Destroyed-buildings detection from VHR SAR images using deep features

Quick identification of post-earthquake destroyed buildings is critical for disaster management. It can be performed in unsupervised way by comparing pre-disaster and post-disaster Very-High-Resolution (VHR) SAR images. Spatial context needs to be modeled for effective change detection (CD) in VHR SAR images as they are complex and characterized by spatial correlation among pixels. We propose a unsupervised context-sensitive method for CD in multi-temporal VHR SAR images using pre-trained Convolutional-Neural-Network (CNN) based feature extraction. The sub-optimal CNN, pre-trained on an aerial optical image dataset and further optimized for using on SAR images by tuning the batch normalization layer of the CNN, enables us to obtain multi-temporal deep features that are pixelwise compared to identify the changed pixels. Detected changed pixels are further analyzed based on the double bounce property of the buildings in SAR images to detect the pixels corresponding to destroyed buildings. Experimental results on a dataset made up of a pair of multi-temporal VHR SAR images acquired by COSMO-SkyMed constellation on the city of L’Aquila (Italy) demonstrates effectiveness of the proposed approach.