Performance Evaluation of Depthwise Separable CNN and Random Forest Algorithms for Landslide Susceptibility Prediction

Landslides are one of the dangerous gravity-driven geological hazards, endangering human life and causing extensive socio-economic and ecological damages. Determining landslide-prone hotspot zones, therefore, is vital for government agencies and local governments for mitigation studies. Lately, deep learning (DL) and machine learning (ML) approaches have been employed in landslide susceptibility analyses for solving challenging problems. Notwithstanding, significant drawbacks have been reported for ML methods, including poor model variance and generalization capabilities. Therefore, DL models have often been preferred to address these limitations. The DL architecture with a separable convolution layer was proposed in this research to enhance predictive performance. The produced results revealed that the suggested DL model outperformed the ML approach by up to 9% in terms of overall accuracy, which was ascertained by the Wilcoxon signed-rank test. The proposed technique also indicated a substantial improvement in susceptibility map accuracy (~6%), according to the area under curve metric.