Satellite Image Segmentation with Deep Residual Architectures for Time-Critical Applications

This work addresses the problem of training a convolutional neural network for segmenting satellite images in emergency situations, where images to be segmented are potentially very different from training images. Such case is particularly challenging due to the large intra-class variations in image statistics between images captured at different locations by different sensors. We propose a convolutional encoder-decoder network architecture where the encoder is built around residual networks. We show that the proposed architecture enable learning features able to generalize the learning process across images with largely different statistics. Our architecture can accurately segment images that have no reference in the training set, whereas a minimal refinement of the trained network significantly boosts the segmentation accuracy.