Advancing landslide inventorying through automated mapping, classification, and volume assessments

Landslides are dangerous hillslope failure phenomena caused by internal and external forces that can cause catastrophic release of masses, capable of destroying housing, infrastructure, and causing loss of life. Their prediction and mapping undoubtedly become a de facto goal from a civil protection perspective however, certain critical challenges persist due to their complex nature Strong events like earthquakes and rainstorms trigger tens of thousands of landslides, leading to widespread damage. Understanding their characteristics—including occurrence patterns, mobility types, kinematic progressions, volumes, and temporal resurgences—is essential for better grasping their general behaviors and hopefully predicting their occurrences. Overall, insights into these characteristics have been obtained in the Ph.D. through scalable and transferable automated methods applicable to diverse global regions. Firstly, this thesis delved into multi-temporal and spatial analysis of landslide dynamics across varying terrains, leveraging deep learning techniques. This extensive study resulted in the ability of mapping landslides both over space and time with scalability, meaning that the method can be applied across diverse regions of the earth. Experiments were conducted on the Nepalese Himalayas (2015 Gorkha earthquake), China (Wenchuan, 2008), Papua New Guinea (2018), and New Zealand (2016) where landslides were mapped over 100,000 km2 using remote sensing imagery; demonstrating the method's applicability and scalability in diverse environments. To capture a broad range of landslide movements, a detailed topological analysis of 3D landslide bodies was conducted, identifying different types of movements such as slides, flows, and falls. In addition, complex movements with multiple coupled mechanisms (e.g., slides followed by falls or flows) were also identified and separated out. The method achieved high accuracy (80-94%) in classifying landslide movements across diverse geographical and climatic regions like Italy, the US Pacific Northwest, Denmark, Turkey, and Wenchuan in China. A practical application on previously undocumented datasets from Wenchuan further demonstrated its effectiveness and adaptability. To differentiate landslide kinematic phases—separating source from runout zones—we used topological information coupled with morphometric properties (travel angle, distance, height, and kinetic proxies) across diverse settings. Calculating landslide volumes through elevation model differencing, rather than relying on total landslide length as shown in previous studies, improved volume estimates, particularly for the 2018 Hokkaido co-seismic landslides. Utilizing empirical scaling relationships on the source zones, a 30% improvement in volume prediction accuracy was achieved, showcasing the substantial potential of integrating topology with morphometric data.