A Feature-Enhanced and Edge-Refined Network for Cropland Parcel Extraction from Sentinel-2 Imagery

Accurate identification of arable land, as the foundation of the high-standard farmland construction, impacts the crop layout, accurate management of water and fertilizers, and intelligent control. Due to the 10-m resolution limitation of Sentinel-2 imagery, there is feature overlap within individual pixels of the satellite imagery. This leads to fragmented semantic features during farmland identification, and adjacent plots often appear unclear and intertwined. To address these issues, a Hierarchical Agricultural Segmentation Network (HASNet) was proposed. Built upon the classic encoder-decoder structure, this HASNet model incorporates an expanded feature enhancer (DFE) module to recover weak features and reconstruct cropland features (e.g., edges and shapes) that are obscured by mixed pixels. It also introduces a lightweight strip spatial attention (LSSA) mechanism to capture long-range features unique to farmland. Furthermore, it used a pyramid decoding module (PDM) to refine cropland parcel boundaries. Taking a farm in Xinjiang Uygur Autonomous Region, a semantic segmentation dataset of cultivated land was constructed based on Sentinel-2 imagery. Through accuracy comparisons, visualizations, and inferences, HASNet achieved an MIoU of 88.52% and a Kappa coefficient of 87.82%, outperforming mainstream models such as Unetformer and MPFUnet. Ablation experiments confirmed the effectiveness of the DFE, LSSA, and PDM modules in feature capture and edge refinement. The large-scale image sliding inference experiment prevented the seam effect and demonstrated its practicality. In summary, HASNet provides low-cost technical and theoretical support for the intelligent monitoring of high-standard farmland.