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

We address the problem of training a convolutional neural network for satellite images segmentation in emergency situations, where response time constraints prevent training the network from scratch. Such case is particularly challenging due to the large intra-class statistics variations between training images and images to be segmented captured at different locations by different sensors. We propose a convolutional encoder-decoder network architecture where the encoder builds upon a residual architecture. We show that our proposed architecture enables learning features suitable to generalize the learning process across images with 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.