Low Resolution for DNN in SAR

Deep Learning offers numerous techniques for data analysis and feature extraction, useful in Earth Observation. However, a persistent problem is the small number of labeled datasets and the variety of the data, due to the different instruments used for collecting scenes. For example, to what extent can a deep neural network handle image patches of different resolutions, and to what extent do we need very high-resolution images? In this paper, we study the impact of the resolution in SAR images on the capability of a convolutional neural network to classify urban scenes. The goal is to find urban classes that are recognizable in images with different resolution. We start by training a convolutional neural network on OpenSARUrban [14], a dataset with patches originating from high-resolution scenes (20 meters), and then we lower the resolution of the patches by several factors and train the same network on them. Finally, we need to apply these findings on “real” data. With this purpose, we present VHRUrbanSAR, a novel dataset containing 8234 SAR image patches of from very high-resolution scenes (5 meters), whose purpose in this work is fine-tuning. Both datasets have similar classes, which means that we can identify those urban classes that are recognized by both image types. We conclude that: (1) Reducing the resolution of the images by half has little effect on the performance; (2) Some classes in urban areas can be classified with a network trained on lower resolution images.