MAEMACD: a MAE-enhanced multiresolution attention network for remote sensing image change detection

The task of remote sensing image change detection aims to identify significant changes between two images, which is crucial for understanding terrestrial dynamic changes. However, detecting multi-scale changes of targets in complex scenes while avoiding non-semantic changes (e.g. illumination, sensor noise) remains challenging. This paper presents a MAE-enhanced multi-resolution change detection model (MAEMACD). The model takes bi-temporal image pairs as input and first extracts features through a shared backbone network (BN), which are then fed into the Multi-order Gated Fusion Module (MGFM) to enhance representation capability. The concatenated features from MGFM are subsequently passed to the Multi-Scale Hybrid Attention Module (MHAM) and decoded by the prediction head to generate the final change map. Meanwhile, the concatenated features from MGFM are also transmitted in parallel to the Masked Autoencoder module (MAE) to produce image differences; by comparing these reconstructed differences with the original bi-temporal differences, the framework further optimizes BN and MGFM to better capture true change regions. Extensive experiments on multiple benchmark datasets verify the effectiveness of our model. Compared with the baseline BIFA model, MAEMACD achieves improvements in F1-score of +0.58%, +3.94%, and +5.30% on the LEVIR-CD, WHU-CD, and DSIFN-CD datasets, respectively.