Frequency-Domain Feature Interaction Combined With Multiscale Attention for Remote Sensing Change Detection

Zhongxiang Xie, Shuangxi Miao, Zhewei Zhang, Xuecao Li, Jianxi Huang

Abstract: Change detection (CD) in remote sensing images has seen significant advancement due to the powerful discriminative capabilities of deep convolutional networks. However, the domain gap and pseudo-changes between the bi-temporal images, caused by variations in imaging conditions such as illumination, shadow, and background, remain a challenge. Furthermore, multiscale variations in complex scenes complicate the accurate identification of change regions and their boundary delineation. To address these issues, this article introduces the frequency-domain feature interaction and multiscale attention mechanism network (FIMANet). Specifically, to mitigate the impact of pseudo-change interference, the FIMANet reduces the domain gap and facilitates information coupling within intralevel representations through frequency-domain feature interaction (FDFI). To prevent information loss and noise introduction, a multiple kernel inception (MKI) module is devised to capture multiscale features and perform progressive fusion. Finally, to enhance the extraction of changes in scale-sensitive regions, the FIMANet constructs a cross-scale feature aggregator (CSFA) module, composed of attention at various scales and a transformer, to capture fine-grained details and global dependencies. Comparative experiments with nine methods on three commonly used datasets validate the effectiveness of FIMANet, achieving the highest ${F}1$ -score of 73.98% on the CLCD dataset, 90.55% on the WHU-CD, and 91.01% on the LEVIR-CD. The code is available at https://github.com/zxXie-Air/FIMANet

External IDs:doi:10.1109/jsen.2025.3583301