Assessment of a farmland-constrained deep learning framework for fine-scale extraction of farmland shelterbelts from Sentinel-2

Farmland shelterbelts are a key component of oasis agricultural ecosystems in arid regions, playing an important role in wind erosion control and farmland ecological security. At fine spatial scales, however, the extraction of narrow and linearly distributed farmland shelterbelts from medium-resolution satellite imagery remains challenging. These challenges arise not only from spectral confusion and severe class imbalance, but also from the tendency of pixel-level misclassifications to propagate into spatially structured errors in the final extraction results. To assess the performance of a farmland-constrained deep learning framework for fine-scale farmland shelterbelt extraction, this study focuses on the 11th Regiment of Alar City in southern Xinjiang, China. Sentinel-2 multispectral imagery acquired between September and November 2023 was used, and reference data were constructed through manual interpretation. By introducing farmland distribution data as spatial constraints, the extraction process was restricted to agriculturally relevant areas to improve the spatial plausibility of the results. Within this framework, two representative semantic segmentation models-a convolution-based U-Net++ and a transformer-based U-Net (TransUNet)-were evaluated to examine differences in extraction performance and spatial error characteristics. Model performance was assessed using conventional pixel-level accuracy metrics, including intersection-over-union (IoU), Dice coefficient, precision, and recall, together with map-level indicators such as producer’s accuracy (PA) and user’s accuracy (UA). The results show that, although both models achieve comparable pixel-level extraction accuracy, their extraction reliability differs markedly at the map level. The TransUNet-based framework produces more spatially coherent and structurally reasonable farmland shelterbelt extraction results, particularly for narrow and fragmented shelterbelts occupying a small proportion of the agricultural landscape. In contrast, the U-Net++ model exhibits pronounced expansion-type commission errors within cropland interiors. These findings indicate that, for fine-scale farmland shelterbelt extraction, framework evaluation should emphasize robustness against spatial error propagation rather than pixel-level accuracy alone, providing methodological insights for reliable farmland shelterbelt extraction in arid agricultural regions.