Multiscale Spatial Frequency Fusion and Prior Change Guidance Network for Remote Sensing Change Detection

Deep learning techniques have made impressive progress in the field of remote sensing change detection (RSCD) in recent years. However, existing RSCD methods still exhibit limitations in bi-temporal feature fusion, making it difficult to adequately mine critical change information. Moreover, they often overlook the semantic inconsistency between features at different levels during feature aggregation, which limits the accurate reconstruction of the internal structure of change objects. To address the above issues, this article proposes a multiscale spatial frequency fusion and prior change guidance network, called MPNet, aiming to enhance the complete reconstruction of change objects. The proposed MPNet has two advantages. First, a multiscale spatial frequency fusion (MSFF) module is proposed to capture the bi-temporal features in the frequency domain and different scale spatial domains, and perform dynamic adaptive fusion through the attention mechanism, thereby realizing the adequate mining of global and local change information. Second, a prior change guidance (PCG) module is designed to generate a prior change mapping by fusing high-level semantic information with low-level texture details. This prior mapping guides multilevel feature learning, effectively correcting semantic discrepancies across different feature layers and enabling the extraction of more discriminative change feature representations. Experimental results on the LEVIR-CD, WHU-CD, and SYSU-CD datasets demonstrate that the proposed MPNet significantly outperforms other state-of-the-art (SOTA) methods in the complete detection of the internal structure of change objects. The code is available at https://github.com/NjustHGWei/MPNet.