Debris flow susceptibility in the Jinsha River Basin, China: A Bayesian assessment framework based on geomorphodynamic parameters

Abstract

A major challenge in assessing debris-flow susceptibility at the scale of large mountainous river basins lies in the excessive reliance on simplified topographic metrics. Existing approaches often fail to account for the cascading and dynamically coupled interactions among channel gradient, discharge, and sediment supply. This oversight limits the accuracy and robustness of spatial predictions. To address this gap, we present a novel framework for debris-flow susceptibility assessment grounded in a process-based indicator system derived from geomorphic dynamics, using the Jinsha River Basin as a case study. Our method integrates key parameters that characterize landscape evolution - including stream power, extreme rainfall events, surface erodibility, and sediment connectivity - into a Naïve Bayes probabilistic classification model. By employing kernel functions, we accommodate both continuous and discrete variables, enabling the probabilistic estimation of debris-flow occurrence across small, medium, and large magnitude classes. Model validation across the Jinsha River Basin yields a prediction accuracy of 63 %. Notably, empirical testing against the "8.21"Jinyang debris-flow event in 2023 reveals a high degree of spatial agreement between predicted high-risk zones and observed disaster footprints. Feature importance analysis indicates that surface erodibility is the dominant contributor to susceptibility, followed by connectivity, stream power, and extreme precipitation. Approximately 32 000 high-risk gullies (> 200 m in length) exhibit a power-law distribution, clustering within a 30 km buffer on both sides of the main stem of the Jinsha and Yalong Rivers in their middle and lower reaches. These regions are shown to be strongly associated with infrequent but high-probability events, which tend to drive large-scale debris-flow disasters. Amid intensifying climate change and the rapid expansion of infrastructure in alpine canyon regions, the dynamic datasets we construct - such as stream power and sediment connectivity - offer a quantitative basis for risk-informed planning and mitigation. This modelling approach represents a scalable and physically grounded paradigm for debris-flow hazard assessment, offering broad applicability to other high-relief mountainous environments worldwide.