Delineating Fish‐Scale Pits to Assess the Soil and Water Conservation Effects by Using Multiscale Attention Networks

Fish‐scale pits (FSPs) are small‐scale soil–water conservation measures often used on the arid and semi‐arid Loess Plateau. However, due to their small size and complex geomorphological background, reliable extraction by remote sensing is difficult. Thus, quantitative assessments of conservation effectiveness are limited. In this study, an enhanced U‐Net framework for FSP extraction was developed from UAV‐derived multispectral and topographic data. Analyses based on SHAP and Spearman for feature selection revealed that the input combination RGB + slope reduced redundancy and improved the spectral and geometric signatures of the FSPs. To further enhance recognition performances for small targets and their outlines, pyramid attention and multilevel feature upsampling modules were integrated into the model. The improved model yielded an IoU of 93% in the test area and exhibited superior performance in boundary reconstruction and target identification under complex terrain conditions. Based on the extracted high‐accuracy spatial distribution of FSPs, their soil and water conservation effects were further quantified using the SCS‐CN and RUSLE models. Under a 50‐mm rainfall scenario, FSPs reduced run‐off depth by 0.29–1.29 mm, corresponding to run‐off reduction efficiencies of 9%–21%. The erosion modulus was reduced by 0.16–1.79 t·hm −2 ·a −1 , with erosion reduction rates of 2%–11%. Collectively, these results demonstrate that high‐precision mapping of FSPs can forge a reliable foundation for assessing the hydrological and erosion‐control functions of microtopographic conservation measures and for optimising regional soil and water conservation planning.