Development of risk maps for flood, landslide, and soil erosion using machine learning model

Natural hazards, such as flood, landslide, and erosion, are the reality of human life. spatial prediction of these hazards and their effectiveness factors are extremely important. The main goal of this study was to prepare multi-hazard probability mapping (flood, landslide, and gully erosion) of the Gorganrood Watershed. In addition, different machine learning models such as Random Forest (RF), Support Vector Machine (SVM), Boosted Regression Tree (BRT), and Multivariate Adaptive Regression Spilines (MARS) were applied. First, a flood, landslide, and gully erosion inventory map was produced using GPS in the field surveys and Google Earth. Factors affecting the hazards were identified, and GIS maps were prepared. The MARS model (AUC = 99.1%) provided the highest predictive performance for flood, landslide, and gully erosion hazards. However, for flood and landslide, the RF model exposed excellent and good performance, respectively. According to the variable importance analysis, drainage density (89.4%), digital elevation model (30.5%), and rainfall (41.7%) were consistently highly ranked variables for flood, landslide, and gully erosion, respectively. Multi-hazard maps can be a valuable tool for the conservation of natural resources and the environment, as well as for sustainable land use planning in multi-hazard-prone areas.