The growing amount of satellite data, increasingly resolved spatially and temporally, represents a high potential of information allowing the accurate characterization of the evolution of an area of interest. For this reason, automatic analysis techniques such as change detection methods are widely investigated. Most of them are based on radiometric changes between remote sensed optical images. These methods are however very sensitive to a significant number of irrelevant changes such as those due to the variation of the geometrical conditions between two different acquisitionsThe objective of this work is then to develop an alternative method based on the elevation change detection. The advantage of using the elevation is that this information is particularly relevant and well adapted in a context of urban monitoring where the elements of interest correspond to buildings that can be constructed, modified or destroyed between two dates.In order to satisfy new needs in image analysis which require quick and reliable results, our method is a complete and automatic processing flow based on the analysis of high resolution satellite stereoscopic couples and the generation of Digital Surface Models (DSM). Stereoscopic DSMs, however, generally suffer from a high number of correlation errors leading to false alarms in the final change detection map. One of the main contribution of this work consisted in increasing the DSM accuracy, especially through a better handling of the occlusion and miss-correlation areas. For this purpose, the image matching technique has been improved and all DSMs computed from the same stereoscopic couple are then fusioned through a new approach, based on an optimization method.The comparison between our DSM with a LiDAR-based DSM indicates that our method largely improves the DSM quality, the amount of correlation errors is decreased while the occlusion areas are accurately localized. The change detection method itself is based on the labelization of the pixels of the differential DSM computed from the DSMs generated at each date of interest. This step, performed through another optimization process, enables to bring forward the relevant changes among the residual noise of the DSMs. The results, obtained for several experimental areas, show that more than 80% of the changes larger than 15 pixels x 15 pixels (100 m² with high resolution images) are detected with our method, with less than 20% of false alarms. We also show that these results mainly depend on the regularization parameter which controls the balance between the amount of false alarms towards the amount of true detections in the final results.
