Abstract: Background Landslides in Leiyang City occur frequently, causing severe threats to local residents' lives and properties, damaging infrastructure such as roads and buildings, and disrupting the regional ecological balance by destroying vegetation and soil structure. Given these challenges, developing an evaluation method with strong adaptability and high universality to improve the accuracy of landslide susceptibility assessment has become a key focus in geological disaster research. Methods This study selects Leiyang City, Hunan Province, as the research area. It constructs an IV-RF model by coupling the Information Value (IV) model and the Random Forest (RF) model: the IV model quantifies the correlation between influencing factors and landslide occurrence through statistical analysis, while the RF model enhances nonlinear relationship recognition using ensemble learning. To optimize negative sample selection, a spatial constraint sampling strategy is introduced, which ensures negative samples are distributed reasonably in non-landslide areas and avoids spatial autocorrelation interference. For model comparison, the Receiver Operating Characteristic (ROC) curve and Area Under the Curve (AUC) value are used to analyze the IV, RF, and IV-RF models. Additionally, a comprehensive performance index integrating Accuracy, F1 score, and Matthews Correlation Coefficient (MCC) is proposed to comprehensively measure model performance from multiple dimensions. Results The research findings are as follows: (1) The IV-RF coupled model shows superior performance over single models. Its AUC value reaches 0.891, significantly higher than the IV model (0.786) and RF model (0.832). The comprehensive performance index (Accuracy+F1+MCC) of the IV-RF model is 2.593, exceeding the IV model (2.105) and RF model (2.347). Spatially, over 85% of landslide points are concentrated in extremely high-high susceptibility zones, while less than 5% are in extremely low-low susceptibility zones, fully verifying the model's strong spatial prediction capability. (2) Lithology is identified as the most critical factor affecting landslide development. Among various lithological units, thin-bedded slate and meta-sandstone interlayers show the highest contribution to landslide susceptibility due to their weak mechanical properties and vulnerability to weathering, followed by factors such as slope gradient and distance to faults. Conclusions The IV-RF coupled model effectively integrates the IV model's advantage in quantitative data interpretation with the RF model's strength in nonlinear pattern recognition, significantly improving the accuracy of landslide identification. These results provide a scientific basis for targeted landslide disaster risk prevention (e.g., early warning system construction in high-susceptibility zones), soil and water conservation measures (e.g., vegetation restoration in fragile lithological areas), and rational land spatial planning. The method demonstrates strong adaptability to complex geological environments with diverse lithology and topographic conditions, offering important technical reference for landslide susceptibility assessment in other regions with similar geological and climatic characteristics.