SAR Image Classification Using Mixed Spatial-Spectral Information and Pre-trained Convolutional Neural Networks

Abstract

The recent technological advancements in remote sensing lead to an increased importance regarding the analysis of satellite data targeting security and surveillance tasks. Although the availability of data products is constantly augmented and the advances in Deep Learning technologies are constant, Synthetic Aperture Radar (SAR) image classification remains a challenge in the remote sensing domain because standard convolutional neural network-based architectures may encounter difficulties in recognizing objects that are characterized by similar texture, but different backscattering patterns. Moreover, training deep learning architectures requires a large volume of annotated data, which, in general, represents an obstacle, especially in the case of the remote sensing domain. This article addresses complex-valued SAR image classification through both spatial and Fourier-domain features, extracted by means of pretrained neural networks. While spatial features allow extracting knowl-edge regarding the structure and texture of the objects from intensity images, the physical properties of the objects are learned from radar spectrograms. In addition, we show that considering different polarizations of the SAR sensor, we are able to obtain better visual classifications. The experiments are conducted over Sentinel-1images, which are freely available for download under the Copernicus initiative.