Application of machine learning methods for landslide risk mitigation

Landslides in complex geological settings are a widespread and dangerous natural hazard that, in the context of climate change, poses significant challenges for modern society. Despite knowledge about triggering mechanisms and monitoring techniques, conducting comprehensive studies of landslide susceptibility and emergency management in regions like Emilia-Romagna, Italy, often entails significant obstacles. This PhD project analyzes the multifaceted aspects of landslide prediction and management in this geologically complex region. The research initially focused on improving rainfall thresholds for forecasting, comparing conventional empirical-statistical approaches with machine learning (ML) techniques. Results demonstrated that ML methods improved predictive accuracy and reduced false positives. However, this advancement also highlighted the persistent challenge of input data quality and volume. A key innovation emerged: developing a unified platform that integrates the improved rainfall thresholds and ML predictions into operational early warning systems. The catastrophic rainfall events of May 2023 in Emilia-Romagna posed a significant challenge to our research, necessitating an unprecedented rapid response. Considering this urgent need, a multi-institutional collaboration was quickly established, leading to a comprehensive landslide inventory documenting 80,000 polygons using high-resolution aerial imagery. Our research pivoted to explore rapid automated mapping techniques, starting with trials using the U-Net neural network in severely affected municipalities like Casola Valsenio. More advanced algorithms, such as SegFormer, were subsequently employed across diverse settings including Modigliana, Predappio, and Brisighella. This urgent application of our research offered valuable insights into the practical challenges of rapid landslide response, informing innovations for future methodologies. In conclusion, this research demonstrates the transformative potential of integrating machine learning with traditional approaches in landslide prediction and management. The improved prediction models, early warning platform, comprehensive landslide inventory, and rapid mapping techniques collectively contribute to building resilient communities capable of effectively responding to and mitigating the impacts of landslides in the face of increasing climate-related risks.