Demystifying SAR with attention

Abstract

Synthetic Aperture Radar (SAR) imagery is indispensable for earth observation, offering the ability to capture data under challenging conditions such as cloud cover and darkness. However, its grayscale format and speckle noise hinder interpretability and pose significant challenges for traditional processing methods. This study introduces an innovative framework for SAR image colorization, leveraging an Attention-Based WGAN-GP (Wasserstein GAN with Gradient Penalty). The model incorporates multi-head self-attention mechanisms to enhance feature extraction, capture long-range dependencies, and dynamically suppress noise through a novel variance-based attention adjustment mechanism.

Extensive evaluations on Sentinel-1 and Sentinel-2 datasets across diverse terrains, including agriculture, urban areas, barren land, and grasslands, demonstrate the model’s superiority over existing approaches. It achieves an LPIPS score of 0.27, SSIM of 0.76, and an average inference time of 0.22 s, showcasing its ability to preserve spatial coherence and perceptual quality even in complex, noisy environments. This capability enables real-time applications in disaster management, flood monitoring, and urban planning, providing actionable insights and advancing the state-of-the-art in SAR image processing.