Economic Risk Assessment of Future Debris Flows by Machine Learning Method

Abstract

A reliable economic risk map is critical for effective debris-flow mitigation. However, the uncertainties surrounding future scenarios in debris-flow frequency and magnitude restrict its application. To estimate the economic risks caused by future debris flows, a machine learning-based method was proposed to generate an economic risk map by multiplying a debris-flow hazard map and an economic vulnerability map. We selected the Gyirong Zangbo Basin as the study area because frequent severe debris flows impact the area every year. The debris-flow hazard map was developed through the multiplication of the annual probability of spatial impact, temporal probability, and annual susceptibility. We employed a hybrid machine learning model—certainty factor-genetic algorithm-support vector classification—to calculate susceptibilities. Simultaneously, a Poisson model was applied for temporal probabilities, while the determination of annual probability of spatial impact relied on statistical results. Additionally, four major elements at risk were selected for the generation of an economic loss map: roads, vegetation-covered land, residential buildings, and farmland. The economic loss of elements at risk was calculated based on physical vulnerabilities and their economic values. Therefore, we proposed a physical vulnerability matrix for residential buildings, factoring in impact pressure on buildings and their horizontal distance and vertical distance to debris-flow channels. In this context, an ensemble model (XGBoost) was used to predict debris-flow volumes to calculate impact pressures on buildings. The results show that residential buildings occupy 76.7% of the total economic risk, while road-covered areas contribute approximately 6.85%. Vegetation-covered land and farmland collectively represent 16.45% of the entire risk. These findings can provide a scientific support for the effective mitigation of future debris flows.