Irregular Agricultural Field Delineation Using a Dual-Branch Architecture From High-Resolution Remote Sensing Images

Current agricultural field (AF) delineation methods using high-resolution remote sensing (RS) images encounter two primary challenges: 1) the interactions between boundary features and field features are often overlooked, resulting in imprecise field boundaries and 2) the efficacy of delineating AFs with fuzzy boundaries, irregular shapes, and small sizes in mountainous and hilly regions necessitates systematic evaluation. To deal with these challenges, we propose a boundary-field interaction network, namely, BFINet, leveraging multitask learning techniques for AF delineation. BFINet comprises two branches: a core branch for AF delineation and an auxiliary branch for boundary prediction that furnishes fine-grained boundary information to enhance geometric feature learning. To model the interactions between boundary semantics derived from low- and high-level semantics obtained from the field extraction branch, we introduce a semantic interaction module (SIM), facilitating AF identification in complex terrains. A boundary enhancement module is integrated to refine spatial details of high-level semantics, thereby alleviating boundary fuzziness. We also design a dense multibranch fusion module (DMFM) to augment the semantic representation of BFINet’s multiscale information for various fields. Extensive experiments were conducted on these datasets from Hunan (HN) and Chongqing (CQ) in China, featuring small AFs with irregular shapes and fuzzy boundaries, and compared BFINet with five existing AF delineation methods. Our findings indicate that BFINet excelled in identifying diverse AFs, achieving superior attribute and geometric accuracies over competing methods. It effectively delineated AFs with various sizes and shapes while minimizing interference from other ground objects.