Improving Landslide Prediction and Rapid Assessment through Machine Intelligence

Landslides represent a significant global threat, standing among the most destructive natural hazards with profound consequences for both communities and ecosystems. The inherent complex, nonlinear, and discontinuous nature of landslide dynamics make it particularly difficult to accurately predict and mitigate their impacts. This thesis addresses these challenges by enhancing automated landslide localization and prediction using state-of-the-art artificial intelligence (AI) algorithms and remote sensing. The primary aim of the thesis is to improve the reliability and adaptability of landslide localization and prediction systems. This objective is pursued through a series of seven interrelated studies, each contributing to a comprehensive framework for landslide localization and prediction. The first study introduces the High-Resolution Global Landslide Detector Database, designed to enhance AI-based landslide mapping across diverse regions and events by leveraging high resolution multispectral satellite imagery. The subsequent three chapters investigate and develop automated methods for detecting and mapping landslides under all-weather, day-and-night conditions using Synthetic Aperture Radar (SAR). Here, deep learning techniques are employed to recognise the landslides footprint in SAR, even in adverse conditions like cloud cover, where multispectral imagery becomes unusable. This research also introduces the SAR-based Landslide Rapid Assessment tool, designed for rapid assessment of large earthquake-triggered landslide events, making it a ready-to-use solution for future disaster response. The thesis further examines AI's application in real-world disaster scenarios, as demonstrated in the Hualien City earthquake in Taiwan, where approximately 7,000 co-seismic landslide failures were localized within two hours of satellite image acquisition. This highlights the effectiveness of automated detection methods in emergency response while addressing still existing challenges. The final chapters develop and evaluate models for forecasting landslide displacement and identifying precursory failure indicators using neural networks and advanced statistical techniques. These models incorporate historical landslides dynamics and a range of triggers, such as rainfall and reservoir water level changes, to enhance predictive accuracy of future landslide movements. Additionally, advanced prediction models are integrated with granular systems theory to detect regime shifts in slow-moving landslides, which may serve as precursors to catastrophic failure. Overall, the findings of this thesis significantly improve disaster preparedness and resilience for communities and infrastructure vulnerable to landslide hazards. By introducing advanced AI-driven methods for landslide detection, mapping, and forecasting, this research offers more reliable and adaptable approaches for mitigating landslide impacts and enhancing early warning systems. Additionally, most of the codes, tools, and datasets developed in this thesis are open source and freely accessible at https://github.com/lorenzonava96.