Assessing glacial lake outburst flood risk in the Eastern Himalayas: a Bayesian neural network framework

Glacial lake outburst floods (GLOFs) represent a growing global hazard with increasing impacts in high-mountain regions across Asia, South America, and the Arctic. In the Eastern Himalayas, the frequency and intensity of GLOFs have risen markedly over the past three decades, driven by accelerated glacier retreat, altered precipitation patterns, and unregulated development in downstream valleys. However, systematic risk assessment remains limited, constrained by insufficient understanding of glacier–lake interactions and the lack of predictive models that integrate hazard complexity and uncertainty. These deficiencies weaken the effectiveness of early warning systems and adaptation strategies, not only in the Himalayas but also in other vulnerable mountain regions worldwide. This study presents a robust, two-stage methodology to assess GLOF risk in the Eastern Himalayas, offering a scalable framework relevant to other glaciated regions. In the first stage, we employ a multi-criteria decision-making (MCDM) approach integrating 14 parameters—including topographic configuration, geomorphology, hazard proximity, meteorological extremes, and anthropogenic exposure—weighted using a hybrid of Shannon Entropy and the Analytic Hierarchy Process (AHP). This model classifies glacial lakes into four GLOF risk categories: low, moderate, high, and very high. In the second stage, we implement a Bayesian Neural Network (BNN) to replicate and refine the initial classification while quantifying predictive uncertainty—marking the first application of BNNs in GLOF risk science. The model achieves an overall accuracy of 93.4%, identifying 66 low-risk, 74 moderate-risk, 56 high-risk, and 18 very high-risk lakes. Importantly, 87.77% of lakes exhibit predictive variance below 0.1, and none exceed 0.165, indicating high model reliability across risk categories. By integrating expert-driven decision-making with uncertainty-aware machine learning, this research offers a globally relevant advancement in GLOF hazard assessment. The methodology provides a transferable template for other climate-sensitive mountain regions facing similar threats. It supports the development of more adaptive, evidence-based disaster risk reduction strategies aligned with global frameworks such as the Sendai Framework for Disaster Risk Reduction and the UN Sustainable Development Goal. In an era of accelerating cryosphere change, such predictive and scalable approaches are critical for safeguarding downstream communities, infrastructure, and transboundary water security.