HSRRSFPD and MEGRDNet: empowering high-resolution farmland parcel mapping in complex terrains

• A large-scale, high spatial resolution remote sensing farmland dataset (HSRRSFPD) covering various terrains was constructed. • A multi-scale Gaussian edge refinement dual-branch network (MEGRDNet) was proposed. • Refined farmland parcel mapping was conducted in complex terrains. Existing methods in farmland extraction tasks still face limitations due to significant scale differences among farmland parcels and insufficient boundary perception capabilities, which undoubtedly hinders the promotion and application of large-scale and detailed farmland parcel mapping. To address these technical challenges of fine-grained farmland extraction in complex terrain areas, this study constructed a high spatial resolution remote sensing farmland parcel dataset (HSRRSFPD) covering multiple geomorphic regions (plains, hills, and mountains). HSRRSFPD includes 8,441 image chips with sizes of 256 x 256 pixels, and its total coverage area exceeds 550.00 km 2 . Based on HSRRSFPD, a novel multi-scale edge Gaussian refinement dual-branch network (MEGRDNet) was proposed for refined farmland parcel mapping. MEGRDNet integrated a multi-scale spatial (MSS) branch to address the challenges posed by the diversity of spatial scales of ground objects. Additionally, an edge Gaussian refinement (EGR) branch was developed to enhance the continuity and integrity of farmland boundaries. Compared with existing state-of-the-art algorithms, MEGRDNet achieved maximum improvements of 20.03% in IOU and 10.45% in F1 score. It significantly enhanced the accurate representation of multi-scale farmland boundaries under different terrains. Furthermore, the synergistic effect of the MSS and EGR branches improved MEGRDNet’s IOU performance by 21.93% compared to the Baseline model. In practical applications, a main grain-producing region (Sichuan Province) in China was used for fine-grained farmland parcel mapping at a 1 m spatial resolution. The results were highly consistent with the local government statistical data (GSD). Compared with existing farmland products, MEGRDNet provided finer boundary information, higher spatial resolution, and more accurate spatial representation. The findings and methods of this study are of great significance for land resource optimization, crop monitoring, and food security assurance in complex terrain regions of China or globally.