Trustworthy and Generalisable AI for Natural Hazards: Foundation Models, Physical Consistency, and the Path from Experiment to Operation

Natural hazards — landslides, debris flows, floods, wildfires, and cascading or compound events — are intensifying under climate change, yet the AI methods used to study them are advancing faster than our ability to trust them. A wave of recent work applies machine learning and remote sensing to hazard detection, susceptibility mapping, and forecasting, but results too often fail to hold up across regions, sensors, events, and timescales. The unresolved problems are now methodological rather than merely applied: weak physical consistency, domain shift and out-of-distribution behaviour, limited and biased labels, poorly quantified uncertainty, low interpretability, and a persistent gap between benchmark accuracy and operational reliability. Foundation models, self-supervised learning, physics-informed and hybrid modelling, and digital twins offer routes forward, but their robustness and transferability remain largely unproven.This Research Topic focuses on the reliability of AI for natural hazards rather than on any single hazard or sensor. It asks how AI can be made transparent, generalisable, and scientifically grounded enough to move from experimental demonstrations into dependable operational systems for Earth-system research and disaster-risk reduction. We are especially interested in contributions that test generalisation across environments and hazard types, embed physical understanding into data-driven models, quantify and calibrate uncertainty, and establish reproducible benchmarks — and in work on compound and cascading hazards, where interactions between processes make trustworthy prediction hardest. Central questions include how AI can be validated across heterogeneous settings, how physical constraints improve robustness, and how outputs can be communicated credibly to practitioners, emergency managers, and decision-makers.We welcome contributions on the following themes:• Foundation models and self-supervised learning for hazards, including data-scarce and under-observed regions• Physics-informed, physics-guided, and hybrid AI that enforces physical consistency in hazard processes• Generalisation across regions, sensors, events, and timescales: domain adaptation, transfer learning, and out-of-distribution detection• Uncertainty quantification, calibration, explainable AI, and trustworthy prediction• Compound, cascading, and multi-hazard interactions and their representation in AI models• Digital twins and intelligent early-warning systems as operational endpoints• From benchmark to operation: reproducible workflows, AI-ready datasets, open benchmarking protocols, and evaluation beyond accuracy• Robust multimodal and multiscale fusion (satellite, UAV, LiDAR, InSAR, geophysical, meteorological, hydrological, in situ) under sensor and label heterogeneity• Effective communication of AI-derived hazard information for decision support and disaster-risk reductionAppendix: We welcome original research, methods and data papers, reviews, perspectives, and application-focused case studies — with a preference for work that demonstrates generalisation, physical consistency, or operational value. Topic editors: Alessandro Novellino, Sansar Raj Meena, Daniel Kibirige, Xiaochuan Tang Source: https://www.frontiersin.org/research-topics/85252/trustworthy-and-generalisable-ai-for-natural-hazards-foundation-models-physical-consistency-and-the-path-from-experiment-to-operation