Abstract: The current landslide susceptibility assessment system lacks unified and scientifically grounded standards for selecting factors that influence landslide development, leading to inconsistencies in evaluation results. To improve the accuracy of the landslide susceptibility assessment system, a method based on machine learning (ML) , which innovatively integrates factor importance selection and soil moisture content was proposed. Focusing on landslides in the Fumin County in Yunnan Province, this study integrated remote sensing data, auxiliary data, and field survey data to compile comprehensive landslide history records. Soil moisture content factors were extracted using SAR satellite backscatter coefficients and DEM data, and 15 evaluation factors were ranked for importance using XGBoost and Lasso regression models. Multicollinearity assessments of landslide influencing factors (LIFs) were conducted to identify the most distinctive factors. The landslide susceptibility in the Fumin County was evaluated using LightGBM and Random Forest models, before and after factor importance selection. The results show a significant improvement in assessment accuracy after factor selection, with the LightGBM model exhibiting superior evaluation performance with an AUC value of 0.91, demonstrating its effective application in landslide susceptibility assessment. This study highlights the impact of factor importance selection on the landslide susceptibility assessment system, effectively incorporates areal soil moisture content factors into landslide influencing factors, and optimizes the precision and reliability of susceptibility assessment results, providing new insights for landslide disaster prevention.