Noise Robust SAR Image Classification Using Siamese Spiking Neural Networks

In advancements in artificial neural network (ANN) learning, significant enhancements have been made to the accuracy and robustness of synthetic aperture radar (SAR) image classification. However, using ANNs involves high computational costs. On the other hand, spiking neural networks (SNNs), recognized as the third generation of neural networks, have been introduced due to their energy efficiency. Recent research has explored the use of SNNs for SAR classification, but deeper architectures and noise-robust settings have not been fully explored, resulting in a gap in accuracy compared to ANN models. In this letter, we propose a deep Siamese SNN with speckle noise augmentation that addresses the limitations of shallow SNNs in previous studies, enhances robustness against noise, and leverages information maximization through Poisson encoding and soft resetting. We have validated the effectiveness of our model against various ANN and SNN models on different variations of the MSTAR dataset, including those with limited samples and noisy images. Our results demonstrate the potential of SNNs to achieve comparable performance to ANNs in SAR image classification.