Weighted Difference Image Guided Siamese Mamba Network for SAR Change Detection

Abstract

Synthetic aperture radar (SAR) change detection is important for real-world monitoring because it can operate under all-weather and day–night conditions, but accurate detection remains challenging due to speckle noise and complex scene variations. This paper presents a Siamese Mamba network for SAR change detection, designed to improve pixel-wise change localisation while maintaining a simple and efficient architecture. The proposed method processes the bi-temporal SAR images through a shared encoder, exploits feature differences between temporal branches, and incorporates a weighted difference image as complementary guidance to enhance change-sensitive representation learning. A decoder then produces the final change probability map, which is refined through thresholding and morphological post-processing. Experiments on three public SAR change detection datasets demonstrate strong and consistent performance. The proposed method achieved F1-scores of 95.42% in San Francisco, 93.61% in Farmland and 92.15% in Yellow River. These results indicate that the proposed Siamese framework provides an effective and practical solution for SAR change detection, offering a good balance between accuracy, efficiency, and simplicity of implementation.