Prediction of debris flow hazard based on multi-source remote sensing data: A case study of Dulongjiang Township, Yunnan Province

Abstract

Debris flows, characterized as a highly destructive natural hazard with sudden onset and extensive impact zones, pose severe threats to mountainous regions worldwide, particularly in geologically vulnerable areas like China's Yunnan Province. Accurate prediction of these events is crucial for disaster prevention and mitigation, yet it remains challenging due to the complex interplay of environmental factors. This investigation presents an advanced debris flow prediction framework for Dulongjiang Township in Yunnan Province by integrating multi-source remote sensing data and machine learning (ML) techniques. By combining optical, Synthetic Aperture Radar (SAR), and topographic datasets, we develop a high-resolution grid-based approach that captures both static predisposing factors and dynamic precursory signals. Among six evaluated ML models, the Multilayer Perceptron (MLP) demonstrated superior performance, achieving an Area Under the Precision-Recall Curve (AUPRC) score of 0.94, with recall and precision of 0.93 and 0.86, respectively, in debris flow prediction. The incorporation of Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR)-based surface deformation significantly enhanced prediction accuracy compared to traditional static-factor models, establishing a novel methodology for improved early warning systems in mountainous regions. This research provides valuable insights for disaster prevention and could be adapted to other geohazard-prone areas worldwide.