FarmCD: Agricultural Information-Guided Gated Network for Farmland Change Detection From Remote Sensing Images

In recent years, farmland change detection from remote sensing images has emerged as an important research topic. Due to the complexity of farmland dynamics, existing change detection methods often perform poorly when directly applied to farmland scenarios. To address these challenges, FarmCD is proposed as a farmland-specific change detection framework that selectively focuses on farmland regions while suppressing irrelevant categories such as buildings. FarmCD integrates several specialized modules to extract and refine farmland-related change information. The Farmland Feature Aware Module (FFAM) extracts farmland features from bitemporal images through two mutually interactive branches, while the Differential Fusion Module (DFM) captures temporal variations and structural differences. To mitigate interference from non-farmland categories, the Farmland Soft-hard Gate Module (FSGM) adaptively suppresses non-farmland changes. In addition, the Multi-scale Farmland Refinement Module (MFRM) progressively enhances farmland boundary details, reducing misdetections caused by irregular farmland shapes. By integrating weakly supervised farmland mask generation and a multi-stage training strategy, the network dynamically filters and enhances farmland-relevant features throughout the change detection process. Experimental results on the public CLCD and CropSCD datasets demonstrate that FarmCD achieves competitive performance, with significant improvements over seven state-of-the-art models. The implementation of FarmCD will be publicly available in the future at https://github.com/lixinghua5540/FarmCD.