User Manual - Institut für Photogrammetrie und GeoInformation
Contents
1. Groups of trees blurred Extracted Table 5 summary of qualitative analysis of the used matching algorithms Building heights determined by matching are compared with reference heights from aerial images with 7cm GSD Some of the buildings have been changed after aerial imaging and before imaging from space The differences of the DSMs from IKONOS and GeoEye 1 minus the reference DSM based on the aerial images may be influenced by building changes Only building heights existing in all data sets have been taken for comparison The newly erected buildings red circle in figure 6 have been avoided in the comparison with reference data as shown in figure 6 Furthermore some gaps and shadow areas in the reference DSM also have been avoided In order to derive building heights from the DSMs the heights of 4 points on the roof and close to the building corners have been averaged for each building The height of the selected buildings ranges from 6 m to 18 m In the Riyadh test area all buildings have flat roofs In order to quantify the performance of matching algorithms the generated building heights from IKONOS by LSM DP and SGM have been compared manually with corresponding heights from reference data For the independent check 75 buildings have been selected These buildings have a random location within the scene The investigated building heights are average values of 4 points on the roof close to the building corners The2. 2 4 2 4 LL LL 2 2 0 a N 0 5 4 3 2 4 0123 4 5 5 4 3 2 1 012 3 4 5 Height differences m Height differences m Fig 7 Histograms of the height differences left LSM right SGM Riyadh height base 1 5 checked with 125 buildings LSM SGM Standard deviation of reference heights 0 2 0 2 RMS of GeoEye 1 height differences 0 7 0 5 SZ of GeoEye 1 height differences 0 6 0 4 RMS of x parallax 0 5 0 3 Table7 Accuracy of GeoEye 1 building heights The vertical accuracies of the GeoEye 1 DHM are below 0 7 m corresponding to 0 4 pixels for the x parallax for LSM and in the range of 0 5 m or 0 3 pixels for the x parallax for SGM Figure 8 shows a Satisfying normal distribution of the results LSM results have a very small bias of 0 07 m and the SGM results 0 06m Histogram of AH LSM reference Histogram of AH SGM reference 15 20 RMS Z 0 7 RMS Z 0 5 Mean 0 07 15 Mean 0 06 gt 10 gt 5 5 3 mi 10 D D E 5 ic 5 0 0 5 4 3 2 1012 3 4 5 5 4 3 2 1 012 3 4 5 Height differences m Height differences m Fig 8 Histograms of the height differences left LSM right SGM 8 CONCLUSION The generation of digital surface models in urban areas based on very high resolution stereo satellites pairs has reached a high level of accuracy The area based least squares matching is not able to generate clear building outlines and is strongly influenced by occlusions Semiglobal
3. 1 w3 pp unpaginated CD ROM Konecny G 1998 Remote Sensing for Sustainable Development State of the Art EARSeL Aavances KrauB T Lehner M Reinartz P 2008 Generation of coarse 3D models of urban areas from high resolution stereo satellite images International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences Vol 37 B1 pp 1091 1098 Kux H Pinho D Souza 2006 High Resolution Satellite Images for Urban Planning International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences vol 36 2 pp 121 124 Mas J 1999 Monitoring land cover changes a comparison of change detection techniques International Journal of Remote Sensing Vol 20 pp 139 152 Otto G amp Chau T 1989 Region growing algorithm for matching of terrain images VC Vol 7 2 pp 83 94 Poli D Remondino F Dolci C 2004 Use of Satellite Imagery for DEM Extraction Landscape Modeling and GIS Applications International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences Vol 36 1 W3 Renouard L Lehmann F 1999 High resolution Digital surface models and orthoimages for telecom network planning Photogrammetric Week 99 pp 241 246 Heidelberg Skidmore A 1997 GIS Applications and use of digital terrain modelling Joint European Conference and Exhibition on Geographical Information Vol 1 pp 442 463 Vienna Thomas N Hendrix
4. results of the analysis of the IKONOS building height are shown in table 6 The RMS of height difference is slightly below 1m corresponding to 0 5 pixels for the x parallax for LSM and in the range of 0 8 m or 0 4 pixels for the x parallax for SGM The building outlines were not tracked perfectly by LSM It was difficult to determine accurate elevations although supported by contour lines The obtained results from SGM are considerably better that obtained by LSM Fig 6 Upper left part of GeoEye 1 image Upper right grey value coded reference DSM lower left grey value coded GeoEye 1 DSM Lower right grey value coded DSM differences Riyadh height base 1 7 checked with 75 buildings LSM SGM Standard deviation of reference heights 0 2m 0 2m RMS of IKONOS height differences 0 9m 0 8m SZ of IKONOS height differences 0 8m 0 7 m RMS of x parallax 0 5 GSD 0 5 GSD Table 6 Accuracy of IKONOS building heights The frequency distribution of the height differences is shown in figure 7 with height intervals of 0 25 m together with the Gaussian distribution red lines The discrepancies between the real and normal distribution can be explained by the limited number of differences Only the results based on LSM shows a small bias of 0 20 m Histogram of AH LSM Reference Histogram of AH SGM reference 10 10 _ __ ___ _ nr RMS Z 0 9 RMS Z 0 8 8 Mean 0 20 8 Mean 0 2 A 3 5 6 S 6
5. C Congalton R G 2003 A comparison of urban mapping methods using high resolution digital imagery Photogrammetric Engineering amp Remote Sensing Vol 69 9 pp 963 972 Welch R 1990 3D terrain modeling for GIS applications GIS World 3 5 pp 26 30 Zhang L Feiner J Louy O Gruen A Schmid W 2002 Photo textured Digital Terrain Models as a Basis for Regional and Local Planning International Workshop on Visualization and Animation of Landscape IAPRS Vol 44 5 w3 Kunming China the 35 Plenary and Working Groups Meeting of ISO TC 211 Jeddah Kingdom of Saudi Arabia 12 S CD
6. Pixel based matching of satellite stereo pairs for build up areas Abdalla Alobeid and Muhamad Alrajhi 1 Leibniz University Hannover Institute of Photogrammetry and Geoinformation alobeid ipi uni hannover de Assistant Deputy Minister for Land amp Surveying Ministry of Municipal and Rural Affairs MoMRA Riyadh KSA mnalrajhi yahoo com Abstract The traditional area based matching is handling the object plane in the area covered by the matching window as a plane In areas with sudden height changes this does not correspond to the reality causing a smoothening of the objects which can be seen especially at buildings Semi Global Matching SGM belongs to the pixel based matching SGM is able to handle sudden height changes as appearing at buildings in the correct manner With images taken by IKONOS and GeoEye 1 3D city models have been generated by area based matching with Least Squares Matching LSM and SGM As expected LSM is flattening the objects while SGM delivers the best result Especially the building shape is more precise based on SGM and SGM has not as much problems with areas of low contrast and shadows Nevertheless with DSMs of both methods differential height models could be generated showing very clear height changes of buildings caused by new eliminated and enlarged buildings Height changes of one floor in the range of 2 7m became obvious Ground control points GCPs for the bias corrected RPC orientation o
7. Remote Sensing Symposium IGARSS 01 IEEE Vol 3 pp 979 981 Falkner E 1995 Aerial Mapping Methods and Applications 7st Ed London pp 1 7 Lewis Publisgers Flamanc D Maillet G 2005 Evaluation of 3D city model production from Pleiades HR satellite images and 2D ground maps International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences Vol 36 8 w27 Tempe AZ USA Fraser C S Baltsavias E Gruen A 2001 3D building reconstruction from high resolution Ikonos stereo imagery In Automatic Extraction of Man Made Objects from Aerial and Space Images A A Balkema Lisse Grodecki J 2001 Ikonos stereo feature extraction RPC approach American Society for Photogrammetry and Remote Sensing Conference St Louis p p 7 on CD ROM Haala N 2009 Comeback of digital image matching in Fritsch ed Photogrammetric Week pp 289 301 Heidelberg Heipke C 1996 Overview of image matching techniques 76th OEEPE Workshop on Application of Digital Photogrammetric Workstation Lausanne Hirschmuller H 2008 Stereo Processing by Semiglobal Matching and Mutual Information IEEE Transactions on Pattern Analysis and Machine Intelligence Vol 30 2 pp 328 341 Jacobsen K B y ksalih G Topan H 2005 Geometric Models for the Orientation of High Resolution Optical Satellite Sensors International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences Hannover 36
8. blurry areas that correspond to shadow areas The buildings close to each other are merged and appear as building blocks where the hidden parts are not matched The border between the houses cannot be clearly distinguished In addition the building outlines are smoothened Summarizing it can be said that LSM generated visually acceptable results However DSMs are blurry and contain large interpolated areas LSM still does not yield clear building shapes The building footprints can be extracted from the generated DSM but not with satisfying detail LSM failed in extracting details of small and very small houses The used LSM program may require several seed points The main source of mismatching was missing texture of some buildings Some errors in the generated DSM correspond to moving objects on the streets and low contrast areas Fig 3 Left hand side GeoEye 1 image grey value coded DSM generated by LSM algorithm on right hand side lighter gray levels correspond to higher elevations 6 Analysis of DSM generated by SGM SGM has been proposed as an alternative solution to overcome drawbacks in LSM SGM requires epipolar images The algorithm was tested with the same high resolution satellite imagery datasets as described before tables 1 and 2 A program written in Visual C was designed for automatic image matching by SGM The parameters of SGM have been determined empirically based on visual inspection of the generated DSMs The param
9. ed in figure 2 upper right corresponding to larger matching gaps appearing where no contrast is available Buildings close to each other are merged and appear as building blocks where hidden parts are not matched On the other hand due to the use of a fixed template size the street level building facades and roof tops are partially merged and LSM leads to an averaged height So the building outlines are smoothened Few or no points are extracted in shadow areas close to the base of buildings The matching in these areas may not be satisfying to model discontinuities such as caused by buildings Fig 2 Left hand side IKONOS image grey value coded DSM generated by LSM algorithm on right hand side lighter gray levels correspond to higher elevations Matching with GeoEye 1 images was also done for the Riyadh test site Of course the GeoEye 1 image with 0 50m GSD includes more details as the IKONOS image The spatial resolution is better because of the view direction the shadow areas are disturbing more as in the used IKONOS stereo pair With the higher resolution of GeoEye 1 of course the result is better as with IKONOS see figure 3 The building shapes are clearer than those generated from the IKONOS stereo pair As expected with the area based LSM the building shape cannot be sharp as in the original form Also some incorrect information on top of the buildings and close to facades remain unresolved The generated DSM has some interpolated
10. eters in automated DSM generation by SGM algorithm are number of accumulated path and penalty for changes in disparity of one pixel P1 and penalty for changes in disparity more than one pixel P2 The number of paths should be at least 8 or16 for providing a good coverage of the 2D image Sixteen paths have lead to better results than eight paths The investigations proved that the best results were achieved for ratio P2 P1 of approximately two Up to 10 of all pixels changed their disparity values in the test area when P1 and P2 were increased above the used values The building shapes were not affected too much by such an increase Most of the changed values were on building outlines and lower buildings The used values of P1 and P2 in the test area are P1 6 and P2 12 Fig 4 Left hand side IKONOS image grey value coded DSM generated by SGM on right hand side lighter gray levels correspond to higher elevations The mentioned IKONOS stereo pair over Riyadh city has been used to generate a DSM by SGM with results as demonstrated in figure 4 Although the images have low contrast a visual inspection shows that SGM is able to generate a DSM with more and clearer details as LSM Most buildings are well represented Building outlines seem to be sharp as in the original image but some inaccurate points appeared especially around building edges due to shadows However trees around buildings may affect the extraction of the correct building sha
11. f the satellite images have been determined in aerial images Based on the GCPs a shift of the models against the direct sensor orientation for IKONOS in X and Y in the range of 7m and in Z of 0 3m and for GeoEye in X and Y below 1m and in Z 2 8m were determined Key Words satellite data digital surface model urban area local and global matching empirical comparison 1 Introduction The extraction of the third dimension from stereoscopic image pairs is a well known technique Photogrammetry is one of the oldest methods which has been used for 3D information generation with developments already beginning around 1840 Falkner 1995 One of the first application of satellite technologies was the military reconnaissance Since 1962 stereoscopic coverage has been available Konecny 1998 but for civilian application use of satellite stereo pairs started with SPOT in 1986 The ground resolution and imaging at different instants limited the photogrammetric space application In 1999 when the IKONOS satellite was launched in order to provide very high resolution stereo imagery a new era was opened with a competition to aerial images Based on such images high resolution DSMs can be generated DSMs got an increasing interest in recent years because of their significant role for many applications e g telecommunication Renouard and Lehmann 1999 urban planning Allam M M 1978 Zhang et al 2002 Thomas et al 2003 Kux et al 2006 map u
12. ght to Angle of Sun Image Description date base h b convergence Elevation quality buildings close iow together 2008 05 24 1 7 31 7 74 7 Aantrasi separate buildings Table 2 Basic characteristics of IKONOS stereo pairs used for DSM generation 3 Reference Data In order to carry out a qualitative and statistical analysis of DSMs generated by the mentioned matching algorithms accurate reference data have to be available A DSM was generated by using aerial photos taken in May 2007 by an analogue aerial camera RC30 The aerial images were available at a scale 1 5500 focal length 303 2 mm scanned at 14 um corresponding to a ground sampling distance GSD of 7 cm The stereo pairs have 60 forward overlap and 20 side lap The image quality of the scanned grainy analogue aerial images was improved by an average Filter 4x4 filter window reducing the image noise The DSMs are generated by LSM with image orientation given from a bundle block adjustment LSM was used because of processing time and the very high resolution of 7 cm The bundle block adjustment is with 2 7 um Oo and mean square discrepancies at GCPs of 1 0 cm for X 1 1 cm for Y and 4 4 cm in Z The DSM has been used as reference to compare the geometric accuracy of DSMs generated by the used matching algorithms from IKONOS and GeoEye 1 images in the Riyadh test area 4 Image Orientation The relation between ground coordinates and its corresponding position in the i
13. image quality allowed a threshold of 0 8 what is high in relation to other areas Table 4 shows the resulting matching success Testarea Riyadh IKONOS Riyadh GeoEye 1 Percentage of accepted points 79 78 Table 4 matching success 100000 350000 2 80000 lt 300000 f 5 5 250000 amp 60000 200000 hi g 40000 4 150000 8 20000 g 100000 FAE 8 9 50000 ty ga 0 S eae 0 95 0 90 0 80 0 70 0 60 0 50 0 40 0 30 0 200 10 0 00 0 95 0 90 0 80 0 70 0 60 0 50 0 40 0 30 0 20 0 10 0 00 correlation coefficient correlation coefficient Threshold of the correlation coefficient 0 8 Threshold of the correlation coefficient 0 8 Riyadh test area IKONOS Riyadh test area GeoEye 1 Fig 1 frequency distribution of correlation coefficients by LSM with used threshold The IKONOS stereo pair over Riyadh city was used to generate a DSM by LSM The images have limited contrast and a satisfying height to base ratio of 1 75 These characteristics make the matching algorithm particularly interesting for the generation of urban DSMs especially in an urban area with high sun brightness As demonstrated with figure 2 the LMS matching results based on IKONOS images are not very good Especially building outlines are fuzzy and deformed into blobs The DSM from LSM makes a determination of satisfying building shapes difficult More problems occur in occluded areas and on homogenous roof tops and streets Black areas as highlight
14. mage has be calculated by using the replacement model of Rational Polynomial Coefficients RPC details can be found in Grodecki 2001 Jacobsen et al 2005 The absolute geo reference for the Riyadh satellite stereo models are GCPs extracted from aerial images By bias based RPC determination the shifts listed in table 3 of the direct sensor orientation were determined The root mean square differences of the GCPs are slightly above 1 GSD mainly caused by limited identification of the GCPs Tad amb l l RMSX RMSY est area ae Be X shift Y shift Z shift TT 1 26m 0 95m IKONOS 3 an ee ae ae 0 61 0 75 Riyadh 22 1 04m 0 7 m 2 5 m i j GeoEye 1 Table 3 results of the bias corrected RPC orientation 5 Analysis of DSM generated by LSM The automatic extraction of DSMs in urban areas has been carried out with high resolution satellite imagery such as IKONOS and GeoEye 1 described in tables 1 and 2 A good choice of the parameters is essential for the success of the algorithm The parameters in automated DSM generation by LSM algorithm are seed points window size matching step distance and threshold for correlation coefficient The image positions of seed points and some check points for the image matching have been measured manually using the Hannover program DPLX Then Hannover program DPCOR embedded in DPLX has been used for automatic image matching The used program for least squares image matchi
15. matching requires epipolar images It leads to improved DSMs compared to LSM In general it is possible to generate DSMs by automatic image matching in urban areas with a quality acceptable for a number of applications using very high resolution satellite images REFERNCES Allam M 1978 DTM s Application in Topographic mapping Proc of the DTM Symposium pp 1 15 Louis Alobeid A 2011 Assessment of Matching Algorithms for Urban DSM Generation from Very High Resolution Satellite Stereo Images In Wissenschaftliche Arbeiten der Fachrichtung Geod sie und Geoinformatik der Leibniz Universitat Hannover ISSN 0174 1454 Nr 293 Dissertation Hannover 2011 Alobeid A Jacobsen K Heipke C 2010 Comparison of Matching Algorithms for DSM Generation in Urban Areas from Ikonos Imagery Photogrammetric Engineering amp Remote Sensing Vol 76 9 pp 1041 1050 Baltsavias E P Stallmann D 1992 From Satellite Images to GIS with Digital Photogrammetry Using SPOT Data In Proc of EGIS 92 Vol 2 pp 945 946 Brito J L N S Silveira M T Jacobsen K Amorim S Mota G L A Feitosa R Q Heipke C 2008 Monitoring of Height Changes in Urban Area from Multi Platform Remotely sensed Data The International Archives of the Photogrammetry Remote Sensing and Spatial Information Sciences vol 37 B1 pp 835 840 Beijing Caetano M Santos T 2001 Updating land cover maps with satellite images Geoscience and
16. ng uses region growing The core of this program was developed by Heipke and Wegmann 2002 see also Heipke 1996 Matching in image space is independent from any orientation information The number of required seed points depends on image similarity and decreases with a larger height to base h b ratio corresponding to a smaller angle of convergence We found empirically that the optimal size of the matching window in the investigated areas was 10 by 10 pixels as usual In particular for larger window sizes adjacent buildings may be merged and appear as a common blob A smaller template may circumvent this problem to a certain extent and thus yield more correct results near height discontinuities but generally leads to a lower accuracy because of the reduced redundancy of the adjustment The step width of region growing should depend upon the roughness of the terrain In images of urban areas there is no significant difference in identifying the borders of the building when the matching step distance was changed from 1x1 pixel to 3x3 pixels Alobeid 2011 In order not to lose any information a point interval of 1 pixel was used in all investigations The threshold for the correlation coefficient plays an important role for the success of this algorithm It can easily be determined empirically The threshold of the correlation coefficient used for LSM was selected based on the histogram of the obtained correlation coefficients figure 1 The
17. pdate and monitoring land cover changes Mas 1999 Caetano and Santos 2001 Brito et al 2008 DSMs also have become an information source for generation of high resolution urban models Fraser et al 2001 Flamanc et al 2005 Krau et al 2008 In addition DSM can be used as information layers in Geographic Information System GIS Welch R 1990 Baltsavias and Stallmann 1992 Skidmore 1997 Poli et al 2004 The applications of DSMs led to intense research with the aim to generate DSMs from very high resolution stereo satellite images HRSI for large areas automatically In this paper our investigation has been initiated for two primary reasons The first reason relates to the use of high resolution satellite images for DSM generation instead of other sources such as aerial images or laser scanner data which in many parts of the world are unavailable expensive or classified Stereo pairs from very high resolution satellites such as IKONOS and GeoEye 1 led the way into a new not restricted era of generating DSMs The second reason is a recent revival of image matching see e g Haala 2009 and thus a renewed interest in the capabilities of recently developed algorithms with respect to traditional solutions An automatic procedure for the generation of DSMs including building shapes based on image matching techniques is highly desirable In this paper we present an analysis and a comparison of two different matching methods fo
18. pe The borders between houses can clearly be distinguished The DSM shows details similar to the original images where the details of building roofs are clearly extracted see figure 4 lower right Although the IKONOS stereo pair was taken with high sun elevation the generated DSM is not affected because SGM is based on mutual information MI Matching cost and MI has been shown to be rather robust with respect to radiometric variations A visual inspection shows that some errors correspond to moving objects on the streets The GeoEye 1 image pair of the Riyadh test site also has been used to generate a DSM by SGM it is shown in figure 5 Also with these images DSMs based on SGM are clearly better as based on LSM A visual inspection shows that most buildings are very well represented and building outlines seem to be sharp as in the original images Trees around the buildings affect the extraction of the correct shape of some buildings However spaces between buildings could clearly be distinguished Black areas appear where no contrast exists or moving objects disturbed the matching Vertical walls are visible very well SGM shows more details on building roofs and around buildings such as parked cars Fig 5 Left hand side GeoEye 1 image grey value coded DSM generated by SGM on right hand side lighter gray levels correspond to higher elevations Summarizing the results it can be said that SGM achieves visually very good results for mos
19. r generating urban DSMs based on very high resolution satellite images least squares matching LSM F rstner 1982 in a region growing fashion Otto Chau 1989 Heipke et al 1996 a local area based method which compares the intensity values within a template to those in a search window semiglobal matching SGM Hirschmiller 2008 which computes conjugate points along multiple conjugate lines hierarchically by using mutual information instead of intensity value differences as dissimilarity measure The mentioned matching algorithms for automated DSM generation from high resolution satellite stereo image pairs are described in Alobeid et al 2010 2 Used data and project areas IKONOS and GeoEye 1 stereo pairs of urban area was used to generate DSMs in Riyadh Saudi Arabia the project areas are located in rolling terrain including densely built up areas Panchromatic IKONOS GEO stereo models with a ground sampling distance GSD of 1m was available also a GeoEye 1 pair with 0 5m GSD Some technical specification of the used GeoEye 1 and IKONOS stereo pairs are shown in table 1 and 2 General Acquisition Height to base Angle of Sun Image Description date h b convergence Elevation quality buildings close together 2009 09 15 1 5 35 9 62 9 good separate building Table 1 Basic properties of Geo Eye 1 stereo pairs used for DSM generation General Acquisition Hei
20. t building shapes The main drawback to the algorithm is caused by buildings without contrast leading to limited results Generally it can clearly be seen that thanks to the pixel based matching and the combination of several 1D paths in the different directions the SGM algorithm is able to produce an improved DSM compared to LSM SGM is able to match complex roof shapes in some situations where LSM fails and shows positions with sudden height changes in the correct manner 7 Qualitative and Quantitative analysis A qualitative and visual evaluation assessment was made by comparing different aspects based by visual inspection It is possible to draw the following summary of the qualitative analysis listed in table 5 Object Buildings outlines LSM Smooth SGM Sharp Flat roofs Smoothened shape Correct shape Gable and hip roofs Acceptable shape Good shape Facades Not vertical Nearly vertical Base of buildings Wider as correct size Similar to correct size Details on roof Not detected Most details detected Poor texture Not matched Acceptable match interpolation Densely located buildings Merged and appear as building blocks Merged and appear as building blocks Borders between usual buildings Often not distinguished Clearly distinguished Bridge over highway Deformed Clearly extracted Bare area Clearly extracted Clearly extracted
Download Pdf Manuals
Related Search