Enhancing Landslide Susceptibility Mapping Through a Hybrid Model Utilizing Bivariate Methods and Convolutional Neural Networks

Landslides frequently result in human casualties and economic losses in mountainous regions, being particularly severely affected in the Himalayan areas. Mitigating hazards and risks serves as an effective means to the use of landslide susceptibility mapping. This study is dedicated to exploring a more advanced model by integrating convolutional neural networks (CNN) with bivariate methods for LSM. Four models were developed and compared: CNN, certain factors (CF) hybridized with CNN, frequency ratio (FR) hybridized with CNN and information value method (IV) hybridized with CNN. Initially, a total of 313 landslide events were identified and systematically incorporated into a landslide inventory map, with 12 predisposing factors concurrently selected for subsequent analysis. Subsequently, the dataset was randomly partitioned into two subsets, wherein 75% was allocated for model training and 25% reserved for validation purposes. Finally, the performance of the models was validated and compared using area under the curve (AUC) and statistical metrics. The results showed that the IVCNN model demonstrated superior performance (AUC 0.974 and accuracy = 93.3%), compared with CNN model (AUC 0.91 and accuracy = 86.9%). Additionally, the LSM generated by the IVCNN model effectively predicted susceptible areas by identifying key factors as maximum elevation difference exceeding 1200 m. In conclusion, the hybrid combines the intuitiveness of bivariate methods with the ability of CNN to process complex patterns and thus significantly enhance LSMs' quality. By analyzing the key factors contributing to landslide occurrences and leveraging the robust image processing capabilities of CNN, an innovative model for assessing landslide susceptibility was developed.