Prediction of the Periglacial Debris Flow in Southeast Tibet Based on Imbalanced Small Sample Data

Abstract

Using data sourced from 15 periglacial debris flow gullies in the Parlung Zangbo Basin of southeast Tibet, the importance of 26 potential indicators to the development of debris flows was analyzed quantitatively. Three machine learning approaches combined with the borderline resampling technique were introduced for predicting debris flow occurrences, and several scenarios were tested and compared. The results indicated that temperature and precipitation, as well as vegetation coverage, were closely related to the development of periglacial debris flow in the study area. Based on seven selected indicators, the Random Forest-based model, with its weighted recall rate and Area Under the ROC Curve (AUC) greater than 0.76 and 0.77, respectively, performed the best in predicting debris flow events. Scenario tests indicated that the resampling was necessary to the improvement of model performance in the context of data scarcity. The new understandings obtained may enrich existing knowledge of the effects of main factors on periglacial debris flow development, and the modeling method could be promoted as a prediction scheme of regional precipitation-related debris flow for further research.