A wavelet-enhanced Mamba network and a high-quality dataset for remote sensing change detection

Recently, the Mamba architecture has been widely adopted in remote sensing change detection (RSCD), owing to its linear-complexity advantage and strong spatiotemporal modelling capability. Given that existing Mamba-based RSCD methods inevitably suffer from structural information loss during global modelling and lack dedicated strategies to address the inherent noise caused by seasonal and illumination variations in bi-temporal images, we propose a wavelet-enhanced Mamba network for RSCD, named WMCD. Specifically, we design a wavelet feature enhancement (WFE) module and integrate it into a Siamese VMamba backbone. This module first performs feature decomposition via discrete wavelet transform, then applies a temporal noise suppression strategy to dynamically eliminate pseudo-change noise in high-frequency subbands, and further introduces a gated fusion unit to modulate the low-frequency subbands. Additionally, considering that existing public RSCD datasets suffer from performance bottlenecks and poor generalization due to registration errors and limited change types, we construct a new small-scale, high-quality, and diversified change detection dataset, BSE-CD. Extensive experiments on the three public datasets, LEVIR-CD, SYSU-CD and WHU-CD, as well as the BSE-CD dataset, validate the effectiveness of the proposed WMCD, achieving the highest F1 scores of 90.80%, 82.81%, 93.88% and 95.60%, respectively.