Spatial prediction and mapping of landslide susceptibility using machine learning models

Spatial prediction and mapping of landslide susceptibility are crucial for landslide risk assessment and management. In this study, three different machine learning techniques, viz. multi-layer perceptron (MLP), self-organizing map (SOM), and classification tree analysis (CTA), are used to predict and map the landslide susceptibility in a typical landslide-prone mountainous region (Hanyuan, Southwest China). Initially, 104 identified historical landslides (centroid) for the study case as an inventory was mapped and randomly partitioned into training and test datasets in a 7:3 proportion. Fifteen conditioning factors were then chosen from different physical-geographical conditions and optimized using multi-collinearity analysis. Subsequently, MLP, SOM and CTA models were employed to carry out the spatial prediction of landslide susceptibility zones and generate corresponding landslide susceptibility maps (LSMs), further categorized into four susceptibility classes from low to very high. The performance estimation, validation and comparison of LSMs were conducted using the success rate (SR), prediction rate (PR) methods and statistical tests. The results reveal that MLP has the highest PR and SR (86.68%; 90.08%), followed by SOM (85.10%; 89.48%) and CTA (74.69%; 86.85%). All three models have very good or good capability to describe landslide susceptibility zones, and can be advanced ML-technique options for generating feasible LSMs. The resulting LSMs can effectively support the landslide risk management decision-making and land use planning formulation in the landslide-prone area.