Local-to-Global Shape-Consistent Learning for Boundary-Reliable Agricultural Object Extraction from High-Resolution Remote Sensing Imagery

Accurate agricultural object extraction from high-resolution remote sensing imagery requires not only high regional segmentation accuracy but also reliable object geometry. Yet most semantic segmentation frameworks are optimized primarily for pixel- or region-level agreement, so high overlap scores can coexist with boundary discontinuities, object merging, and structurally inconsistent predictions. To address this mismatch, we propose Attention-enhanced DeepLab V3+ with Local-to-Global Shape Consistency (ADLSC), a local-to-global framework that coordinates semantic discrimination, adaptive boundary representation, and complete-object geometric consistency. Target-aware multiscale context strengthens target–background discrimination, contour-adaptive representation adjusts sampling to local boundary variation while preserving narrow inter-object separation, and a staged hierarchical objective extends optimization from regional agreement to curvature- and shape-related consistency. On cropland parcel and greenhouse facility datasets, ADLSC improves IoU by 8.28 and 15.92 percentage points and matching rate by 24.32 and 42.88 percentage points, respectively, relative to DeepLab V3+, while reducing curvature and shape errors. The larger gains in object matching than in regional overlap show that explicitly coupling local contour adaptation with object-level geometric constraints improves aspects of segmentation quality that pixel-oriented objectives alone do not capture, supporting more reliable object-based agricultural mapping and spatial analysis.