Controls of Morphometric and Climatic Catchment Characteristics on Debris Flow and Flood Hazard on Alluvial Fans in High Mountain Asia: A Machine Learning Approach

Abstract

Debris flows and floods pose considerable hazards to populated areas of High Mountain Asia (HMA). Debris flows are generally more hazardous than floods, and therefore identification of process type is important for hazard assessment and mitigation. Prior statistical assessments, though informative, typically considered a limited number of parameters, excluded climatic variables, and failed to address classification probability and uncertainty. Here we developed a machine learning model to determine process type and its likelihood for a diverse set of 1,793 catchments in HMA using a wide range of morphometric and climatic parameters. We classified the alluvial fans of these catchments as either debris flow or flood dominated based on surface morphology. A data set of morphometric (e.g., catchment area, slope, relief, Melton ratio) and climatic features (e.g., temperature and precipitation regime, freeze–thaw cycles, glacier and permafrost presence) per catchment was subsequently built, and a CatBoost machine learning model to quantify debris flow and flood probabilities was employed. The CatBoost model has a high classification accuracy compared to traditional approaches, and offers the advantage of providing classification uncertainty. Results show that catchment slope, area, and perimeter are the main morphometric controls on process type across HMA, in line with previous work, and further show that including climate information leads to a minor improvement of model performance. These findings shed light on controls on debris flow and flood occurrence in mountainous area, showcase the potential of machine learning models in mountain hazard research, and provide insights for assessing risks.