Contrastive Learning for Ship Classification Using Real and Complex SAR Imagery

We explore the potential of contrastive learning for the classification of ships using real and complex Synthetic Aperture Radar (SAR) modalities. We employ a network that takes SAR images as input and learns a latent space for the classification of maritime objects, fishing and non-fishing vessels. In contrast with the prior work on SAR data, the model presented in this paper can operate entirely in the complex domain. We leverage existing work on the complex analogs of real-valued input, processing and output layers to deliver a contrastive learning approach that is capable of working with data that is naturally represented in the complex domain. We investigate different types of data augmentation that can be applied to real valued Ground Range Detected (GRD) images, complex valued Single Look Complex (SLC) data, and the magnitude component of SLC data and compare relative performance for contrastive learning. Our results on widely available SAR satellite data show our method can operate on complex-valued imagery without compromising accuracy compared to equivalent realvalued baselines.