A progressive framework combining unsupervised and optimized supervised learning for debris flow susceptibility assessment

This research aims to improve the reliability of debris flow susceptibility (DFS) assessment, which is crucial for disaster prevention and mitigation in mountainous regions. A progressive framework was proposed and applied to Pinggu District, Beijing, China. First, 16 debris flow predisposing factors (DFPFs) were selected, and the slope structure and soil stability were incorporated to account for material sources. Watershed units instead of grid units were used to extract data. Then, the multi-collinearity among the factors was reduced by using variance inflation factors (VIF) and information gain (IG), and 13 DFPFs were retained. Three unsupervised learning algorithms (i.e., affinity propagation (AP), Gaussian mixture model (GMM) and self-organizing maps (SOM)) were used to optimize the sampling strategy of non-debris flow units. Subsequently, the occurrence probability of debris flows in each unit was predicted by using four supervised learning algorithms (i.e., logistic regression (LR), random forest (RF), adaptive boosting (ADAB) and extreme gradient boosting (XGB)). They were optimized by a state-of-the-art meta-modeling approach. Finally, the DFPF's importance was ranked. The main contributions of our framework are establishing a high-quality data set and optimizing the prediction algorithms. The results show that the tree-based models perform well, and the boosting-based algorithms outperform the bagging-based algorithms. Supervised learning is more suitable for DFS assessment than unsupervised learning. Debris flows are most likely to occur on a dolomite consequent or diagonal slope with a relief amplitude above 540 m.