Abstract
Cropland change detection (CD) in high-resolution remote sensing images is critical for cropland protection and food security. However, style differences caused by inconsistent imaging conditions (such as season and illumination) and ground object scale differences often lead to high numbers of false and missed detections. Existing approaches, predominantly relying on spatial domain features and a multiscale framework, struggle to address these issues effectively. Therefore, we propose FDFENet, incorporating a Frequency Domain Feature Exchange Module (FFEM) that unifies image styles by swapping the low-frequency components of bitemporal features. A Frequency Domain Aggregation Distribution Module (FDADM) is also introduced as a comparative alternative for handling style discrepancies. Subsequently, a Multiscale Feature Enhancement Module (MSFEM) strengthens feature representation, while a Multiscale Change Perception Module (MSCPM) suppresses non-change information, and the two modules work cooperatively to improve detection sensitivity to multiscale ground objects. Compared with the FDADM, the FFEM exhibits superior parameter efficiency and engineering stability, making it more suitable as the primary solution for long-term deployment. Evaluations on four CD datasets (CLCD, GFSWCLCD, LuojiaSETCLCD, and HRCUSCD) demonstrate that FDFENet outperformed 13 state-of-the-art methods, achieving F1 and IOU scores of 77.09% and 62.72%, 81.81% and 73.63%, 74.47% and 59.32%, and 75.95% and 61.23%, respectively. This demonstrates FDFENet’s effectiveness in addressing style differences and ground object scale differences, enabling high-precision cropland monitoring to support food security and sustainable cropland management.