Check dams are widely used to regulate floods and intercept sediment by altering local topography and promoting vegetation growth. With increasing extreme rainfall events and aging dam structures, systematic evaluation of sediment interception efficiency has become crucial for mitigating soil loss and enhancing basin flood control. However, quantifying sediment retention efficiency under rainstorm conditions remains challenging. Here, a small-watershed soil loss model was developed at rainfall-event scale using an explainable machine-learning algorithm, based on field observations from the Yanhe River Basin and Chabagou watershed on the Loess Plateau. The model was applied to assess check dam efficiency and its influencing factors during the “7·26” rainstorm on July 26, 2017 in the Chabagou Watershed. The algorithm demonstrated strong potential for elucidating soil loss mechanisms, revealing that rainfall factors dominated and their positive effect on soil loss increased exponentially. During non-rainstorm periods, individual check dams achieved 100% interception efficiency, while the dam system reached 84.09%. However, during the “7·26” rainstorm, individual dam efficiencies were 1.82–57.11%, and the dam system intercepted 20.89%. On average, failed dams had only 5.13% sediment interception efficiency, which was significantly lower than that of non-failed dams (26.90%). The presence of upstream check dams enhanced efficiency (17.82% vs. 5.75%). Grassland-dominated (13.11%) and grassland–cropland co-dominated (20.93%) watersheds exhibited markedly greater sediment interception efficiencies than bare land–grassland-dominated watersheds (2.75%). This study provides recommendations for soil loss control and check dam layout optimization to reduce soil loss and enhance the flood control capacity of watersheds on the Loess Plateau.
