Objective Due to their special geological conditions and coal mining background, coal mining subsidence areas are prone to suffer from ground cracks and subsidence, which tend to further induce geological disasters such as landslides. This makes it extremely necessary to conduct landslide sensitivity assessment of these areas using appropriate influencing factors and training models.

Methods This study investigated the Xishan coal mine area in Beijing as an example to conduct landslide sensitivity assessment of coal mining subsidence areas by coupling slope units with mathematical statistical and machine learning models. With 10 influencing factors, including topographic and geomorphic factors such as slope, aspect, and surface wetness, geological background factors like formation lithology and fault throw, and coal mining background factors such as distances from coal mine roadways and wellheads, as assessment indices, this study optimized the landslide sensitivity assessment systems using correlation analyses. Meanwhile, based on the topographic slope units determined using hydrological analysis, this study predicted the spatial landslide sensitivity using the information value (I) model, the information value - random Forest (I-RF) coupling model, and the information value - multi-layer perceptron (I-MLP) coupling model individually. Results and conclusions The results indicate that the I-RF and I-MLP coupling models exhibited higher accuracy than the independent I model. The I-RF, I-MLP, and I models exhibited areas under the curve (AUCs) of 0.861, 0.845, and 0.761, respectively, suggesting that the I-RF model enjoys the highest predictive ability and accuracy. Additionally, the effects of landslide sensitivity assessment were improved due to the introduction of geological and coal mining background factors. To further verify the practicality of the landslide sensitivity zoning, this study, using the “23.7” extreme rainfall event in Beijing- as a case study, compared the landslide sensitivity results obtained using the models and the landslide hazards determined using techniques such as large-scale aerial photogrammetry and time-series InSAR. The verification results indicate that the landslide sensitivity assessment results based on slope units and coupling models agree well with the landslide hazards triggered by the rainfall event. Therefore, the landslide sensitivity assessment results can reflect the landslide probability of various slopes to a certain extent, thus serving as a reference for the prediction and prevention of landslide hazards in coal mining subsidence areas.