DEM Generation Using a Digital Large Format Frame Camera

DEM Generation Using a Digital Large Format Frame Camera Joachim Höhle Abstract Progress in automated photogrammetric DEM generation is presented. Starting from the procedures and the performance parameters of automated photogrammetric DEM generation, the results of some practical tests with large scale images are presented. The DEMs are derived from images taken by a digital large-frame aerial camera and checked by reference data of superior accuracy. In average, a vertical accuracy of s h 13 cm or 0.20 per thousand of the mean flying height above mean terrain has been achieved. Some recent innovations in digital large-frame cameras and in the processing software give hope for even better results. In comparison with results from film-based cameras, it can be stated that both technologies are able to produce very dense and accurate DEMs. Introduction The generation of Digital Elevation Models (DEMs) has become important in recent years. Applications such as the production of orthoimages and of 3D city models require a high accuracy and especially higher production rates. For many years, photogrammetry has been used as the standard method of DEM generation. Manual methods proved to be accurate, but are too slow. Correlation of overlapping images enabled automated procedures but requires editing. This semi-automatic production can now be improved by means of new tools and new procedures. Digital large-frame cameras seem to enable a more accurate and more reliable DEM generation due to their higher radiometric and geometric resolution. Two types of cameras are in use for the accurate photogrammetric DEM generation: large-format frame cameras and line scanner cameras. There are advantages in both types, but this article will deal with frame cameras only. The occurrence of airborne laser scanning gave new possibilities to acquire DEMs, especially in difficult landscapes like forests or built-up areas. The generation of DEMs may require only one of the two technologies, but also a combination of both methods may be useful from a technical point of view. The performance of automated photogrammetric DEM generation has to be re-evaluated today because various new tools are available. It is the objective of this article to investigate the performance of a digital largeformat camera with regard to DEM generation. A comparison with the results obtained with analogue cameras may show whether the new type of cameras has the same or even a better performance. Recently announced innovations in the processing software will also be presented in order to judge Aalborg University, Department of Development and Planning, Research Group of Geoinformatics, Fibigerstraede 11, DK9220 Aalborg, Denmark (jh@land.aau.dk). the current potential of the photogrammetric method for DEM generation. First the categories of DEMs and the procedures in the automated photogrammetric DEM generation will be summarized in order to understand its performance parameters. Categories of DEMs and Their Applications There are various ways to categorize DEMs. First of all, there are two main types, the Digital Surface Model (DSM) and the Digital Terrain Model (DTM). The DTM is the bare ground, and the DSM contains elevations on top of buildings and vegetation as well. The applications of the two types are different. The design of highways and other engineering tasks require a DTM, and the generation of true orthoimages and the planning of towers for cell phones require a DSM. Furthermore, the accuracy of DTMs and DSMs can be very different. Highly accurate DEMs may have standard deviations of less than 0.5 m; less accurate DEMs may have a standard deviation of more than 0.5 m. The density of the points will then differ as well. The data can be collected and stored in different formats. The applications of the DEM require either a triangle (TIN) structure or a grid structure. Table 1 gives an overview of the various DTM categories and their applications. The method of acquisition will be different according to the accuracy requirements. The type of terrain may have influence on the selection of the data acquisition method. Procedures in the Automated Photogrammetric DEM Generation The present procedures with up-to-date photogrammetric systems will be outlined in the following. The DEM generation may start from scratch and a new DEM will be derived by means of at least two images, each one in several levels of resolution. The method and results with it are described in the literature, for example in Gülch (1994) and Heipke (1995). Professional photogrammetric workstations contain elaborated software packages, and the photogrammetric practice uses the automated photogrammetric generation of DEMs on a routine basis with images taken by analogue cameras. If a DEM already exists, its accuracy and completeness may also be improved. The existing heights are used as approximations in the generation of the DEM. The corrections to an existing DEM can also be derived from horizontal Photogrammetric Engineering & Remote Sensing Vol. 75, No. 1, January 2009, pp. 87 93. 0099-1112/09/7501 0087/$3.00/0 2009 American Society for Photogrammetry and Remote Sensing PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING January 2009 87

TABLE 1. OVERVIEW ON THE VARIOUS DEM CATEGORIES AND THEIR APPLICATIONS accuracy density format applications acquisition method DTM 0.5 m 1 5 m TIN design of highways and laser scanning & other engineering tasks photogrammetry 0.5 m 5 50 m grid orthoimages photogrammetry DSM 0.5 m 1 5 m TIN true orthoimages laser scanning & photogrammetry 0.5 m 5 50 m grid planning of cell phone towers photogrammetry parallaxes between two overlapping orthoimages. The process can be carried out automatically. This approach was first presented in Norvelle (1994) and further refined and tested by other authors, for example Georgopoulos and Skarlatos (2003) and Potuckova (2004). Both approaches have the following steps of production: Acquisition of images, georeferencing of images, DEM generation or DEM improvement, editing, and quality control. These steps are shortly described in the following in order to discuss the performance parameters and the test results thereafter. Acquisition of Images The acquisition of images needs careful planning of the flights. The sensors and the flying altitude have to be chosen so that the required DEM accuracies can be achieved. Cameras have to be calibrated and their relationship to other sensors such as Global Positioning Systems (GPS) and Inertial Measurement Units (IMU) has to be known. The image acquisition has to be completed when vegetation has not yet grown. Furthermore, the atmospheric conditions and sun angle have to enable good image quality (brightness, contrast, short length of shadows). The film-based images must be scanned with a precision scanner. The digital images then have to be converted into image pyramids, compressed, and saved. Georeferencing of Images The digital aerial images need to be supplemented with the data of interior and exterior orientation. The exterior orientation of the images is usually determined by aerotriangulation. A few ground control points (GCPs) are required. The use of GPS and IMU can further reduce the number of necessary GCPs or avoid them completely. This direct georeferencing requires calibration of the bore sight, which has to be determined by means of a test field. The socalled Integrated Sensor Orientation uses the GPS/IMU data and the tie-point measurements in a joint adjustment (EuroSDR 2002). This approach enables high accuracy and is becoming the standard method for the georeferencing of images to be used in the generation of highly accurate DEMs. DEM Generation The generation of DEMs is accomplished in softcopy workstations by means of stereo models. It can be done manually, semi-automatically, or fully automatically. The manual work consists of collecting breaklines and mass points which both characterize the terrain. The semi-automated approach guides the operator to the next point, where it is decided whether the automatic measurements should be used or not. For the collection of DTMs, for example, the operator will not use measurements on top of houses and trees. The automated approach uses correlation techniques. The matching of corresponding image parts can be area-based or feature-based. Features are, for example, edges which are extracted from the images in advance. Parameters in the software packages have to be selected according to the terrain type and image quality. The automatically generated DEMs represent the surface; they are therefore DSMs. Blunders may occur at image areas with low texture and low contrast. Editing The conversion of the DSM to the DTM (bare earth data) and the removal of blunders requires editing of the raw data. Also, thinning of the data may be necessary. The visualization of the automatically derived heights as colored dots in the image pair makes it possible to detect errors by stereo vision. The human operator can remove erroneous heights, add manually measured spot heights, and breaklines. Derived contour lines may also be used for checking and editing of the generated DEM. These visual procedures can be supplemented by automated filtering of the DEM, which will reduce the DSM to bare earth data. Also, blunders in the digitally correlated data set may be detected and removed automatically. Quality Control Quality control is carried out by means of accurate reference data. Checkpoints can be derived in the process of aerotriangulation or by ground measurements. Their accuracy should be higher than the accuracy of the DEM points, and the sample size should be sufficiently large. The assessment of the accuracy should include the vertical as well as the horizontal accuracy. The accuracy measures are derived for only a few samples. The checking of large DEM areas requires automated methods. In EuroSDR (2006) various automated checking methods are described and tested. They use aerial images of the same (large) scale as for orthoimage production. Performance Parameters of the DEM Generation The imagery should be taken with the proper sensors. There is a strong movement to using digital cameras so that a completely digital work flow can be achieved. One of the advantages of digital cameras is the higher radiometric resolution than in film-based cameras. Many more (4,096) intensity values of 12-bit (or more) data are stored per spectral band. Details in the shadows are therefore easier to recognize. The geometric resolution (pixel size) of digital frame cameras is very small, for example 9 mm. The format of the output image and the parameters of the lens system (camera constant, lens distortion, resolution) are important for accurate and economic work. If stereo pairs with 60 percent overlap are used, the base/height ratio is considerably smaller than at analogue cameras of the 230 mm 230 mm format and equipped with wide-angle lenses (typical camera constant 153 mm). This unfavorable base/height ratio for the digital large-format frame cameras may reduce the accuracy of the DEMs, if the matching accuracy cannot be improved by a factor of two when using the same image scale. Table 2 shows some of the performance parameters for existing digital large-format frame cameras. The deliverable images are panchromatic, normal color, and false-color, and are all produced from a single flight 88 January 2009 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

TABLE 2. LARGE-FORMAT FRAME CAMERAS AND THEIR PERFORMANCE PARAMETERS pixel image size image base at camera base to size [mm] 60% overlap constant height producer camera name [mm] [mm] [mm] ratio Intergraph DMC 12 165.9 92.2 36.9 120 0.31 Vexcel / MS UltraCam D 9 103.5 67.5 27.0 101 0.27 DiMAC Systems DiMAC 2.0 6.8 71.4 49.0 19.6 80 0.25 mission. The color images are produced by pan-sharpening (for example, the DMC or UltraCam) or directly by means of the CCD sensors (for example, DIMAC). In the UltraCam camera the images of four spectral bands (red, green, blue, and near-infrared) are taken by four additional lenses with smaller camera constants. The eight CCDs capture images which are used to generate three large-format images of different spectral characteristics. The panchromatic image is fused by means of four images and nine CCD images Additional technical details on digital aerial cameras can be found in the literature, for example in Sandau (2005). In order to meet a specified DTM accuracy, the flying altitude above mean terrain can be calculated by formula: h s h # c # b h s px where h is the flying altitude above mean terrain, s h is the required DEM accuracy (standard deviation), c is the camera constant, b/h is the base to height ratio, and s px is the matching accuracy (standard deviation). The matching accuracy or parallax accuracy (s px ) depends on the quality of the cameras, the contrast and texture in the image, the pixel size, and the applied algorithm. It can be derived by formula: s h # b h s px where m b is the image scale number. If the matching accuracy and the required DEM accuracy are known for a camera system, flying height and image scale can then be found for a DEM project (with its specified accuracy) using Equations 1 and 3: M b 1 m b c h. The pixel size on the ground or the Ground Sampling Distance (GSD) is then: m b GSD pel # mb where pel is the pixel size of the digital image. With the known GSD and the number of pixels across the flight direction, the coverage across track (swath width) can be calculated. All the large-frame cameras use the longer side of the rectangular format across-track in order to reduce the number of strips. This fact is, however, not in favor of a high DEM accuracy. The parameters in the automated DEM generation consist of the number of levels in the image pyramids, the size of the matching windows, (1) (2) (3) (4) thresholds for the correlation coefficient, and thresholds for the accuracy of the least squares matching. The selection of proper values for these parameters has a strong influence on the results. The density of grid posts can be very high; practically each pixel can have an elevation. Normally, certain grid spacing is selected and an elevation is determined for each grid post. Elevations can only be determined at positions where good conditions for correlation exist. The estimated accuracy of the determined elevation can be visualized by colored dots which are displayed on top of the stereo pair or on top of an orthoimage. The visual inspection will detect errors or problems. The automatic detection and elimination of blunders is an important feature of advanced DEM extraction programs. The applied methods compare a generated elevation with the elevations of the neighborhood. A statistical evaluation will decide whether the generated height is a blunder. The editing of the raw DEM data is an interactive process. Editing tools include the visualization of the DEM as colored dots or contour lines on top of the stereo model. Profiles or perspective views may also be used to detect problem areas. The operator re-measures spot elevations and supplements with breaklines and spot elevations. New contour lines are generated on the fly, and the operator can make decisions whether the generated DEM can be accepted. The checking of the DTM can also be done automatically. If map data are available, logical checks can be carried out by software. For example, the elevations at shorelines of lakes should be equal, and heights at creeks and rivers should continuously decrease in the direction in which water flows. Filtering of the DEM with a proper algorithm can automatically reduce the DSM to bare earth data. The performance parameters in the editing are the times necessary to correct the raw data. The amount of data which can be handled is a performance parameter as well. If the programs of automated DTM generation produce a minimum of blunders and gaps, then the editing time will also be short. The possibility of using other data (vector maps, orthoimages) during the editing, and/or to represent the DEM in various forms is an important feature of such editing programs. To support the quality control of DEMs, the vertical accuracy has to be determined. The accuracy measures to be derived include the Root Mean Square Error (RMSE), the number of blunders, the systematic error (average error), and the standard deviation. Checkpoints of superior accuracy (at least by a factor of 3) and of a sufficient number (20 or more) are usually used in order to derive the accuracy measures of a DTM (Maune, 2007). The checkpoints can be derived by field measurements, for example, by means of GPS. The formulas for the aforementioned accuracy measures are given in Table 3. Practical Tests As it was shown above, the results of the DEM generation depend on many parameters. In order to meet the specifications of a DEM task, it is necessary to know the flight height which is based on the matching accuracy of the applied PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING January 2009 89

TABLE 3. ACCURACY MEASURES FOR DEMS Difference from reference data (for a pair i) Number of tested points Root Mean Square Error Maximum difference Definition of a blunder (threshold) Number of blunders Number of points without blunders Mean h i n RMSE K ƒ h max ƒ ƒ hƒ 3 RMSE N n n N n a h i i 1 m n n a h 2 i i 1 n Standard deviation s K n a i 1 ( h i m) 2 (n 1) camera system. It is therefore an objective to know the matching accuracy for the digital camera used. Practical tests will be carried out in order to determine this value for the digital large-format camera used. The tests are based on large-scale imagery taken by the digital large-frame camera, UltraCam D. The swath width was 1,150 m, and the ground sampling distance about 6 cm. Altogether, five models were evaluated. The test areas are situated in the suburbs of Aalborg, Denmark, and all of them can be characterized as open terrain covered with grass. A few houses, trees and paths were also present. The differences in elevation do not exceed 30 m, the average slope was below 10 percent, and the form of the terrain is rather smooth. Figure 1 depicts one of the test areas. The test area has texture and contrast, and the quality of the image is good. The other test areas were similar. Each test area covered about 4 percent of the generated DEM. The test areas were also surveyed by means of GPS/RTK, and about 60 terrain points could be used for checking. Their accuracy was s 2 cm (standard deviation) and thereby of superior accuracy. The generation of the DEMs was fully automatic using the program Image Station Automatic Elevations, version 5.1, of Z/I Imaging (Z/I imaging, 2006). The spacing of the derived grid posts was different in the five models averaging 5 m. Editing and filtering of the DEM data was not done. Checkpoints were on the terrain ( bare earth ) only. The data of the test are summarized in Table 4. The accuracy of the DEMs can be found by comparing the elevations of the test points with the values obtained by interpolation in the (automatically) derived DEM grid. Blunders are detected by a threshold which is defined by ƒ hƒ 3 RMSE. For the computation of the systematic error (bias) the elevation errors larger than the specified threshold are removed. The remaining errors are reduced by the systematic error, and standard deviations are calculated. The formulas used can be seen in Table 3. Several models have been processed, and the RMSE, average error and standard deviation are presented in Table 5 and depicted in Figure 2. The average value for the root mean square error amounts to RMSE 14 cm or 0.20 per thousand of the flying height above average terrain. The standard deviation is about the same (s 13 cm) because the systematic error is very small. The matching accuracy after Equation 2 amounts to 6 mm or 0.6 pixel. This accuracy is based on ground truth and can therefore be used for the planning of the flight mission using Equations 1 and 3. A required DEM accuracy can then be achieved. For example, if a specified accuracy Figure 1. Orthoimage of the Test Area #5 at the DEM test with the UltraCam D. TABLE 4. DATA OF THE TESTS WITH THE DIGITAL LARGE-FORMAT FRAME CAMERA ULTRACAM D terrain type built-up area flight altitude 662 m image scale 1:6530 base/height ratio 0.28 GSD 6 cm swath width 1,150 m DEM generation grid spacing 5 m method fully automatic, no approximations threshold correlation coefficient ( 0.75) editing non checkpoints number 25 accuracy 2 cm (standard deviation) of s h 0.185 m ( 0.61ft) has to be met, the required flying height (above average terrain) for the UltraCam D camera (used with 60 percent forward overlap) is then: h 0.185m # (101 # 0.27/0.006) 0.185m # 4545 841m. The images taken have then a scale of M b 1:8325 and a ground sampling distance of GSD 0.0009 8325 7.5 cm. Comparison with Results from a Film-based Camera In order to evaluate the results with the digital camera, a few tests with a film-based camera (Zeiss RMK-TOP, format 23 cm 23 cm, camera constant 153 mm) have been carried out as well. The base to height ratio at the usual 60 percent forward overlap is b/h 0.60. The images were 90 January 2009 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

TABLE 5. RESULTS OF DEM GENERATION WITH THE DIGITAL LARGE-FORMAT FRAME CAMERA ULTRACAM D accuracy measure / test # 1 2 3 4 5 average number of check points 156 25 49 25 111 73 RMSE [cm] 23 11 6 17 12 14( 0.22 h) average error [cm] 2 2 4 10 1 1 standard deviation [cm] 23 11 4 13 12 13 ( 0.20 h) matching accuracy [mm] 10 5 3 7 5 6 ( 0.6 pel) taken from different flying heights, and their scale varied between 1:3 000 and 1:25 000. The photographs were digitized in a precision scanner with a pixel size of 15 mm and 21 mm. The three test areas can be characterized as a mixture between open terrain and built-up areas. The DTM generation took place with different spacing of the grid posts (1 m and 25 m). Elevations of new points were interpolated at the position of the checkpoints. The number of checkpoints was above 90, and their accuracy was at least three times better than the automated generated elevations. The results of the comparison between the elevations of checkpoints and the automatically generated (and interpolated) elevations are presented in Table 6. The achieved accuracy for the three models was in average RMSE 0.19 per thousand of the flying height above average terrain elevation. The computed matching accuracy was in average s px 17 mm or 1 pixel. The result for the matching accuracy is based on accurate reference values, which are determined from images of a much lower flying height or from ground truth. Details on this investigation are published in Höhle and Potuckova (2006). Other investigations on DEM generation with analogue cameras derived similar results. In Saleh and Scarpace (2000) for example, scanned photographs with standard base/height ratio (b/h 0.6), but of various pixel size, were used. An average matching accuracy of s px 18 mm can be derived from the published RMS values and the flight parameters. A matching precision has been derived for terrain of different slope in Karras et al. (1998). The achieved values differ between 0.4 and 0.7 pixels. The reference values were derived from manual measurements with the same images. Discussion of the Results The achieved accuracy of DEMs with a digital large-frame camera and large-scale images (m b 6300 and GSD 6 cm) was s 13 cm or 0.20 per thousand of the flying height above average terrain. There was no significant difference from the results with a film-based camera. The digital camera used (UltraCam D ) can obviously compensate for its drawbacks (unfavorable base/height ratio and smaller format) by means of a higher matching accuracy. The computed matching accuracy in the tests amounted to s px 6 mm at the digital camera and s px 17 mm for the analogue camera. The reasons for the higher matching accuracy are very likely the higher radiometric resolution and the way in which the forward motion is compensated. The analogue cameras move the film mechanically, but the digital cameras integrate the intensity values (digital numbers) of the pixels during the time interval of exposure. This electronic approach leads to higher image quality and the matching accuracy certainly benefits from it. The investigation with the digital large-format camera is based on five stereo models. They contained open terrain and built-up areas. Other categories of landscape, e.g., wooded areas, will produce different results. In order to specify accuracy standards (as it is required in many countries), tests with similar and other terrain types have to be carried out. New Developments In recent months, some improvements in the tools regarding the DEM generation have become known. Furthermore, the digital camera is widely accepted as the image acquisition tool, and the production of DEMs, orthoimages, and photorealistic 3D models can now be realized in a completely digital workflow with a high degree of automation and high production rates. The new UltraCAM X of Microsoft Photogrammetry has a small pixel size in the nine CCDS (7.2 mm), a new lens system (c 100 mm for panchromatic and c 33 mm for color and color-infrared images), and a frame rate of 1.35 seconds (Gruber, 2007). These improvements may change the methodology in DEM generation. Pan-sharpened color images of 80 percent forward overlap at high ground resolution (for example with GSD 5 cm) can be produced. The DEM can then be derived from multiple images, especially if a side overlap of 60 percent is used. Blunders and gaps in the DEM will be reduced. The accuracy of the DEM can further be improved by applying a calibration grid. Lens distortion and possible shifts of the CCDs due to a temperature change during the photo flight can then be corrected. The correction grid may be created by TABLE 6. RESULTS OF DEM GENERATION WITH A FILM-BASED LARGE-FORMAT FRAME CAMERA (RMK-TOP) accuracy measure/test # 1 2 3 average Figure 2. Examples of achieved absolute DEM accuracy with large-scale images (1:6 300) with the digital largeformat camera UltraCAM D. number of check points 2,033 116 94 748 RMSE [ h] 0.18 0.13 0.25 0.19 standard deviation [ h] 0.18 0.09 0.07 0.11 matching accuracy [mm] 17 12 23 17 matching accuracy [pel] 0.8 0.6 1.5 1.0 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING January 2009 91

self-calibration in the aerotriangulation program and then applied in the DEM program (Jacobsen, 2007). Other improvements in software concern the DEM programs. In several programs (e.g., Match-T 5.1 of Inpho GmbH), the single stereo model is replaced by a number of images which contribute to the generation of a very dense point cloud. Besides the X-Y-Z coordinates, each point has also information about color, correlation, and accuracy values. From the dense point cloud, an accurate and reliable DEM is derived by noise filtering, reduction to bare earth, and thinning. Improvements are achieved in the matching algorithms. In the program NGATE of BAE systems, for example, the matching takes place for each pixel which leads to very dense point clouds. Furthermore, the size of the search window is adapted to the height differences in the window area. In built-up areas, edges are derived from the images first. Each pixel lying on an edge is then matched with a pixel of the corresponding edge. The results are more reliable so that considerably less editing work is required. More information can be found in Zhang et al. (2007). New developments in the editing of DEM data happened as well. Special editing stations like DTMaster of Inpho GmbH were created. The program package (version 5.01) can handle up to 50 million points at the same time and uses stereo vision and stereo measurements for checking and improving of DEMs in an interactive process. Automated features are also integrated, for example, plausibility checks, deletion of double measured points, and detection of blunders and gaps. Such editing stations can handle point clouds from photogrammetry and from laser scanning. The tasks and the problems in editing are the same for both technologies. The developments of editing programs in both fields have stimulated each other. Improvements in the editing are also very necessary because this work is the bottle neck in the efficient generation of accurate and reliable DEMs. The quality control of DEMs requires reference data of superior accuracy. Ground surveying is usually used for accurate DEMs, but it is expensive and only a few checkpoints are used. Automated methods of checking are developed and tested in a recent EuroSDR project (EuroSDR, 2006). The methods are based on photogrammetry. The applied imagery is taken from lower flying heights than the one for DEM generation. The results showed that an automated quality control is possible and efficient. More details on this research work are published in Höhle and Potuckova (2006) and Höhle (2007). Relation of Photogrammetric DEM Generation and Airborne Laser Scanning Accurate DEM data of high-density can also be derived by laser scanning. This acquisition method does not need sunlight and texture on the surface. The connection of strips cannot be done by single points but by area elements only. The achievable accuracy depends on the performance of the GPS/IMU. An in-flight calibration of the boresight is necessary and requires a suitable test field. Positional errors may occur and should always be checked. Laser scanning has advantages in urban and in forest areas. When applying photogrammetry, the images within and across the strip are connected by tie-points. This network of rays can be used to determine the parameters of exterior orientation. GPS/IMU data are not necessary. The structure and breaklines of the terrain can be derived and contribute to DEMs of high quality. The images taken may be used for compilation of vector maps and for orthoimage production as well. Both technologies have their advantages and the tasks will decide which one should be used. A combined use would be ideal. Laser scanners already use digital medium-format cameras. By means of the images, the point clouds can be interpreted much better. One GPS/IMU unit can be shared when both systems are combined. The processing of the data including editing has also to be integrated into one system. Whether such a combination is an economic solution has to be evaluated. Conclusions During the last few years many changes have occurred in the generation of DEMs. Automated photogrammetric methods can now use images of digital large-format cameras as well as efficient matching and blunder-detection algorithms. The editing tools have become more efficient using automatic filtering and interactive procedures. The presented results of practical tests with the UltraCam D camera indicate that vertical accuracies of s 0.13 m (or 0.2 per thousand of the flying height above average terrain) can be achieved using images with a ground resolution of GSD 0.06 m. The derived matching accuracy is in average s px 6 mm. This accuracy is based on ground truth and can therefore be used for planning of the flying height (or the image scale and ground sampling distance) in order to meet a required accuracy. The spacing of DEM posts can be as small as the ground sampling distance. Additional investigations revealed that there was no significant difference from the results with a film-based aerial camera which has a larger format and a longer focal length than the used digital camera. Recent developments concern improvements in the geometric resolution (pixel size) at the digital large-format frame cameras. Systematic image deformations can be handled by improved postprocessing and use of calibration grids. A change from the stereo model approach to a multi-image approach may also improve the accuracies of DEMs. Images should then be taken with a higher longitudinal and lateral overlap. Changes in procedures concern the use of approximate DEM data and the derivation of corrections for an approximate DEM. This can efficiently and accurately be done by means of automated parallax measurements in two orthoimages. The editing of the raw DEM data remains the bottle neck in the automated DEM production; it requires interaction with the human operator. Such interactive editing can use stereo vision and map data for checking, decision-making, and re-measurement. Digital Surface Models and Digital Terrain Models can be produced at the same time. In comparison to lidar, the photogrammetric method has several advantages, for example, easy interpretation of objects, measurement of accurate break lines and of characteristic terrain points, and universal use of images for other mapping tasks. The digital large-frame camera is also a valuable supplement to a lidar system, and a simultaneous use of both systems may happen in the future. Acknowledgments The author wants to thank M. Potuckova for support in programming. Students of Aalborg University contributed to this article with ground surveying and processing of the data. B. Nørskov is thanked for her improvements of the English language. References EuroSDR, 2002. Integrated Sensor Orientation, Test Report and Workshop Proceedings, EuroSDR Official Publication No. 43, ISSBN- 3-89888-864-9, 297 p. 92 January 2009 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

EuroSDR, 2006. The EuroSDR Test: Checking and Improving of Digital Terrain Models, Official Publication No. 51, ISSN 0257-0505, 142 p. Georgopoulos, A., and D. Skarlatos, 2003. A novel method for automating the checking and correction of digital elevation models using orthophotographs, The Photogrammetric Record, 18(102):156 163. Gruber, M., 2007, UltraCam X, The new digital aerial camera system by Microsoft Photogrammetry, Proceedings of the Photogrammetric Week 2007, pp.137 145. Gülch, E., 1994. Erzeugung digitaler Geländemodelle durch automatische Bildzuordnung (Generation of digital terrain models by means of automatic image matching), Deutsche Geodätische Kommission, Reihe C., Heft 418, ISBN 3 7696 9462 7, 167 p. Heipke, C., 1995. State-of-the-art of digital photogrammetric workstations for topographic applications, Photogrammetric Engineering & Remote Sensing, 61(1):49 56. Höhle, J., and M. Potuckova, 2006. The EuroSDR Test: Checking and Improving of Digital Terrain Models, EuroSDR Official Publication No. 51, ISSN 0257-0505, 136 p. Höhle, J., 2007. The EuroSDR Project: Automated checking and improving of digital terrain models, Proceedings of the ASPRS 2007 Annual Conference, Tampa, Florida, (American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland), unpaginated CD-ROM. Jacobsen, K., 2007. Geometry of digital frame cameras, Proceedings of the ASPRS 2007 Annual Conference, Tampa, Florida, (American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland), unpaginated CD-ROM. Karras, G.E., D. Mavrogenneas, D. Mavrommati, and N. Tsikonis, 1998. Tests on automatic DEM generation in a digital photogrammetric workstation, International Archives of Photogrammetry and Remote Sensing, Vol. XXXII, Part 2, Commission II, pp.136 139. Maune, D.F. (editor), 2007. Digital Elevation Model Technologies and Applications: The DEM User Manual, Second edition, ISBN 1-57083-082-7, 655 p. Norvelle, F.R., 1994. Using interactive orthophoto refinements to generate and correct digital elevation models (DEMs), Proceedings of Mapping and Remote Sensing Tools for the 21 st Century, American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, pp. 134 142. Potuckova, M., 2004. Image Matching and its Application in Photogrammetry, Ph.D. dissertation, Czech Technical University in Prague, 132 p. Saleh, R., and F. Scarpace, 2000. Image scanning resolution and surface accuracy: Experimental results, International Archives of Photogrammetry and Remote Sensing, Amsterdam, Vol. 32(B2). Sandau, R., 2005. Digitale Luftbildkamera - Einführung und Grundlagen (Digital Aerial Camera - Introduction and Basics), Wichmann Verlag, ISBN 3-87907-391-0, 342 p. Zhang, B., S. Miller, S. Walker, and K. Devenecia, 2007. Next generation automatic terrain extraction using Microsoft Ultra- Cam imagery, Proceedings of the ASPRS 2007 Annual Conference, Tampa, Florida (American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland), unpaginated CD- ROM. Z/I imaging, 2006. User s Guide for Image Station Automatic Elevations (ISAE). (Received 15 November 2007; accepted 10 January 2008; revised 18 January 2008) PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING January 2009 93