Unsupervised SAR Imagery Feature Learning with Median Filter-Based Loss Value

Abstract

The scarcity of open SAR (Synthetic Aperture Radars) imagery databases (especially the labeled ones) and sparsity of pre-trained neural networks lead to the need for heavy data generation, augmentation, or transfer learning usage. This paper described the characteristics of SAR imagery, the limitations related to it, and a small set of available databases. Comprehensive data augmentation methods for training Neural Networks were presented, and a novel filter-based method was proposed. The new method limits the effect of the speckle noise, which is very high-level in SAR imagery. The improvement in the dataset could be clearly registered in the loss value functions. The main advantage comes from more developed feature detectors for filter-based training, which is shown in the layer-wise feature analysis. The author attached the trained neural networks for open use. This provides quicker CNN-based solutions implementation.