Debris-flow susceptibility assessment using deep learning algorithms with GeoDetector for factor optimization

Abstract

Accurate assessment of debris flow susceptibility is crucial for disaster prevention and evaluation in mountainous regions. This study proposes a debris flow susceptibility assessment method based on the Convolutional neural network (CNN)-bidirectional long-short-term memory neural network (BiLSTM)-attention mechanism (CNN-BiLSTM-attention) deep learning model that combines geographic information technology and artificial intelligence algorithms. The study area is the Xiaojiang River watershed in the Yunnan-Guizhou Plateau, a region prone to frequent debris flows. First, a debris flow conditioning factor system and a debris flow sample set are constructed using multi-source data, including remote sensing, geological, and precipitation data, with the watershed unit as the assessment unit. Second, the geoDetector method is adopted to explore the optimal combination of conditioning factors. Finally, the CNN-BiLSTM-Attention model is applied to quantitative analyze debris flow susceptibility, and its performance is compared against three deep learning models and three machine learning models. The findings are summarized as follows. Lithology, elevation difference, average slope, 24H maximum precipitation, average elevation, average modified normalized difference water index (MNDWI), Melton ratio, average land surface temperature (LST), and channel gradient are the dominant factors influencing the debris flows development. The prediction performance of CNN-BiLSTM-Attention is significantly better than that of the other six models. Its area under the receiver operating characteristic curve (AUC), accuracy (ACC) and mean absolute error (MAE) reach 0.903, 0.953 and 0.165 respectively, demonstrating excellent prediction accuracy and generalization performance. This study offers new insights for debris flow susceptibility analysis.