Debris-Flow Susceptibility Assessment in China: A Comparison between Traditional Statistical and Machine Learning Methods

Debris flows, triggered by dual interferences extrinsically and intrinsically, have been widespread in China. The debris-flow susceptibility (DFS) assessment is acknowledged as the benchmark for the mitigation and prevention of debris flow risks, but DFS assessments at the national level are lacking. The role of human activities in the DFS assessment has always been overlooked. On the basis of a detailed inventory of debris-flow sites and a large set of environmental and human-related characteristics, this research presents the comparative performance of the well-known information value (IV), logistic regression (LR) and random forest (RF) models for DFS assessments in China. Twelve causative factors, namely, elevation, slope, aspect, rainfall, the normalized difference vegetation index (NDVI), land use, landform, geology, distance to faults, density of villages, distance to rivers and distance to roads, were considered. Debris-flow susceptibility maps were then generated after the nonlinear relationship between the debris-flow occurrence and the causative factors was captured. Finally, the predictive performance of the three maps was evaluated through receiver operating characteristic (ROC) curves, and the validation results showed that areas under the ROC curves were 81.98%, 79.96% and 97.38% for the IV, LR and RF models, respectively, indicating that the RF model outperformed the other two traditional statistical methods. The importance ranking of the RF model also revealed that distance to roads, slope and rainfall dominated the spatial distribution of debris flows. This is the first experiment to compare between the traditional statistical and machine learning methods in DFS studies for the whole of China. Our results could provide some empirical support for China’s policymakers and local practitioners in their efforts to enable residents to be less vulnerable to disasters