Assessing the Impact of Multi-Resolution DEMs and Sampling Strategy Uncertainty on Deep Learning-Based Debris Flow Susceptibility Mapping

Abstract

Debris-flow susceptibility mapping requires precise topographic representation and balanced sampling strategies to ensure model robustness and generalizability. This study systematically quantifies the influence of digital elevation model (DEM) resolution (6.5, 12.5, 30, and 90 m) and sampling-strategy uncertainty on the predictive performance of advanced deep-learning (DL) architectures, including one and two-dimensional convolutional neural networks (CNN1D, CNN2D), recurrent neural networks (RNN), and long short-term memory (LSTM) models. A field-verified debris-flow inventory comprising 108 catchments and 13 conditioning factors derived from multi-resolution DEMs and remote-sensing datasets was used for model construction. To evaluate sampling uncertainty, 100 symmetrical iterations of debris-flow and non-debris-flow samples were executed, resulting in a 6.7-10.5% increase in mean accuracy with optimized sample selection. Factor importance derived from Random Forest and Frequency Ratio analyses identified rainfall as the dominant control on debris-flow occurrence. Model performance, assessed through multiple statistical metrics, revealed that the LSTM consistently outperformed other architectures, achieving a maximum accuracy of 0.929 and AUC of 0.973 at 12.5 m resolution. The proposed framework provides a reproducible and scalable approach for multi-resolution DFSM and quantification of sampling-related uncertainty in complex mountainous terrain.