Unsupervised change detection from multitemporal SAR images based on a detail preserving approach and a robust threshold estimation

In environment monitoring and disaster management, Synthetic aperture radars (SARs) have shown their great efficiency due to the fact that they provide short revisit times and they can operate in day and/or night and their independence of weather conditions. Different applications have been addressed a lot in recent years, but the one that receives a lot of attention is the change detection of the observed earth surface by exploiting the multitemporal SAR images. In this paper, we propose an unsupervised method for the change detection from multitemporal SAR images that does not require any speckle filtering. This method is based on: i) generating a multiresolution set of the single-channel log ratio image using stationary wavelet transform (SWT); ii) applying the T-point algorithm for all the images of the multiresolution sets; and iii) fusing the obtained images at the optimum reliable scale to generate the change map. The proposed method was experimentally validated using semisimulated and real SAR images acquired by RADARSAT-2 satellite in the region of Algiers.