A Deep Learning Method for Field Boundary Delineation From Remote Sensing Imagery With High Boundary Connectivity

Abstract

Accurate delineation of agricultural field boundaries from remote sensing imagery is fundamental for precision agriculture and policy-making. However, existing deep learning methods struggle with this task due to the difficulty of balancing semantic context with fine detail. Their unique imaging features (e.g., intricate linear nature, frequent occlusions) further challenge the detection of continuous boundaries. To address these issues, we propose a High Connectivity and Accuracy Field Boundary Network (HCFNet). HCFNet employs a novel two-stage framework where a lightweight region segmentation task provides explicit semantic constraints to a dedicated boundary delineation task. The boundary task is powered by an enhanced edge network featuring four key innovations: a Linearly Connected Deformable Convolution (LCDConv) to capture fine boundary morphology, a Non-Local Operation (NLO) to resolve local ambiguities using global context, a Continuity Restoration Decoder (CRD) that expands the receptive field and bridge occluded segments, and a Hierarchical Attention Fusion Module (HAFM) for adaptive multi-scale feature fusion. We conducted extensive experiments on diverse datasets, including high-resolution Gaofen-1, Google Earth, UAV imagery, and the global FTW benchmark. HCFNet consistently outperformed state-of-the-art models, achieving superior boundary accuracy (F1 score) and significantly improved connectivity. Our results demonstrate that HCFNet provides a robust and efficient solution for field boundary delineation, with its specialized modules proving particularly valuable in data-scarce scenarios common in practical applications. The code for HCFNet is available at (https://github.com/BNU-zhu/HCFNet).