A DeepLabV3+-Based Framework for Terraced Field Extraction from High-Resolution Remote Sensing Imagery

Abstract

Terraced fields are difficult to delineate from high-resolution imagery because narrow and irregular boundaries, fragmented patches, and visual similarity to adjacent cropland can lead to omission errors and boundary confusion. This study adapts DeepLabV3+ for terraced field extraction from 2 m GF-1 RGB imagery through a task-oriented integration of spatial-detail preservation, contextual modeling, and boundary refinement. The resulting framework incorporates a dilated residual encoder, a gated global–local context module, and a boundary-supervised PointRend refinement head to address spatial-detail loss, background confusion, and boundary uncertainty in terrace mapping. On the re-evaluated test set, the framework achieved 86.68% precision, 87.77% recall, 87.22% F1 score, and 77.34% target-class IoU. Relative to the unmodified DeepLabV3+ baseline, the F1 score and IoU increased by 10.01 and 14.46 percentage points, respectively. These results indicate that task-oriented integration of spatial-detail preservation, contextual modeling, and boundary refinement can improve fine-scale terraced-field delineation from high-resolution optical imagery.