Deep Learning (DL) is increasingly being integrated into natural hazard management, providing new capabilities for monitoring, forecasting, and mitigating disasters such as wildfires, floods, and volcanic activity. However, applying DL in this domain presents key challenges, including handling vast Earth Observation datasets, addressing data imbalance due to the rarity of extreme events, and ensuring models are both explainable and uncertainty-aware for reliable decision-making.

This webinar will provide an overview of recent advances in DL techniques for natural hazard applications, highlighting three core topics. First, the use of spatiotemporal datacubes enables structured representation of geophysical data, facilitating more effective machine learning workflows. Second, Bayesian deep learning and uncertainty quantification techniques enhance model trustworthiness, a crucial requirement for operational disaster response. Finally, emerging AI methodologies, including self-supervised learning and teleconnection-informed forecasting, offer new avenues for improving predictive accuracy and spatio-temporal generalization.

Real-world applications will be discussed, including the development of AI-driven wildfire risk assessment models, flood mapping using Synthetic Aperture Radar (SAR) time series, and volcanic unrest detection with deep learning. The webinar will conclude with insights into current research challenges and future directions in AI-driven hazard forecasting, emphasizing the potential for scalable, explainable, and robust machine learning models in disaster management.