Speckle-Robust SAR Image Change Detection for Real-World Monitoring Using Learnable Wavelet Transforms and Attention

Synthetic Aperture Radar (SAR) change detection (CD) supports real-world monitoring by identifying land-cover changes from multi-temporal acquisitions under all-weather, day‑night conditions. However, practical deployment remains challenging due to speckled noise and scene variability, which often lead to false alarms and missed subtle changes. This paper presents WASNet (Wavelet-AttentiveSEBAM Network), which combines learnable wavelet downsampling and reconstruction with dual-domain attention to improve robustness and precision. A Wavelet-based Transform Module (WBTM) performs lossless multi-scale analysis with trainable wavelet filters, preserving low- and high-frequency cues that are important for small changes. To enhance discrimination in noisy scenes, the SEBlock and MHSA-enhanced Bi-dimensional Aggregation Module (SEBAM) decouples channel recalibration (squeeze-and-excitation) and spatial dependency modelling (multi-head self-attention), improving feature selectivity while suppressing speckle-driven artefacts. We further employ a composite MSE–KL loss to balance reconstruction fidelity and distribution alignment. Experiments on four public SAR CD benchmarks demonstrate consistent improvements over recent methods, particularly under strong speckle and subtle structural variations, supported by quantitative metrics and qualitative change maps. Source code is available at https://github.com/Mohamed-DL/WASNet.