|TSM/TDM Science Team Meeting 2016|
|Science Team Meetings|
|How to Submit a Proposal|
|COFUR Price List (scientific use)|
|TanDEM-X Science Service System|
|Title||Urban mapping for risk management|
|Investigator||Spigai, Marc - Alcatel Alenia Space, OS/PS/I|
|Summary||The objective of the project is to evaluate the potential of high resolution X-band TERRASAR data
for mapping and change detection in urban areas.
In order to be able to monitor urban areas, high resolution data are required. By using one or more high resolution images on selected urban areas, a general classification will be processed, which allows to obtain a first analysis of the scene : buildings, main roads, etc. Then a finer extraction of the road network will be realized and change detection will be processed, with a particular goal concerning risk management (flooded area, traficability, etc.).
Classification of the scene and fine extraction of road network will be based on methods studied at Alcatel Alenia Space which use a supervised Bayesian classifier with Fisher distribution on amplitude (estimation with log-cumulants). Fisher distribution domain allows to enlarge the Gamma and Inverse Gamma domains to a more general domain including high resolution radar phenomena. Fine extraction of road network will be based on tree-search algorithm constrained and initialized by the classification of the scene.
Change detection will be based on both methods studied at CNES and Alcatel Alenia Space which use classical methods based on specific radar correlation tools but also with more recent methods based on primitive/features estimation coupled with similarity measurement (based on a set of distance such as Kullback distance).
The project duration is two years (after the commissioning phase). The amount of Alcatel Alenia Space funding for the project will be 10 man-months for the duration of the project
Data required for the project are 24 images which will cover three different French cities:Toulouse, Cannes and Aix-en-provence for which Alcatel Alenia Space has ground truth (digital maps, very high resolution optical images and 3D databases). The 24 images are composed of : 16 High resolution Spotlight SSC images single polarization at two different incidences, 4 High resolution Spotlight SSC images quad polarization at the same incidence and 4 Stripmap SSC images at single polarization at two different incidences.
Deliverables of the project are the following :
If any problem appear during the process of reading and first short analysis, a short report will be send one month after reception of the last image of the first group of data.
End of first year of the project : First report on the preliminary results obtained during the first year of the project.
End of second year of the project : Final report of the project. Publication and participation to scientific meeting(s). In particular, Alcatel Alenia Space participates since the beginning (year 2000) to the CNES-DLR annual workshop "Information Extraction and Scene Understanding for Meter Resolution Images". This meeting could give the opportunity to Alcatel Alenia Space to present the state of the project.
|Detailed report||Main results are given hereafter. A set of figures will be given in one of the next chapters of the final report. REFERENCES [REF 1]: Progress Report project MTH0161 : AO TerraSar-X Urban Mapping for Risk management Project. TAS-09-OS/PG/I-7. 2 March 2009. [REF 2]: Proposal MTH0161 “Urban mapping for risk management” [Amb1] Virginie Amberg, Marc Spigai, Philippe Marthon “Structure extraction from high resolution SAR data on urban areas Application to road extraction“. IGARSS, Alaska, 2004. [Chri1] E. Christophe and J. Inglada, “Robust road extraction for high resolution satellite images,” in Proc. IEEE Int. Conf. Image Process. (ICIP), San Antonio, Texas, USA, Sep. 2007, vol. 5, pp. 437–440. [Hab1] T. Habib, J. Inglada, G. Mercier, J. Chanussot. Speeding up Support Vector Machine (SVM) image classification by a kernel series expansion. International Conference on Image Processing, 2008, ICIP 2008 IEEE International. [Nic1] J.M. Nicolas. A fisher-map filter for sar image processing. IEEE Proc. IGARSS’02, International Conference on Geoscience and Remote Sensing, Toronto, Canada, 4, 24-28 June 2002. [Nic2] J.M. Nicolas. Introduction aux statistiques de deuxième espèce : applications aux lois d’images rso. Rapport Technique ENST2002D001, GET-Télécom Paris, Février 2002. [Pou1] Vincent Poulain, Jordi Inglada, Marc Spigai, Jean-Yves Tourneret and Philippe Marthon, « High resolution optical and SAR image fusion for road database updating », IEEE Geoscience and Remote Sensing Symposium (IGARSS10), July 2010, Honolulu, Hawaii, USA. [Pou2] V. Poulain, J. Inglada, M. Spigai, J. Y. Tourneret, and P. Marthon, “Fusion of high resolution optical and SAR images with vector data bases for change detection,” in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Cape Town, South Africa, Jul. 2009. [Sim1] E. Simonetto, H. Oriot, R. Garello, Extraction of industrial structures and DEM from airborne SAR images, PSIP'01, Marseille, january 2001. [Tup1] F. Tupin, B. Houshmand, M. Datcu: Road detection in dense urban areas using SAR Imagery and the usefulness of multiple views. IEEE Trans. GRSS, Vol.40, N°11, November 2002. [Wied1] C. Wiedemann, “External evaluation of road networks,” Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., vol. 34, pp. 93–98, 2003. TAS SOFTWARE TOOL The method that was used in Thales previous work ([Amb1]) was described in [REF 1]. As a remind figures 1 to 7 are given in the last part of the report which illustrates the different phase of processing : Classification of the scene and fine extraction of road network. RESULTS 1 : Road network extraction from a single image We have split the problem in two areas of the Toulouse city, illustrated on Figure 8 and Figure 9 : ->Area 1 which is a peri-urban area with “big” roads which represent, a priori, the kind of area where we would like to detect road network with TERRASAR-X. Moreover, there are in this area : motorway, entrance ramp, street in a dense urban area, and bridle path of a racecourse. ->Area 2 which is more a dense urban area for which it is, a priori, very difficult for one single SAR image to extract the road network. There are in this area : a motorway perpendicular to the azimuth, Streets lined with houses. Area 1 : Peri-urban The use of the processing chain is illustrated hereafter. The classification (Figure 10) shows the mix of road in “specular” objects present in the scene. Figure 11 shows the segments of initialization obtained after Geometrical filtering and Hough transform Figure 12 and Figure 13 show the road network finally extracted. Road network is represented in blue line. Red bullets represent pixels selected in the first time to extend a road segment but in the second time they are not retained by the cost function It is interesting to see that the main big roads are quite well extracted with TERRARSAR-X images, smaller roads are though detected even if the detection rate is of medium quality. The quantitative results give mainly the typical following value of 45% of correctness and 30% of completeness. So these values are not sufficient by themselves but can give important information useful for a human operator of for a multi-sensors fusion system. Area 2 : Urban dense Figure 12 shows the road network extracted in the urban area 2. Road network is represented in blue line. The extraction of road in a dense urban area is quite difficult to obtain with 1-meter resolution TERRARSAR-X images. The motorway is detected by the road extraction algorithm, but only part of the main streets are retrieved. It is due to the signal response of an urban area, which is very complex. The specular effect of road is often merged with shadow or masked by a layover effect. In summary the detection rate is quite low and it is obvious that the extracted information is relatively poor in this kind of area. A priori information, such as an existing database, should be necessary to help the extraction of the road network, this needs another study and is out of the scope of the project. RESULTS 2 : Road network extraction from multi-temporal images We leave here the rapid mapping to enter the domain of cartography update with 3 images taken in a one-year period. We do not intend here to deal with the problem of change detection but to give a feeling of what kind of improvement can bring a multi-temporal approach. To reduce the speckle noise, a multi-temporal filter is applied (a simple mean of the images), the three input images having the same orbit direction (ascending). This configuration aims at highlighting specular effect observed on road areas, the probability to detect a road would be better. Element on a multi-temporal approach with an ascending and a descending orbit are shown is a second step. ->Use of 3 multi-temporal ascending images The results are presented in Figure 15. Detection is better than the ones obtained with the single image configuration (completeness grows from 30% with a single image until 40% with 3 images). The improvement is mainly observed for instance on the left of the area, in the upper part the roundabout and the across roads are clearly detected and on the lower part the main road around an industrial zone is well detected. -> Use of 2 multi-temporal ascending and descending images To illustrate another configuration of multi-temporal images, the algorithm has been performed with an image in descending orbit and another image in ascending orbit. The objective is to highlight the impact of acquisition configuration on the road network extraction. The two phenomena, layover and shadow, are particularly depicted hereafter. On Figure 16, trees along the motorway generate a layover effect on the left side in the ascending orbit image (a) and on the right side in the descending orbit image (b). (c) represents the superposition of the two image, red elements are due to high level radiometry in the descending orbit image and cyan elements high level radiometry in ascending orbit image. The layover produced by the trees along the motorway (in top of the Figure 17) disturbs the road network extraction algorithm and the localization of the extracted road differs according to the direction of the orbit during the acquisition (Figure 17). The second geometric effect which can also disturb the extraction of roads is shadows of high buildings. Though most of shadows are removed with the geometric filtering (§ 188.8.131.52.), some of them are assimilated as a road. It is the case in the Figure 19 where shadow effects in the pink circle are detected as a road. In the Figure 19. It is worst noting that shadow localization is dependant of the orbit direction. We know apply a multi-temporal filter from the ascending orbit image and the descending orbit image in order to remove shadow effects. For this configuration, the ascending orbit image and the descending orbit have been used as inputs of the multi-temporal filter. The multi-temporal filter using two images acquired with opposed orbit allows removing the shadows of building. But the layover effects (Figure 16) add noise on the both side of the motorway in the top of the Figure 20. As illustrated in the Figure 21. Detection are globally not better, even worse, than the one obtained with the single image configuration EXTENSION TO OPTIC/SAR FUSION As additional work, we give here some elements of a multi-sensors optic/SAR approach, typically for rapid mapping. The work done by TAS is strongly linked to the scientific approach done in the V. Poulain co-funded Phd CNES/TAS “High resolution image analysis with exogenous data” with IRIT French Laboratory as scientific partner. The general principle is explained hereafter, see [Pou1]&[Pou2] for details. The Phd technical work deals with the problem of cartographic databases creation/update in urban environment with high resolution optical images (in the range 70cm up to 2,5m) and radar satellite images (1m). In our study, we focus on the creation of the road network database, that means to build the road network without any prior information coming from an existing database. Moreover, due to its interest highlighted during the PhD work, we focus on the following scenario : extraction of the road network with an optical multispectral image at 2,5 meters resolution (Pleiades simulated image, copyright CNES) and a 1 meter TERRASAR-X image (among those available for our study). The principle is to construct road “candidates” in both optical and SAR image. Then a set of features is computed for each road candidate and a fusion of the feature is then done allowing to confirm or infirm the road candidates. For the optical image the road candidates are build by using the method described in [Chri1], which is also available in the CNES ORFEO Toolbox free software http://www.orfeo-toolbox.org. In the SAR image, the method used is the one described in the previous chapter of the study. For each candidate, a set of feature is then processed, such as : ratio of intensity, vegetation indexes, and so on (see [Pou1]). The feature fusion gathers all the available knowledge to obtain a score for each object. Based on this score, a decision can be taken to include or not a road section in the database. The Dempster-Shafer evidence theory has interesting properties for data fusion (see [Pou2]). It can be considered as a generalization of Bayesian theory allowing one to handle imprecise information. However, it requires less prior information than Bayesian theory, facilitating the integration of new features in the chain. Figure 22 Illustrated the extraction of roads in the optical image (a), in the SAR image projected in the optical image for comparison (b) and finally the result of the fusion. The quantitative values are : Optical image : Correctness = 60%, Completeness = 37% SAR image : Correctness = 45%, Completeness = 30% (as previously obtained) Optical + SAR : Correctness = 60%, Completeness = 53% So the use of multi-sensors Optical+SAR image slightly enhances the completeness of the detection. CONCLUSION For rapid mapping with a single SSC image, the results (Correctness = 45%, Completeness = 30%) show that TERRASAR-X image cannot solve the problem of road extraction by itself but it can bring relevant information in peri-urban situations where “big” roads are present in the scene. Enhancement with a multi-temporal approach where images in the same orbit (ascending or descending) are recommended to be used. For dense urban area, a priori information, such as an existing database, should be necessary to help the extraction of the road network, this needs another study and is out of the scope of the project. Additional work on multi-sensor approach optical (2,5m)/SAR (1m) can bring enhancement (correctness = 60%, Completeness = 53%).|
Back to list of proposals
|© DLR 2004-2016||