MTFSR: Multitemporal and Spatial Feature Reconstruction Denoising Network for Remote Sensing Change Detection

With the widespread application of Convolutional Neural Networks (CNNs) in Remote Sensing (RS) technologies, Change Detection (CD) has attracted increasing attention in environmental monitoring research. In remote sensing image change detection tasks, researchers often face challenges such as multi-scale change capture inadequacy due to feature noise interference and difficulties in spatiotemporal feature alignment. To address these issues, this paper proposes a Multi temporal and spatial feature reconstruction denoising network for remote sensing change detection (MTFSR). MTFSR first utilizes ResNet50 to extract deep features from dual-temporal images and innovatively introduces Bézier curves to smooth the features through feature redistribution, along with brightness enhancement to effectively suppress noise interference and enhance change areas. Meanwhile, a spatial attention mechanism and Discrete Cosine Transform (DCT) are incorporated to further improve the spatial expression and frequency domain perception of the features. Then, multi-scale filtering is applied to accurately align the frequency and spatial domains, effectively solving the spatiotemporal feature inconsistency problem. On this basis, the model designs a Dynamic Attention Unit (DAU) that includes both local and global attention to generate comprehensive feature maps and calculates difference weights through linear fusion, further strengthening the feature expression ability to more accurately capture change information. Finally, the decoder generates high-precision change detection results. To verify the effectiveness of MTFSR in Remote Sensing Change Detection (RSCD) tasks, we conducted tests on three challenging datasets (LEVIR-CD, GZ-CD, and WHU-CD). The experimental results show that MTFSR significantly outperforms traditional methods and most existing advanced techniques in change detection performance.