SAR Image Classification with Knowledge Distillation and Class Balancing for Long-Tailed Distributions

Abstract

The classification of data following long-tailed distributions is a challenging problem that is encountered in practical situations. In the case of Synthetic Aperture Radar (SAR) image classification, the problem is even more difficult due to noisy data and lack of large datasets for training. In this paper, we explore a cross-modal framework for knowledge distillation from Electro-Optical (EO) data to SAR data, with class balancing considerations. Our methodology leverages coupled EO-SAR data and involves the following stages: (1) EO network training, (2) SAR network training with response-based knowledge distillation from the EO network, and (3) class balanced training of the SAR network to account for long-tailed distributions in the data. Our model can be used with a different network backbones without placing any constraints on the network architectures for EO or SAR. Class balancing is utilized during training to deal with imbalance across classes and an equal diversity loss is considered for improved training. We test our approach on an EO-SAR coupled dataset that is highly imbalanced and demonstrate performance gains in each stage of our model.