A Lightweight Wavelet-Aligned Difference and Mask-Guided Fusion Network for Change Detection

Remote sensing change detection (CD) plays an important role in many application domains. To meet the practical demand for both high accuracy and efficiency, and address the fundamental challenges of bitemporal misalignment, insufficient difference representation, and change and non-change coupling, this article proposes a lightweight wavelet-aligned difference and mask-guided feature fusion network (WDMF-Net). WDMF-Net first employs a Siamese MobileNetV2 with an improved adjacent-level aggregation module to extract multiscale complementary features. A wavelet-enhanced feature alignment module (WFAM) is then introduced to enhances local boundaries and global structures of these features in frequency domain, followed by context-consistent spatial alignment to reduce bitemporal misregistration. On this basis, a collaborative feature difference module (CFDM) is designed to model the intermediate complementary relationship between channel and spatial dimensions, enabling accurate localization and capture of contextual differences. Finally, a mask-guided feature enhancement module (MFEM) is employed, where foreground and background mask priors are progressively involved in multiscale difference fusion to explicitly decouple changed and unchanged regions. Experiments demonstrate that WDMF-Net outperforms 16 SOTA methods on five public CD datasets, and achieves a favorable balance among detection accuracy, computational efficiency, and generalization capability. The code is available at https://github.com/LYT-Works/WDMF-Net.