FL-SCD: a semi-supervised change detection method based on frequency domain convolution and local feature enhancement

Utilizing high-resolution remote sensing images (HRSIs) as data sources for change detection (CD) is one of the most prominent research directions in remote sensing. Many high-precision CD networks are based on supervised learning methods; when faced with a large amount of unlabeled HRSIs, they struggle to adequately extract valuable geometric and visual attributes. To address this issue, we propose a semi-supervised CD method based on frequency domain convolution and local feature enhancement (FL-SCD), which integrates frequency and spatial domain features to enhance feature extraction capability, consisting of two main components: the frequency domain convolution branch (FDCB) and the local feature enhancement module (LFEM). The FDCB captures global frequency features beneficial for frequency-domain change recognition and then fuses them with deep spatial features to promote global information extraction from unlabeled data. Meanwhile, the LFEM further aggregates locally similar structural features of ground objects, capturing the visual features and local geometric structures of target objects, and fusing them with shallow features, thereby effectively extracting valuable geometric and visual attributes. Experimental results on three publicly available high-resolution CD datasets demonstrate that our approach achieves promising performance, even with limited labeled data.