Answers are grounded in this paper and cite their sources.
Gully erosion severely destroys farmland and reduces agricultural productivity worldwide, and accurate gully extraction in Southwest China is hindered by complex terrain, vegetation cover, and heterogeneous surface environments. This study proposes a novel deep learning model MSAR-Net for precise gully erosion extraction from high-resolution UAV orthophotos obtained from Yanyuan County, Sichuan Province. Built on a ResNet34 encoder, the model integrates a hybrid attention module and multi-scale dilated convolution to suppress background noise, capture multi-scale contextual features, and improve the detection of tiny gullies. Comparative experiments with mainstream segmentation models (FCN, U-Net, U-Net++, MapNet, TransUNet, SwinUNet) verified the superiority of MSAR-Net. The model achieved optimal performance with an 88.34% F1-score, 79.12% IoU, and 85.32% recall, providing an effective tool for intelligent gully erosion monitoring in complex mountainous areas of Southwest China.
No takes yet. Share an insight, caveat, or question.
Haoyu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: