Soil erosion is a global ecological and environmental issue that severely degrades terrestrial ecosystems. A range of soil and water conservation measures, notably the construction of check dams in gullies, have been widely implemented to mitigate soil erosion and sustain agricultural productivity. In this study, Ningxia province in China was selected as the study area. High-resolution Google Earth imagery and digital elevation model (DEM) data were integrated with three representative deep learning semantic segmentation models—FCN, U-Net, and DeepLab v3+—to achieve automatic extraction and spatial distribution analysis of engineered check dams. Model performance was quantified using overall accuracy (OA), F1-score, and mean intersection over union (mIoU), among other metrics. The results demonstrated that U-Net outperformed FCN and DeepLab v3+ across all evaluation metrics. On the test dataset, U-Net’s F1-score exceeded those of FCN and DeepLab v3+ by 3.89% and 7.08%, while mIoU increased by 2.17% and 6.57%, demonstrating superior boundary delineation. Based on the precise area extraction by U-Net, a piecewise empirical equation was subsequently developed to relate predicted silted land area to actual sediment volume, achieving R2 values of 0.92 for small dams and 0.96 for large dams. Spatial distribution analysis revealed that check dams are predominantly concentrated in the southern mountainous and hilly-gully regions, moderately distributed in the central areas, and relatively sparse in the northern plains. Overall, this study demonstrates the feasibility and effectiveness of deep learning-based semantic segmentation for automated check dam mapping and sediment volume estimation.
Meng et al. (Mon,) studied this question.