.

To overcome this difficulty, we propose a new templatematching approach that transforms the image patches withinthe matching window. To this end, we first carried out epipolarimage resampling in order to accurately align the viewingangles of the stereo pair parallel to the image rows. Notethat epipolar resampling removes all y-parallax (image rowdifference between the stereo pair) and leaves only x-parallax,which is directly related to height or elevation [19]. We thenused the NCC by transforming the image patches within thematching window along the sample (scaling), line (skewing),and diagonal (scaling and skewing) directions in order to im-prove the match rates over the large geometric dissimilaritywithin the steep areas.

II. PROPOSED METHOD

A flowchart of the proposed method is presented in Fig. 2.Given the high-resolution stereoscopic satellite images with ra-tional polynomial coefficients (RPCs), the positional accuracyof the original RPCs was enhanced using the bias-correctionprocedure with ground control points [18]. The optimal pa-rameterization for the correction may vary depending on thesatellites; however, one pixel level of positional accuracy canbe achieved by introducing only the shift terms for IKONOS.Following the correction of the RPCs, epipolar image resam-pling was carried out. In the epipolar image space, the conjugatepoints in the stereo images are aligned in the same image linesuch that a faster point search is enabled. In addition, the terraindistortion can be easily handled due to the steep slopes becausethe major terrain reliefs are aligned parallel to the image row.To create the epipolar resampled images, a piecewise approachwas followed because either the RPCs or the rigorous sensormodel can be used.

We applied template-based stereo matching, NCC, to theGaussian resampled images using the proposed image patchtransformation scheme. First, NCC matching was carried outfor the image pair at the lowest image pyramid. At this stage,image transformation was not enabled because the resolutionwas too low to consider the local terrain. We then continuedNCC matching in the next image pyramid. If the best cor-relation was higher than the predefined threshold (e.g., 0.8),then the point pair was accepted for the terrain reconstruction.Otherwise, the terrain was analyzed using the lower scalestereo matching results. The image patch of the left image wasthen incrementally transformed while considering the terrainaspect. If the slope was along the sample direction in theepipolar image, the image patch was rescaled along the sampledirection. Next, we skewed the image patch if the slope wasalong the line direction; if the slopes were in the diagonaldirections, the image was resized along the sample direction,and the image skewing was then applied while considering thedirection. As a result, the left image patch was transformedincrementally to a predefined level, and these transformed

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 3

Fig. 2. Flowchart of the study.

image sets were used for the conjugate point search in theright epipolar image. If the best correlation was higher thanthe established threshold, the point pair was accepted forground reconstruction. We continued stereo matching up tothe largest pyramid image to the generated DEM of full res-olution of the input images, e.g., 1-m resolution for IKONOSimages.

A. Epipolar Image Resampling

Epipolar image resampling has three major advantages inthe proposed stereo image matching approach. First, the re-sampling process helps to improve the matching speed byreducing the search area and minimizing the y-parallax. Ac-curate epipolar image resampling was often carried out for amaximum line difference between the stereo pair of one pixellevel. Therefore, the search range can be set to only one pixel

larger than the matching window height. Second, the minimizedsearch area helps to limit the potential inclusion of the matchingoutliers. Third and most importantly, the resampling enablesthe easy estimation and handling of image dissimilarity fromthe diverse viewing directions because a major dissimilarityappeared along the left and right directions in the epipolarresampled image. The epipolar resampled images from Fig. 1are shown in Fig. 3. Note that the viewing angles of the stereopair are aligned along the horizontal direction, which is theflight pass of the satellite.

The epipolar geometry of a satellite pushbroom sensor differssomewhat from that of frame cameras. Previous studies (suchas [20] and [21]) have shown that the epipolar curve shape ofthe pushbroom sensor is hyperbola-like and that the epipolarcurve pairs do not exist for the entire scene. These propertieslimit the accurate epipolar image resampling of the satelliteimages. Thus, in a couple of studies, the method used has

4 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 3. Epipolar resampled images of Fig. 1.

relied on approximate models such as 2-D affine [22], [23];however, this method has shown rather poor results for highapparent-field-of-view sensors with a dynamic trajectory. Thus,a piecewise method was developed to accurately determineepipolar geometry and to resample the epipolar images [19].The piecewise method is based on iterative projections from animage to the ground and from the ground to the paired image.The projection can be carried out using rigorous physical sensormodels or RPCs without further approximations.

B. Template Matching

In-track high-resolution stereoscopic satellite images showpractically no spectral differences between the image pairs; fur-thermore, feature matching methods that are more sophisticatedand computationally expensive are usually not required. There-fore, NCC can be used for stereo image matching using (1).Epipolar image resampling and the pyramid image approachsignificantly reduce the search space along the line and sampledirections, respectively. In addition, the convolution theoremcan be utilized for a better matching speed [24]

ncc=

∑wi=1

∑wj=1

[(Lij−L)(Rij −R)

]√[∑w

i=1

∑wj=1(Lij−L)2

] [∑wi=1

∑wj=1(Rij−R)2

](1)

where L is the image patch in the left epipolar image, R is theimage patch of the search region in the right epipolar image,Lij and Rij are the elements of image patches L and R,respectively, at line i and sample j, L and R are the averagesof all intensity values in images L and R, respectively, and w isthe image patch size.

C. Matching by Transforming the Image Patch

Area-based matching methods, including NCC, inherentlyassume a smooth surface within a small matching window, suchas 7 × 7 pixels or a larger size. This assumption may not be truefor image patches over a steep terrain because different satelliteviewing angles may introduce different terrain relief patterns inthe matching windows.

These problematic relief distortions are explained in theepipolar resampled image, assuming that the terrain has a steepbut constant slope over the window (see Fig. 4). Fig. 4(a)

presents the case of a steep slope facing the left or right. Notethat the dotted arrow indicates the orthogonal direction to theslope, and the elevation along the line is assumed to be constant.The left-facing slope is wide in the left epipolar image, while itappears to be narrow in the right epipolar image. The conversewould be true for a right-facing slope. Due to such dissimilarity,template-based matching would not be successful. Fig. 4(b)shows the case of a steep slope facing up or down. In theright epipolar image, the higher terrain is displaced to the leftcompared to the lower terrain due to the viewing direction.Steep slopes facing the diagonal directions are depicted inFig. 4(c) and (d). The steep slope facing the upper right is widerin the right image, and the higher terrain in the lower left ismore displaced to the left [see Fig. 4(c)]. Conversely, the slopefacing the lower left appears to be narrower, and the upper rightterrain is more displaced to the left. This contrasts the imagein Fig. 4(d), where the slopes facing the upper left and lowerright appear to be narrow and wide, respectively. Furthermore,the high terrain also appears to be more displaced to the left.

To overcome these issues, we transformed the image patcheswithin the matching window. The benefit of this process was tominimize the geometric dissimilarity between the stereoscopicdata. For example, in Fig. 4(a), the simple scaling of the leftimage patch along the sample (horizontal) direction wouldreduce the difference. In the cases presented in Fig. 4(b)–(d),skewing the image patch would also be required. Thus, optimalimage transformation strongly depends on the local slope di-rection. Therefore, we incrementally transformed the left imagepatch along the sample (scaling), line (skewing), and diagonal(scaling and skewing) directions for the best image matching.

The aforementioned image transformation method requiresthe estimation of the slope pattern in order to select an optimalimage transformation. In the pyramid image approach, wecomputed the slope information from the lower scale imagematching. However, we did not need to compute the real groundelevation in this stage; the slope was analyzed in the epipolarimage space itself. Because the x-parallax from stereo matchingwas proportional to the ground height in the epipolar image, weused the x-parallax itself for the slope estimation.

D. Filtering and Refinement

The proposed method did not guarantee any outlier-freematching results; therefore, we refined those results. The firsttarget included the points that had a matching correlationhigher than the threshold but were not located correctly in theright image, i.e., false positives. We filtered out all the pointsthat showed abnormal x-parallaxes if the parallax was largerthan the median x-parallaxes of the neighboring points, e.g.,smoothing using a 3 × 3 median filter. Second, we continuedthe refinement of the conjugate point pairs with low cross-correlations, implying a high possibility of a mismatch. In thisstage, we can use a better slope estimation and a more mini-mized search space from the full-scale image matching results.Note that the lower scale image matching results were usedfor this information in the aforementioned pyramid approach.The refinement was carried out iteratively until no furtherimprovement was found.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 5

Fig. 4. Relief distortion of steep slopes in the epipolar resampled images. (a) Slope facing left or right. (b) Slope facing up or down. (c) Slope facing the upperright or lower left. (d) Slope facing the upper left or lower right.

Fig. 5. Test data location in Antarctica.

III. EXPERIMENT

A. Data

The test area was located in Terra Nova Bay, where JangBogo Station (the second Korean research base) was com-pleted in February 2014 (see Fig. 5). Terra Nova Bay is about64 km long and is located between Victoria Land and RossSea, Antarctica. The Italian and German research stations arealso located on Terra Nova Bay. The temperature rises abovefreezing during the summer (from November to January), andlarge rock and soil areas are clearly visible when the snowmelts. In-track IKONOS stereoscopic data were acquired onJanuary 16, 2011; the specifications are presented in Table I andFig. 6, portraying the IKONOS image. In the previous research,4-m multispectral information of this data was used to monitorvarious land cover types and to define the accurate detailed landcover boundaries [25]. Here, most surfaces are covered withice, and some bare surfaces can be observed at the lower rightcorner where Jang Bogo Station is built. The steep slopes areobserved at the center and the upper left of the image. Also,cliffs are located near Ross Sea at the bottom of the image. Theproduct level of the images was Geo, which is map projected.Note that, in Fig. 5, the upper direction of the IKONOS image is

TABLE ISPECIFICATIONS OF IKONOS TEST DATA

the lower right direction because the upper direction of Fig. 5 isoriented to 0◦ longitude while the IKONOS image is located inthe Universal Transverse Mercator (UTM) 58-Zone (longitude162–168◦). The image coverage is about 11 km × 11 km, andthe target terrain elevation range is approximately 0–1000 m.The convergence angle of the stereo pair was computed to beabout 30◦ from the azimuth and elevation angles. Based onthe preinstalled Global Navigation Satellite System (GNSS)base station (8 km away), four ground control points wereGNSS surveyed in real-time kinematic (RTK) for 10 min usingTrimble R8 at the German station, which is shown in the lowerright corner of Fig. 6(a), with the point locations highlighted inthe boxes in Fig. 6(b). Note that the photogrammetric groundcontrol points require at least less than a half pixel of thespatial resolution, e.g., less than 50 cm for the tested IKONOSimage. This positional accuracy is obtained with GNSS RTKsurveying, which shows the centimeter level of 3-D accuracy.The corners of the building were used for the controls inorder to easily identify the images [see Fig. 6(b)–(c)]. Thesecontrol points enabled a bias correction of IKONOS RPCs by

6 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 6. Test data. (a) IKONOS image in UTM58. (b) Ground control points.(c) GNSS measurements at German base.

Fig. 7. Epipolar image resampling: (a) Resampling each image and(b) anaglyph image that can be stereo-viewed using cyan–red color glasses.The cyan-colored area around the center and the upper right corner of theimage indicates the high elevation region that produced large x-parallaxes inthe stereoscopic data.

introducing only the shift term. After the bias correction, thepositional residuals of the control points dropped from aboutseven pixels to less than one-half of a pixel. In addition, wemanually observed the conjugate point pairs over the entireimage [small circles in Fig. 6(a)] in order to verify the stereomatching quality. A total of 58 points were taken in the digitalphotogrammetric workstation.

B. Epipolar Generation

Using the bias-compensated RPCs, epipolar image resam-pling was carried out using the piecewise approach. Consid-ering the elevation range of the terrain, the minimum andmaximum elevation for the piecewise approach was set at 0 and1000 m, respectively. Next, a second-order polynomial equationwas used for the image transformation. The transformationresiduals were 0.02–0.1 pixels, indicating that practically notransformation errors were introduced. Epipolar image resam-pling conveyed the maximum y-parallax of 1.1 pixels. Fig. 7(a)

Fig. 8. Successful matching points without patch transformation (correlationwas larger than 0.8 in the white region).

shows the epipolar image resampling for the left and rightimages, and Fig. 7(b) shows the anaglyph image. Note that weplaced the cyan and red colors to the left and right of the epipo-lar images, respectively, in order to generate an anaglyph image.Therefore, the cyan colored areas around the center and theupper right corner of the image indicate a high elevation region,which produced large x-parallaxes in the stereoscopic data.

C. Stereo Matching Without Image Patch Transformation

We generated an image pyramid for the hierarchical stereomatching. The image pyramid of the five stages was createdusing Gaussian image resampling such that the lowest imagehad a spatial resolution 16 times lower than the original image.For stereo matching, we used a matching window of 21 ×21 pixels in order to increase the possibility of including ter-rain primitives over the radiometrically homogeneous surface.Window sizes that are too large may show negative effectsbecause the window tends to include various x-parallaxes overthe rugged terrain, thus leading to the failure of stereo matching.

For a selected image patch in the left epipolar image, the con-jugate patch was searched in the right image. The smaller searchregion was preferred for faster computation and for mitigatingthe matching outliers. The line and sample sizes of the searchrange were established using the y-parallax error information ofthe data and the terrain elevation range, respectively. Therefore,the line size of the search region was set to one pixel largerthan that of the image patch. The sample size of the searchregion was then derived from the lower scale image match-ing result. Hence, the results from the lower scale matchingx-parallax estimates over the stereoscopic data, and the initialx-parallax of an image patch in the current matching stage couldbe derived using a linear interpolation. For the lowest scaleimage, the search range along the sample direction was set to−44 to +11 pixels in order to correspond to the elevation rangeof 0–1000 m, respectively. In Fig. 8, the white region indicatesa correlation larger than 0.8, while the black region showsa low correlation in the left epipolar image. Overall, stereomatching did not produce satisfactory results over the ruggedterrain region in the center and in the upper part of the image.Note that the lower left of this image shows the Ross Sea.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 7

Fig. 9. Improvement of matching correlation by transforming the imagepatch. The original image patches are shown in the left column, the transformedpatches are in the middle column, and the conjugate images are shown in theright column. (a) Scaling for the sample direction slope (correlation increasedfrom 0.31 to 0.94). (b) Skewing for the line direction slope (correlationincreased from 0.36 to 0.89). (c) Scaling and skewing for the slope facing theupper right (correlation increased from 0.55 to 0.83). (d) Scaling and skewingfor the slope facing the upper left (correlation increased from 0.56 to 0.95).

D. Stereo Matching With Image Patch Transformation

Stereo matching without a patch transformation did notproduce dense matching point pairs over the steep terrain.Therefore, we applied the image transformation method for theregion where stereo matching failed. In the lowest scale image,only stereo matching was carried out without an image patchtransformation because the relief was not significant at that spa-tial resolution. In the next pyramid stage, we first analyzed theslope aspect from the lower scale stereo matching results andclassified them into four major cases: sample, line, diagonal-1(upper right to lower left), and diagonal-2 (upper left to lowerright). We then applied the image transformation accordingly.

The transformation level was incrementally increased, andeach transformed image was used for NCC matching in orderto compute the correlations. Next, the location showing the bestcorrelation was accepted for the successful matching point inthe right image. Fig. 9(a)–(d) shows the sample image patchesusing this approach. In the three columns, the left column isthe original image patch from the left epipolar image, andthe middle image shows the transformed left image. The right

Fig. 10. Region where stereo matching without patch transformation showedpoor results [white regions in (a) and box regions in (b)]. (a) White regions arewhere the correlation was less than 0.8, but the proposed method gave a highercorrelation. (b) Box regions showing less than 70% success rate with matchingwindows of 21 × 21 pixels.

image of the search region is the right epipolar image. Fig. 9(a)shows the scaling of the image patch along the sample direction.Although stereo matching using the original image patch wasunsuccessful, with a correlation of 0.31, the scaled image inthe middle produced a correlation of 0.94 with the right imagefor a successful conjugate point location. Fig. 9(b) presentsdown-facing slope images. The matching without patch trans-formation indicated a correlation of 0.36; however, the skewingimproved the correlation to 0.89. Note that the upper sideimage was more displaced to the left. Fig. 9(c) and (d) showsthe diagonal direction slopes. Here, we can observe that theimage patches were skewed and then scaled along the sampledirection; moreover, the correlation for the probable conjugatepoint locations was increased from 0.55 to 0.83 and from 0.56to 0.95, respectively.

We then filtered and refined the matching results. The whiteareas in Fig. 10(a) depict the region where stereo matchingwithout patch transformation showed a correlation of less than0.8; however, the proposed method gave a higher correlation.Moreover, the points were dense over the steep terrains, wherethey are located in the center and the upper right corner of theimage. This result corresponds to the black region in Fig. 8.Furthermore, Fig. 11 shows the derived DEM with 1-m spacingfor the test area in Terra Nova Bay.

E. Comparison and Evaluation

Stereo matching without patch transformation produced100 833 114 matching points out of the 107 896 865 points at-tempted, while excluding those produced in the sea region, andthe overall success rate was 93.5%. The proposed transforma-tion method without refinement produced 102 402 187 points,and the success rate was 94.9%. Hence, we obtained a 1.4%increase in the number of matching points. We could also haveobtained 104 613 870 matching points when the refinement wasenabled, which corresponds to a success rate of 97.0% and a netgain of 3.5%.

We continued the comparison by narrowing the regionwhere stereo matching without patch transformation showedunsatisfactory results. Fig. 10(b) shows the sample regionswhere stereo matching without patch transformation had asuccess rate of less than 70%. The overall success rate of thematching without patch transformation was 58.7% over the

8 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 11. One-meter resolution DEM in Terra Nova Bay (UTM58, unit: meter).

Fig. 12. DEM samples from (left) stereo matching without patch transforma-tion and (right) this study.

region; however, when the image patch transformation wasutilized without and with refinement, the success rate increasedto 67.7% and 79.26%, respectively. This indicated that theproposed approach facilitated a net improvement of over 20%in areas that are difficult for stereo matching without patchtransformation.

Fig. 12 presents the DEM samples in order to show theimprovements. The left figures are those from matching withoutpatch transformation, which show topographic noises creatingfalse linear and rectangular features over the steep terrain be-cause the matching point density was sparse. The DEM sampleson the right were obtained from the proposed approach, whichprovided more detail over the steep slopes.

TABLE IIELEVATION DIFFERENCE FROM THE MANUALLY ACQUIRED ELEVATION

To overcome the lack of GNSS measurements over theentire image region for validating the elevation reliability of thegenerated DEM, we utilized the manually observed checkpointspreviously shown in Fig. 6(a). The checkpoints were acquiredby a skilled operator considering the terrain types and thendistributed over the entire image region. The aim of this studyis to develop an automated stereo matching algorithm that hasa performance close to image matching by human operators.Therefore, we compared the elevation of the automaticallygenerated DEM to the elevation of the manually acquiredcheckpoints; the resulting elevation differences are presented inTable II. Stereo matching without patch transformation showedan elevation difference of 2.2 m in rms and a maximum of11.4 m from the manually acquired elevation. However, theproposed method with refinement reduced the difference to1.3 m in rms and the maximum difference to 4.7 m. Fromthis result, it can be concluded that the proposed approach canautomatically generate DEM a few meters close to that obtainedby human operators; this is known as a 1-m level of verticalaccuracy, as presented in [16].

IV. CONCLUSION

To generate a 1-m resolution DEM over the steep terrainin the Antarctic from the in-track high-resolution satellite im-ages, we proposed a hierarchical template matching algorithmaugmented by transforming the template images. Using theepipolar image, we classified slopes into four cases: sample,line, diagonal-1 (upper right to lower left), and diagonal-2(upper left to lower right). Given the assumption that the slopewas steep but constant, the template image was transformedfor stereo matching. The slope information could be obtainedusing the slope analysis from the lower scale hierarchical stereomatching. We tested the proposed method using AntarcticIKONOS stereo images. The result showed that the overallmatching success rate improved from 93.5% to 97.0% for allimage pixels. For the area where stereo matching without patchtransformation produced sparse matching point density, thesuccess rates showed an improvement from 58.7% to 79.26%.This improvement will significantly reduce the workload ofmanual editing over the area. Finally, we generated a 1-mresolution DEM over the glaciated high relief terrain, whichshowed a maximum elevation error of 4.7 m. This was animprovement over the DEM, which had a maximum elevationerror of 11.4 m. Overall, major improvements were made overthe steep slope, as shown in Fig. 12.

A future study will include an absolute accuracy assessmentusing dense GPS control points. In addition, template-basededge matching with transformation for highly shaded regionswill be performed. Heterogeneous satellite image pairs will alsobe studied for greater economical DEM generation.

LEE et al.: AUTOMATED GENERATION OF DEM OVER STEEP TERRAIN IN ANTARCTICA 9

REFERENCES

[1] R. R. Muskett et al., “Airborne and spaceborne DEM- and laser altimetry-derived surface elevation and volume changes of the Bering Glaciersystem, Alaska, USA, Yukon, Canada, 1972–2006,” J. Glaciol., vol. 55,no. 190, pp. 316–326, 2009.

[2] D. G. Vaughan, J. L. Bamber, M. Giovinetto, J. Russel, andA. P. R. Cooper, “Reassessment of net surface mass balance in Antarc-tica,” J. Climate, vol. 12, no. 4, pp. 933–946, Apr. 1999.

[3] E. Berthier, E. Schiefer, G. K. C. Clarke, B. Menounos, and F. Rémy,“Contribution of Alaskan glaciers to sea-level rise derived from satelliteimagery,” Nat. Geosci., vol. 3, no. 2, pp. 92–95, Feb. 2010.

[4] J. L. Bamber, “A digital elevation model of the Antarctic ice sheet de-rived from ERS-1 altimeter data and comparison with terrestrial measure-ments,” Ann. Glaciol., vol. 20, no. 1, pp. 48–54, Jun. 1995.

[5] J. L. Bamber and J. L. Gomez-Dans, “The accuracy of digital elevationmodels of the Antarctic continent,” Earth Planet. Sci. Lett., vol. 237,no. 3/4, pp. 516–523, Sep. 2005.

[6] J. P. DiMarzio, GLAS/ICESat 500 m Laser Altimetry Digital ElevationModel of Antarctica. Boulder, CO, USA: National Snow and Ice DataCenter, 2007.

[7] C. A. Shuman et al., “ICESat Antarctic elevation data: Preliminary pre-cision and accuracy assessment,” Geophys. Res. Lett., vol. 33, no. 7,pp. L07501-1–L07501-4, Apr. 2006.

[8] T. Toutin, “Digital elevation model generation over glacierized regions,”in Encyclopedia of Snow, Ice and Glaciers. New York, NY, USA:Springer-Verlag, 2011, pp. 202–213.

[9] J. L. Bamber, J. L. Gomez-Danz, and J. A. Griggs, “A new 1 km digitalelevation model of the Antarctic derived from combined satellite radar andlaser data—Part 1: Data and methods,” Cryosphere, vol. 3, no. 1, pp. 101–111, May 2009.

[10] H. Zwally et al., “ICESat’s laser measurements of polar ice, atmosphere,ocean, land,” J. Geodyn., vol. 34, no. 3/4, pp. 405–445, Oct./Nov. 2002.

[11] J. Korona, E. Berthier, M. Bernard, F. Remy, and E. Thouvenot, “Spirit.Spot 5 stereoscopic survey of polar ice: Reference images and topogra-phies during the fourth international polar year (2007–2009),” ISPRS J.Photogramm. Remote Sens., vol. 64, no. 2, pp. 204–212, Mar. 2009.

[12] T. Toutin, C. Schmitt, E. Berthier, and D. Clavet, “DEM generation overice fields in the Canadian Arctic with along-track SPOT5 HRS stereodata,” Can. J. Remote Sens., vol. 37, no. 4, pp. 429–438, Aug. 2011.

[13] M. Urai, T. Tachikawa, and H. Fujisada, “Data acquisition strategies forASTER global DEM generation,” ISPRS Ann. Photogramm. Remote Sens.Spatial Inf. Sci., vol. 1–4, pp. 199–202, Jul. 2012, 2012 XXII ISPRSCongress, Melbourne, Australia, 25 Aug.–01 Sep..

[14] A. J. Cook, T. Murray, A. Luckman, D. G. Vaughan, and N. E. Barrand,“A new 100-m digital elevation model of the Antarctic Peninsula derivedfrom ASTER global DEM: Methods and accuracy assessment,” EarthSyst. Sci. Data, vol. 4, no. 1, pp. 129–142, Jan. 2012.

[15] T. Shenk et al., “DEM generation from the Antarctic LiDAR data,” OhioState Univ., Columbus, OH, USA, 2004.

[16] C. C. Porter, P. J. Morin, I. M. Howat, S. Niebuhr, and B. E. Smith, “DEMextraction from high-resolution stereoscopic Worldview 1 & 2 imagery ofpolar outlet glaciers,” presented at the American Geophysical Union, FallMeeting, San Francisco, CA, USA, Dec. 5–9, 2011, Paper C41E-0461.

[17] K. Jacobsen, “High resolution imaging satellite systems,” presented at theEarsel Workshop Remote Sensing Use 3rd Dimension Remote SensingPurposes, Porto, Portugal, 2005.

[18] C. S. Fraser and H. B. Hanley, “Bias-compensated RPCs for sensor ori-entation of high-resolution satellite imagery,” Photogramm. Eng. RemoteSens., vol. 71, no. 8, pp. 909–915, Aug. 2005.

[19] J. Oh, W. H. Lee, C. K. Toth, D. A. Grejner-Brzezinska, and C. N. Lee,“A piecewise approach to epipolar resampling of pushbroom satelliteimages based on RPC,” Photogramm. Eng. Remote Sens., vol. 76, no. 12,pp. 1353–1363, Nov. 2010.

[20] R. Gupta and R. Hartley, “Linear pushbroom cameras,” IEEE Trans.Pattern Anal. Mach. Intell., vol. 19, no. 9, pp. 963–975, Sep. 1997.

[21] T. Kim, “A study on the epipolarity of linear pushbroom images,”Photogramm. Eng. Remote Sens., vol. 66, no. 8, pp. 961–966, Aug. 2000.

[22] T. Ono, “Epipolar resampling of high resolution satellite imagery,” inProc. Joint Workshop ISPRS WG I/1, I/3/IV/4 Sensors Mapping Space,Hannover, Germany, 1999, [CD-ROM].

[23] M. Morgan, “Epipolar resampling of linear array scanner scene,” Ph.D.dissertation, Univ. Calgary, Calgary, AB, Canada, 2004.

[24] J. Oh, C. K. Toth, and D. A. Grejner-Brzezinska, “Automatic geo-referencing of aerial images using high-resolution stereo satellite im-ages,” Photogramm. Eng. Remote Sens., vol. 77, no. 11, pp. 1157–1168,Nov. 2011.

[25] S. H. Kim and C. H. Hong, “Antarctic land-cover classifier using IKONOSand Hyperion data at Terra Nova Bay,” Int. J. Remote Sens., vol. 33,no. 22, pp. 7151–7164, Nov. 2012.

Changno Lee received the B.S. and M.S. degreesin civil engineering from the Seoul National Univer-sity, Seoul, Korea, in 1987 and 1989, respectively,and the Ph.D. degree from Purdue University, WestLafayette, IN, USA, in 1999.

He was a Senior Researcher at the Electronicsand Telecommunications Research Institute and cur-rently works as an Associate Professor in Civil Engi-neering at the Seoul National University of Scienceand Technology. His research interests are sensormodeling, robust estimation, and automated feature

extraction from imagery.

Jaehong Oh received the B.S. and M.S. degrees incivil engineering from the Seoul National University,Seoul, Korea, in 1999 and 2001, respectively, andthe Ph.D. degree from The Ohio State University,Columbus, OH, USA, in 2011.

He was a Senior Researcher at the Space Imag-ing Asia and Electronics and TelecommunicationsResearch Institute and currently works as an Assis-tant Professor in Civil Engineering at the ChonnamNational University, Gwangju, Korea. His researchinterests cover georeferencing techniques including

photogrammetric sensor modeling and stereo image processing.

Changhee Hong received the B.S. and M.S. degreesin geospatial information technology from Inha Uni-versity, Incheon, Korea, in 1997 and 1999, respec-tively, and the Ph.D. degree from Seoul NationalUniversity, Seoul, Korea, in 2006.

He works as a Senior Researcher at the Informa-tion and Communication Technology Convergenceand Integration Research Division in the Korea Insti-tute of Construction Technology, Goyang-si, Korea.His research interests cover ecology-based remotesensing and extreme cold region monitoring using

remote sensing.

Junhee Youn received the B.S. and M.S. degreesin civil engineering from Yonsei University, Seoul,Korea, in 1996 and 1998, respectively, and the Ph.D.degree from Purdue University, West Lafayette, IN,USA, in 2006.

He was a Senior Consultant at SAMSUNG SDSand currently works as a Senior Researcher atthe Korea Institute of Construction Technology,Goyang-si, Korea. His research interests cover dis-aster management with geospatial data, man-madeobject extraction from multiple sensors, and real-

time photogrammetry with mobile mapping system.