Snow avalanche susceptibility mapping from tree-based machine learning approaches in ungauged or poorly-gauged regions

Abstract

Mapping avalanche susceptibility is essential for disaster management, especially in ungauged or poorly-gauged regions. Most existing machine-learning methods have a shortfall in multi-source, heterogeneous spatial data requirements and overfitting. Faced with these deficiencies, this study has developed four novel tree-based machine-learning models (Catboost, LightGBM, RF, and XGBoost) to predict avalanche susceptibility, along with the data support of remote sensing images (e.g., Superview-1, Sentinel-2, and Calibrated Enhanced-Resolution Passive Microwave Daily Brightness Temperature), Meteorological data (e.g., WRF simulation), topography, and limited observation in Tianshan Mountains, China. There are 21 causative factors in the selection, training, and validation processes. Then, the SHAP value is introduced to quantify the variable-based share of contribution to the possibility of avalanches on both global and local scales. The results demonstrate the following: (1) All four models are potent in assessing avalanche susceptibility due to similar patterns, with Catboost being the best with its eight indices: Accuracy (0.9249), TPR (0.9920), FAR (0.0080), TNR (0.8904), PPR (0.8229), NPR (0.0046), CSI (0.8175), and HSS (0.8404) superior to others; (2) The Catboost provides a reliable susceptibility map illustrating moderate and high susceptible areas up to 54.12% of the total, mainly concentrated in the west and southeast; (3) Topographic factors (distance to rivers, aspect, and relative slope position) and meteorology (precipitation) are the most effective for susceptibility modeling. The results above indicate the significant potential of tree-based machine-learning with multi-source and heterogeneous spatial data in obtaining high-quality susceptibility maps without too much field observation. It is of great importance for avalanche prevention in regions with complex terrain conditions and sparse data.