Consideration of spatial heterogeneity in landslide susceptibility mapping using geographical random forest model

Most previous studies of landslide susceptibility mapping (LSM) have not contemplated spatial heterogeneity and the commonly used models for LSM are aspatial, which could reduce model performance. Therefore, aiming to evaluate the applicability of spatial algorithms to predict landslide susceptibility, the performance of geographical random forest (GRF) was evaluated, in comparison to random forest (RF) and extreme gradient boosting (XGBoost). Based on the results, GRF presented the better performance (AUC = 0.876), followed by RF (AUC = 0.748) and XGBoost (AUC = 0.745). GRF also provided the most suitable susceptibility map. While RF and XGBoost presented almost 50% of the study area as susceptible, the GRF presented more concentrated susceptibility areas spatially, with a reasonable area for moderate (15.55%), high (8.73%) and very-high (2.59%) susceptibility classes. Finally, it can be inferred that spatial assessment may improve model performance, and that spatial models have a great potential for LSM.

Automated summary

This study evaluated the performance of geographical random forest (GRF) compared to random forest (RF) and extreme gradient boosting (XGBoost) for landslide susceptibility mapping. GRF showed better performance in predicting susceptibility and provided the most suitable susceptibility map with concentrated vulnerability areas. The spatial assessment improved model performance and spatial models have great potential for landslide susceptibility mapping.