BUILDING DETECTION AND HEIGHT RETRIEVAL ... - Semantic Scholar
BUILDING DETECTION AND HEIGHT RETRIEVAL IN URBAN AREAS IN THE FRAMEWORK OF HIGH RESOLUTION OPTICAL AND SAR DATA FUSION H. Sportouche, F. Tupin
L. Denise
Institut Telecom; Telecom ParisTech TSI Department 46 rue Barrault - 75013 Paris, France [email protected] fl[email protected]
Thales Communications; Land and Joint IMINT Department 1-5 avenue Carnot - 91883 Massy, France [email protected] [email protected]
ABSTRACT In this paper, we propose a symmetrized version of a semiautomatic processing chain, able to provide the simple 3D reconstruction of buildings in urban scenes, from highresolution optical and SAR imagery. The new elaborated chain gives an equivalent part to the optical and SAR components, in order to fully exploit complementary information provided by proper building features in both images. First, the initial processing chain is reminded and completely illustrated on a studied scene on real data. Then, three points of improvements by process symmetrization are discussed: an augmentation of the detection rate in the footprint extraction step, an increase of the reliability attached to the estimated building heights and a joint improvement of the steps of building validation and qualification. Its is shown that the appropriate combination of optical and SAR features, inside some processing steps, could give better results of reconstruction. Index Terms— Building detection, height estimation, 3D reconstruction, optical-SAR data fusion, urban areas 1. INTRODUCTION During the last years, new approaches, exploring the high detail level characterizing high-resolution (HR) optical and SAR images provided by current spaceborne sensors, have been proposed for object detection and reconstruction in urban areas. Especially, challenges are born in the fields of building extraction and building height estimation for 3D reconstruction of urban scenes. Some technics such as stereoscopy, radargrammetry and interferometry, have been applied to deal with this problematic and have given satisfying results. Nevertheless, some difficulties can be still pointed out: optical couple availability, need of auxiliary input data, incomplete or noisy results from SAR images, necessity of the intervention of an operator for manual proceedings.
978-1-4244-9566-5/10/$26.00 ©2010 IEEE
3660
In this context, news methods have been recently developed in the frame of optical-SAR data fusion, that focus, most of the time, independently on different individual subparts of the 3D building reconstruction, such as the step of building detection [1], the step of footprint boundary extraction [2] and the step of height estimation (from one single SAR image [3] or from a couple of interferometric SAR images). In this paper, we propose a new version of a complete and semi-automatic processing chain, able to provide the simple 3D reconstruction of buildings in large urban scenes. The process takes only as input one HR optical image and one HR SAR image of the same area. It tends to fully exploit complementary information provided by proper building features of both images in a data fusion framework. A simple parallelepipedic building model is used for the reconstruction. In a first part, we remind the initial version of the processing chain, previously described in [4] and decomposed into three steps: the building footprint extraction on the monoscopic optical image, the joint height estimation and building validation from the SAR image and the qualification of the reconstructed buildings on SAR data. The intermediate results, obtained at the end of each step, are illustrated on a scene of interest in a couple of Quickbird and TerraSAR-X images, located in the area of Marseille (France). A 3D view of the reconstructed buildings is finally proposed. In a second part, we discuss a new version of the chain, based on a symmetrization process, that gives an equivalent part to the optical component and to the SAR component. This idea is to combine optical and SAR information inside a same step of the chain in order to achieve better performances of building reconstruction. Three aspects are handled: - an augmentation of the detection rate, in the footprint extraction step, by the introduction of complementary SAR features; - an increase of the reliability attached to the retrieved building heights, by the use of a statistical criterion jointly computed on the optical and SAR images; - an improvement of the steps of building validation and
IGARSS 2010
reconstruction qualification, ensured by a robust process of optical-SAR crossed verification and the computation of scores of confidence from both data. 2. INITIAL PROCESSING CHAIN FOR 3D BUILDING RECONSTRUCTION 2.1. Step 1: Building detection on the optical image The first step aims to extract building boundaries from the monoscopic optical image without auxiliary data. The principle is the following: In a first time, a set of rectangular windows of interest is globally identified on the optical image. In a second time, a local refinement process is performed around each window in order to precisely extract building boundaries. 2.1.1. Identification of windows of interest The set of rectangular windows of interest, is identified by combination of morphological and geometrical tools. As described in [4], the Differential Morphological Profile (DMP) of the optical image is first built to provide a simplification of the scene at different scales. Then, a criterion about geometrical adequation is hierarchically tested on the different objects present in the DMP. This provides a map of windows likely to contain buildings. 2.1.2. Building footprint extraction A contour-based approach is processed around each area provided by the previous ”window map”. The optimal rectangular building footprint, generated from the main local direction, is determined by the minimization of an energetic cost, computed on the optical thresholded gradient image. The output, called the ”boundary map”, identifies precise location of potential buildings. 2.1.3. Application on the studied scene and evaluation of the performances The complete building detection process is applied on a studied urban scene (Quickbird optical sensor, resolution of 0.6 meters, area of Marseille in France). Figure 1 presents the ”boundary map”, obtained in the optical image referential. Among the eight potential footprints detected on the studied scene, we can distinguish seven good detections and one false alarm. Concerning good detections, accurate results of location are provided. The false alarm is due to an homogeneous rectangular parcel on the ground. By comparison with ground truth photography, we can enumerate two non detected footprints, that are due to buildings with weak contrast or imperfect boundaries. A detection rate of 78 % has thus been obtained.
3661
Fig. 1. Result of building footprint extraction, superimposed c to the optical scene (DigitalGlobe). The extracted boundaries are represented in red. 2.2. Step 2: Height retrieval and building validation on the SAR image The second step has two objectives, based on the exploitation of SAR data: firstly, strengthening building presence hypothesis to deliver a decision of validation; secondly, providing a height information for each extracted building. The principle is the following: In a first time, a phase of projection and registration beetween homologous building features is done, to combine two kinds of information issued from both images. In a second time, the proper building validation and the height retrieval are simultaneously performed, through the optimization of SAR criteria. 2.2.1. Building footprint projection and registration on the SAR image We are led to project the rectangular optical footprints into the SAR image and to match them with their homologous SAR features, i.e. with the bright linear echos coming from double ground-wall reflexions. The original method of projection and registration refinement, presented in [4], is inserted in the processing chain. It can be decomposed as follows: First, the building boundary projection is applied by using the available physical parametric joint model of projection. Then, the ground feature registration refinement is performed by using a radiometric criterion computed on the SAR image. 2.2.2. Height estimation and validation from SAR data The original method of height estimation and validation, developed in [4], is integrated in the processing chain. The proposed approach is based on the combination of two SAR criteria, depending on the building height. It can be summarized as follows: First, a statistical criterion, defined as the global Log-Likelihood, is optimized according to a scheme of ”Height hypothesis-Partitioning generation-Likelihood maximization”. It provides a set of candidate building heights, that guarantee a good geometric adequation between the predicted partitioning into characteristic building areas (”layover area”, ”single roof area”, ”shadow area”...) and the real building signature. Then, a
radiometric criterion, defined as a local contrast testing the presence of a predicted discontinuity between the ”background area” and the ”layover area”, is used. It permits to select the candidate height running to a best criterion value and to estimate the building height. Finally, a thresholding on the best contrast enables to validate or not the building. The output is a grey level map, called the ”height map”, indicating the relative estimated heights for the validated buildings. The ”final 3D view” of reconstructed buildings is easily achieved by combining planimetric information issued from the ”boundary map” and altimetric information issued from the ”height map”.
Fig. 3. Result of height estimation and building validation. A grey level scale is used to indicate the relative building heights, that have been retrieved for the validated buildings. The rejected buildings are designed by a cross.
2.2.3. Application on the studied scene and evaluation of the performances The steps of registration, building validation and height estimation are performed on the corresponding SAR scene (TerraSAR-X SAR sensor, resolution of 1.1 meters). Figures 2, 3 and 4 present respectively the ”registration map”, the ”height map” and the textured ”final 3D view”, obtained in the SAR image referential. After registration, a fine adjustement of the projected footprint location is obtained in the SAR image. Concerning the building validation, six correct decisions (validation or reject) among eight have been taken; only one non-detection and one false alarm have occured. Concerning the height estimation, the retrieved building heights appear quite reliable with a global root mean square error of 0.85 meters. A globally satisfying 3D reconstruction of the whole scene is provided.
c Fig. 4. Result of 3D building reconstruction (Infoterra). 2.3.2. Application on the studied scene and analysis Scores of confidence have been computed for the validated buildings on the scene. All scores are strictly higher than 0.5 (reaching values between 0.58 and 0.83), except the score of one building that effectively corresponds to the false alarm. The method of qualification appears thus quite relevant. These scores could be used to lead a manual verification for buildings presenting a too low score. 3. CHAIN SYMMETRIZATION FOR PERFORMANCE IMPROVEMENT 3.1. Augmentation of the detection rate
Fig. 2. Result of the footprint projection and registration, suc perimposed to the SAR scene (Infoterra). The extracted boundaries are represented in red.
2.3. Step 3: Qualification of the 3D reconstruction from SAR data 2.3.1. Definition of scores of confidence on SAR data To provide a quality measure of each reconstruted building, a global score of confidence is defined between 0 (weak confidence) and 1 (strong confidence), by merging individual scores. These last ones attest the reliability of the optimal peaks in the evolution curves of both SAR criteria (in function of the height). They take into account different aspects (peak shape, peak location and peak value).
3662
A new scheme for a symmetrized footprint extraction, taking also into account SAR features, can be proposed. The idea is to independently extract a ”window map” from the optical image and an other one from the SAR image and to merge them together into a global one. The SAR ”window map” can be built as follows: Firstly, some characteristic features such as double echos and shadows are extracted on the SAR image (respectively, with a line detector [5], for the bright echos, and, with the segmentation algorithm of active grid [6] followed by a thresholding on the radiometric means in regions, for the shadows); secondly, some couples of associated features (double echos-shadows) are identified by considerations about their relative location and their proximity; thirdly, the thresholded SAR image of linear extracted structures is used to generate SAR rectangular windows through an energetic cost minimization. This may allow to cope with the high building signature
variability, that characterizes the visibility of optical and SAR features (indeed, optical building boundaries or SAR echos are not always well distinguishable on data). Figure 5 presents, on the studied scene, an example of two building boundaries, that have not been detected on the optical image, but that can be extracted with the introduction of such SAR features in the detection step.
Fig. 5. Result of SAR window detection using the proposed symmetrized scheme, superimposed to the SAR scene c (Infoterra). Zoom on two buildings of interest. The new window boundaries are represented in blue.
3.2. Increase of the reliability of the estimated building heights The principle is to combine the presented SAR criteria with new optical criteria, depending on the building height. For instance, the Log-Likelihood criterion, computed on a test partitioning of the building SAR signature, for a given 3D building model and a supposed building height, could be similarly calculated on the optical image. A combined LogLikelihood criterion, defined as the weighted sum of the initial SAR Log-Likelihood and the new optical one, could thus provide more reliable candidate building heights. In this way, when the characteristic SAR areas would appear as not clearly imaged on SAR data, a strong weight would be given to the optical component and vice versa. 3.3. Improvement of building validation and reconstruction qualification The building validation could be made more robust by mixing optical and SAR clues accumulated throughout the chain. For instance the false alarm rate could be reduced by a crossedvalidation process, checking the presence of discriminant features on both data (SAR echos, SAR shadows and optical shadows into neighbourhoods and strong optical gradients along potential building boundaries). In the same way, the qualification could be improved by taking into account measures relative to the representativity of optical and SAR optimal criteria. A combined score of confidence could be proposed by fusing the initial SAR score with a new optical score. This last one would be relative to the optimal energetic cost, computed on the optical gradient image, for a considered footprint during the detection step. This should help to provide a more relevant qualification.
3663
4. CONCLUSION AND FURTHER WORKS An initial processing chain, able to operationally produce a satisfying 3D reconstruction of buildings from optical and SAR metric data, has been first reminded. Its performances have been evaluated on a Quickbird and TerraSAR-X scene. A new symmetrized version of the chain has then been proposed. The interest of this elaborated version, that combines in an appropriate way optical and SAR complementary features inside some processing steps, has been discussed. It has been demonstrated that this new chain could give better qualitative and quantitative results for 3D reconstruction: indeed, the building detection would be more complete and the height estimation should be more reliable. In future work, this new version should be applied and evaluated on large urban scenes. Moreover, the complete methodology could be extended to more sophisticated buildings. An other interesting aspect would be to generalize the approach when an additional interferometric height information is available as input. 5. REFERENCES [1] V. Poulain, J. Inglada, M. Spigai, J. Y. Tourneret, and Marthon P., “Fusion of high resolution optical and SAR images with vector data bases for change detection,” in IEEE International Geoscience and Remote Sensing Symposium, 2009. IGARSS’09. Proceedings, July 2009, vol. 4, pp. 956–959. [2] F. Tupin and M. Roux, “Detection of building outlines based on the fusion of SAR and optical features,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 58, no. 1-2, pp. 71–82, 2003. [3] D. Brunner, G. Lemoine, L. Bruzzone, and H. Greidanus, “Building height retrieval from VHR SAR imagery based on an iterative simulation and matching technique,” IEEE Transactions on Geoscience and Remote Sensing, vol. 48, no. 3, pp. 1487–1504, March 2010. [4] H. Sportouche and F. Tupin, “A processing chain for simple 3D reconstruction of buildings in urban scenes from high resolution optical and SAR images,” in 8th European Conference on Synthetic Aperture Radar. EUSAR 2010. Proceedings, June 2010. [5] F. Tupin, H. Maˆıtre, J. F. Mangin, and J. M. Nicolas, “Detection of linear features in SAR images: application to road network extraction,” IEEE Transactions on Geoscience and Remote Sensing, vol. 36, no. 2, pp. 434–453, March 1998. [6] O. Germain and P. R´efr´egier, “Statistical active grid for segmentation refinement,” Pattern Recognition Letters, vol. 22, no. 10, August 2001.
L. Denise
Institut Telecom; Telecom ParisTech TSI Department 46 rue Barrault - 75013 Paris, France [email protected] fl[email protected]
Thales Communications; Land and Joint IMINT Department 1-5 avenue Carnot - 91883 Massy, France [email protected] [email protected]
ABSTRACT In this paper, we propose a symmetrized version of a semiautomatic processing chain, able to provide the simple 3D reconstruction of buildings in urban scenes, from highresolution optical and SAR imagery. The new elaborated chain gives an equivalent part to the optical and SAR components, in order to fully exploit complementary information provided by proper building features in both images. First, the initial processing chain is reminded and completely illustrated on a studied scene on real data. Then, three points of improvements by process symmetrization are discussed: an augmentation of the detection rate in the footprint extraction step, an increase of the reliability attached to the estimated building heights and a joint improvement of the steps of building validation and qualification. Its is shown that the appropriate combination of optical and SAR features, inside some processing steps, could give better results of reconstruction. Index Terms— Building detection, height estimation, 3D reconstruction, optical-SAR data fusion, urban areas 1. INTRODUCTION During the last years, new approaches, exploring the high detail level characterizing high-resolution (HR) optical and SAR images provided by current spaceborne sensors, have been proposed for object detection and reconstruction in urban areas. Especially, challenges are born in the fields of building extraction and building height estimation for 3D reconstruction of urban scenes. Some technics such as stereoscopy, radargrammetry and interferometry, have been applied to deal with this problematic and have given satisfying results. Nevertheless, some difficulties can be still pointed out: optical couple availability, need of auxiliary input data, incomplete or noisy results from SAR images, necessity of the intervention of an operator for manual proceedings.
978-1-4244-9566-5/10/$26.00 ©2010 IEEE
3660
In this context, news methods have been recently developed in the frame of optical-SAR data fusion, that focus, most of the time, independently on different individual subparts of the 3D building reconstruction, such as the step of building detection [1], the step of footprint boundary extraction [2] and the step of height estimation (from one single SAR image [3] or from a couple of interferometric SAR images). In this paper, we propose a new version of a complete and semi-automatic processing chain, able to provide the simple 3D reconstruction of buildings in large urban scenes. The process takes only as input one HR optical image and one HR SAR image of the same area. It tends to fully exploit complementary information provided by proper building features of both images in a data fusion framework. A simple parallelepipedic building model is used for the reconstruction. In a first part, we remind the initial version of the processing chain, previously described in [4] and decomposed into three steps: the building footprint extraction on the monoscopic optical image, the joint height estimation and building validation from the SAR image and the qualification of the reconstructed buildings on SAR data. The intermediate results, obtained at the end of each step, are illustrated on a scene of interest in a couple of Quickbird and TerraSAR-X images, located in the area of Marseille (France). A 3D view of the reconstructed buildings is finally proposed. In a second part, we discuss a new version of the chain, based on a symmetrization process, that gives an equivalent part to the optical component and to the SAR component. This idea is to combine optical and SAR information inside a same step of the chain in order to achieve better performances of building reconstruction. Three aspects are handled: - an augmentation of the detection rate, in the footprint extraction step, by the introduction of complementary SAR features; - an increase of the reliability attached to the retrieved building heights, by the use of a statistical criterion jointly computed on the optical and SAR images; - an improvement of the steps of building validation and
IGARSS 2010
reconstruction qualification, ensured by a robust process of optical-SAR crossed verification and the computation of scores of confidence from both data. 2. INITIAL PROCESSING CHAIN FOR 3D BUILDING RECONSTRUCTION 2.1. Step 1: Building detection on the optical image The first step aims to extract building boundaries from the monoscopic optical image without auxiliary data. The principle is the following: In a first time, a set of rectangular windows of interest is globally identified on the optical image. In a second time, a local refinement process is performed around each window in order to precisely extract building boundaries. 2.1.1. Identification of windows of interest The set of rectangular windows of interest, is identified by combination of morphological and geometrical tools. As described in [4], the Differential Morphological Profile (DMP) of the optical image is first built to provide a simplification of the scene at different scales. Then, a criterion about geometrical adequation is hierarchically tested on the different objects present in the DMP. This provides a map of windows likely to contain buildings. 2.1.2. Building footprint extraction A contour-based approach is processed around each area provided by the previous ”window map”. The optimal rectangular building footprint, generated from the main local direction, is determined by the minimization of an energetic cost, computed on the optical thresholded gradient image. The output, called the ”boundary map”, identifies precise location of potential buildings. 2.1.3. Application on the studied scene and evaluation of the performances The complete building detection process is applied on a studied urban scene (Quickbird optical sensor, resolution of 0.6 meters, area of Marseille in France). Figure 1 presents the ”boundary map”, obtained in the optical image referential. Among the eight potential footprints detected on the studied scene, we can distinguish seven good detections and one false alarm. Concerning good detections, accurate results of location are provided. The false alarm is due to an homogeneous rectangular parcel on the ground. By comparison with ground truth photography, we can enumerate two non detected footprints, that are due to buildings with weak contrast or imperfect boundaries. A detection rate of 78 % has thus been obtained.
3661
Fig. 1. Result of building footprint extraction, superimposed c to the optical scene (DigitalGlobe). The extracted boundaries are represented in red. 2.2. Step 2: Height retrieval and building validation on the SAR image The second step has two objectives, based on the exploitation of SAR data: firstly, strengthening building presence hypothesis to deliver a decision of validation; secondly, providing a height information for each extracted building. The principle is the following: In a first time, a phase of projection and registration beetween homologous building features is done, to combine two kinds of information issued from both images. In a second time, the proper building validation and the height retrieval are simultaneously performed, through the optimization of SAR criteria. 2.2.1. Building footprint projection and registration on the SAR image We are led to project the rectangular optical footprints into the SAR image and to match them with their homologous SAR features, i.e. with the bright linear echos coming from double ground-wall reflexions. The original method of projection and registration refinement, presented in [4], is inserted in the processing chain. It can be decomposed as follows: First, the building boundary projection is applied by using the available physical parametric joint model of projection. Then, the ground feature registration refinement is performed by using a radiometric criterion computed on the SAR image. 2.2.2. Height estimation and validation from SAR data The original method of height estimation and validation, developed in [4], is integrated in the processing chain. The proposed approach is based on the combination of two SAR criteria, depending on the building height. It can be summarized as follows: First, a statistical criterion, defined as the global Log-Likelihood, is optimized according to a scheme of ”Height hypothesis-Partitioning generation-Likelihood maximization”. It provides a set of candidate building heights, that guarantee a good geometric adequation between the predicted partitioning into characteristic building areas (”layover area”, ”single roof area”, ”shadow area”...) and the real building signature. Then, a
radiometric criterion, defined as a local contrast testing the presence of a predicted discontinuity between the ”background area” and the ”layover area”, is used. It permits to select the candidate height running to a best criterion value and to estimate the building height. Finally, a thresholding on the best contrast enables to validate or not the building. The output is a grey level map, called the ”height map”, indicating the relative estimated heights for the validated buildings. The ”final 3D view” of reconstructed buildings is easily achieved by combining planimetric information issued from the ”boundary map” and altimetric information issued from the ”height map”.
Fig. 3. Result of height estimation and building validation. A grey level scale is used to indicate the relative building heights, that have been retrieved for the validated buildings. The rejected buildings are designed by a cross.
2.2.3. Application on the studied scene and evaluation of the performances The steps of registration, building validation and height estimation are performed on the corresponding SAR scene (TerraSAR-X SAR sensor, resolution of 1.1 meters). Figures 2, 3 and 4 present respectively the ”registration map”, the ”height map” and the textured ”final 3D view”, obtained in the SAR image referential. After registration, a fine adjustement of the projected footprint location is obtained in the SAR image. Concerning the building validation, six correct decisions (validation or reject) among eight have been taken; only one non-detection and one false alarm have occured. Concerning the height estimation, the retrieved building heights appear quite reliable with a global root mean square error of 0.85 meters. A globally satisfying 3D reconstruction of the whole scene is provided.
c Fig. 4. Result of 3D building reconstruction (Infoterra). 2.3.2. Application on the studied scene and analysis Scores of confidence have been computed for the validated buildings on the scene. All scores are strictly higher than 0.5 (reaching values between 0.58 and 0.83), except the score of one building that effectively corresponds to the false alarm. The method of qualification appears thus quite relevant. These scores could be used to lead a manual verification for buildings presenting a too low score. 3. CHAIN SYMMETRIZATION FOR PERFORMANCE IMPROVEMENT 3.1. Augmentation of the detection rate
Fig. 2. Result of the footprint projection and registration, suc perimposed to the SAR scene (Infoterra). The extracted boundaries are represented in red.
2.3. Step 3: Qualification of the 3D reconstruction from SAR data 2.3.1. Definition of scores of confidence on SAR data To provide a quality measure of each reconstruted building, a global score of confidence is defined between 0 (weak confidence) and 1 (strong confidence), by merging individual scores. These last ones attest the reliability of the optimal peaks in the evolution curves of both SAR criteria (in function of the height). They take into account different aspects (peak shape, peak location and peak value).
3662
A new scheme for a symmetrized footprint extraction, taking also into account SAR features, can be proposed. The idea is to independently extract a ”window map” from the optical image and an other one from the SAR image and to merge them together into a global one. The SAR ”window map” can be built as follows: Firstly, some characteristic features such as double echos and shadows are extracted on the SAR image (respectively, with a line detector [5], for the bright echos, and, with the segmentation algorithm of active grid [6] followed by a thresholding on the radiometric means in regions, for the shadows); secondly, some couples of associated features (double echos-shadows) are identified by considerations about their relative location and their proximity; thirdly, the thresholded SAR image of linear extracted structures is used to generate SAR rectangular windows through an energetic cost minimization. This may allow to cope with the high building signature
variability, that characterizes the visibility of optical and SAR features (indeed, optical building boundaries or SAR echos are not always well distinguishable on data). Figure 5 presents, on the studied scene, an example of two building boundaries, that have not been detected on the optical image, but that can be extracted with the introduction of such SAR features in the detection step.
Fig. 5. Result of SAR window detection using the proposed symmetrized scheme, superimposed to the SAR scene c (Infoterra). Zoom on two buildings of interest. The new window boundaries are represented in blue.
3.2. Increase of the reliability of the estimated building heights The principle is to combine the presented SAR criteria with new optical criteria, depending on the building height. For instance, the Log-Likelihood criterion, computed on a test partitioning of the building SAR signature, for a given 3D building model and a supposed building height, could be similarly calculated on the optical image. A combined LogLikelihood criterion, defined as the weighted sum of the initial SAR Log-Likelihood and the new optical one, could thus provide more reliable candidate building heights. In this way, when the characteristic SAR areas would appear as not clearly imaged on SAR data, a strong weight would be given to the optical component and vice versa. 3.3. Improvement of building validation and reconstruction qualification The building validation could be made more robust by mixing optical and SAR clues accumulated throughout the chain. For instance the false alarm rate could be reduced by a crossedvalidation process, checking the presence of discriminant features on both data (SAR echos, SAR shadows and optical shadows into neighbourhoods and strong optical gradients along potential building boundaries). In the same way, the qualification could be improved by taking into account measures relative to the representativity of optical and SAR optimal criteria. A combined score of confidence could be proposed by fusing the initial SAR score with a new optical score. This last one would be relative to the optimal energetic cost, computed on the optical gradient image, for a considered footprint during the detection step. This should help to provide a more relevant qualification.
3663
4. CONCLUSION AND FURTHER WORKS An initial processing chain, able to operationally produce a satisfying 3D reconstruction of buildings from optical and SAR metric data, has been first reminded. Its performances have been evaluated on a Quickbird and TerraSAR-X scene. A new symmetrized version of the chain has then been proposed. The interest of this elaborated version, that combines in an appropriate way optical and SAR complementary features inside some processing steps, has been discussed. It has been demonstrated that this new chain could give better qualitative and quantitative results for 3D reconstruction: indeed, the building detection would be more complete and the height estimation should be more reliable. In future work, this new version should be applied and evaluated on large urban scenes. Moreover, the complete methodology could be extended to more sophisticated buildings. An other interesting aspect would be to generalize the approach when an additional interferometric height information is available as input. 5. REFERENCES [1] V. Poulain, J. Inglada, M. Spigai, J. Y. Tourneret, and Marthon P., “Fusion of high resolution optical and SAR images with vector data bases for change detection,” in IEEE International Geoscience and Remote Sensing Symposium, 2009. IGARSS’09. Proceedings, July 2009, vol. 4, pp. 956–959. [2] F. Tupin and M. Roux, “Detection of building outlines based on the fusion of SAR and optical features,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 58, no. 1-2, pp. 71–82, 2003. [3] D. Brunner, G. Lemoine, L. Bruzzone, and H. Greidanus, “Building height retrieval from VHR SAR imagery based on an iterative simulation and matching technique,” IEEE Transactions on Geoscience and Remote Sensing, vol. 48, no. 3, pp. 1487–1504, March 2010. [4] H. Sportouche and F. Tupin, “A processing chain for simple 3D reconstruction of buildings in urban scenes from high resolution optical and SAR images,” in 8th European Conference on Synthetic Aperture Radar. EUSAR 2010. Proceedings, June 2010. [5] F. Tupin, H. Maˆıtre, J. F. Mangin, and J. M. Nicolas, “Detection of linear features in SAR images: application to road network extraction,” IEEE Transactions on Geoscience and Remote Sensing, vol. 36, no. 2, pp. 434–453, March 1998. [6] O. Germain and P. R´efr´egier, “Statistical active grid for segmentation refinement,” Pattern Recognition Letters, vol. 22, no. 10, August 2001.
