Objectives Earthquake-induced landslides is a highly destructive and dangerous geologic hazard, and the accurate assessment of their susceptibility is crucial for post-earthquake relief, mitigation, and reconstruction.

Methods To address the limitations of convolutional neural network (CNN), such as susceptibility to overfitting and uncertainty in multi-source data, a novel approach based on a convolutional fuzzy neural network (CFNN) is proposed. The proposed method combines the efficient feature extraction capability of CNN with the strong uncertainty handling capability of fuzzy logic to achieve accurate modeling and feature optimization of landslide influencing factors. Additionally, the Shapley additive explanations (SHAP) feature optimization strategy is introduced to quantitatively evaluate feature importance, enabling the selection and optimization of model inputs. The 2017 Jiuzhaigou Ms 7.0, the 2022 Lushan Ms 6.1, and the 2022 Luding Ms 6.8 earthquakes are selected as the study area, which includes a total of 5 990 landslide instances covering an area of 72.147 km². A total of 16 landslide impact factors are selected for comprehensive analysis. The CFNN model is constructed and optimized to evaluate earthquake-induced landslide susceptibility in the study area.

Results The results indicate that the distance to rivers, peak ground acceleration, elevation, soil cover type, and distance to faults are the key factors influencing landslide development. Compared with the traditional CNN models, the CFNN model demonstrates superior performance and exhibits significant overfitting resistance. Furthermore, after incorporating the SHAP-based feature optimization strategy, the performance of the CFNN model is further enhanced, achieving an area under the receiver operating characteristic curve of 0.942 0, indicating excellent predictive accuracy.

Conclusions The results verify the reliability of the CFNN model in earthquake-induced landslide susceptibility evaluation, providing a new theoretical basis and technical support for the prevention and monitoring of such disasters. Moreover, these results provide a valuable reference for advancing the application of deep learning models in geohazard studies.