UC Berkeley

This dissertation incorporates artificial intelligence (AI) techniques on remote-sensing data for ground failure detection, mobility assessment, and infrastructure monitoring. First, the use of AI for landslide detection is investigated. Although an increasing body of work is observed on this topic, a systematic investigation of the factors (input and algorithms) that affect the accuracy of the machine-learning co-seismic landslide detection model has not been attempted. This study leverages the state-of-the-art detailed 3D inventory of more than 700 landslides triggered by the Mw 6.5 Lefkada earthquake on November 17, 2015. The result highlights that feature selection is the most essential factor for successful landslide detection, but the number of features needed is not particularly high. The geospatial distribution and size of the training sample are also important. Input data resolution and machine learning algorithms are the secondary factors that influence detection accuracy. Geospatial distribution affects the training sample size needed to create an accurate landslide detection model, and a wider geospatial distribution of training samples generates a more precise landslide detection model. The work is expanded to consider the generality of the above results for two additional co-seismic landslide events, namely the 2016 Kaikōura earthquake and the 2021 Nippes earthquake, with the goal to identify the commonalities and differences in the success of the machine learning-based landslide detection model. It is found that although feature selection is the most vital factor in the landslide detection model, both topographic and spectral features are useful, with spectral features being most significant in two of the study areas due to their geologic and climatic setting. The input data resolution and training sample size similarly influence the model performance for the three earthquake events, but the importance of segmentation and machine learning algorithms varies across events.Next, a simple mechanistic-model that is based on the Voellmy friction law as incorporated in Rapid Mass Movement Simulation Debris Flow (RAMMS-DF) is tested against statistically significant observations of landslide runout for hundreds of mapped rock avalanches triggered by the Mw 6.5 Lefkada earthquake on November 17, 2015. It is found that the dry-Coulomb friction (μ) controls the simulation's performance, whereas the simulation is less sensitive to viscous-turbulent friction (ξ), especially for large values of ξ. The simulation's accuracy positively correlates with landslide source area, height, and 3D travel distance. The model does not match very well landslides with small source areas (<4,000 m2), but in these cases, it systematically overestimates landslide runout, i.e., it is inherently conservative. The fourth part of this dissertation leverages lessons learned from the damage observed along Highway 1 during the January 2021 atmospheric river event. A remote sensing-based methodology is developed for system-level monitoring and assessment following natural disasters. It is shown that remote sensing indicators of vegetation loss can detect the occurrence of debris flows and ground failure and indicate the severity of highway damage. Damage severity is correlated to increasingly broader distribution and a lower minimum value of the vegetation loss curve. The last part of this dissertation aims to develop a methodology for fully autonomous remote-sensing-based monitoring of mines. Specifically, the detection of mining instability using high-resolution satellite imagery for eight recent failure cases is considered: the 2022 Jagersfontein tailings dam failure, the 2022 Pau Branco iron ore mine landslide, the 2020 Carmen copper mine landslide, the 2020 Singrauli fly ash dam breach, the 2019 Córrego De Feijão tailings dam failure, the 2018 Cadia gold mine tailings dam failure, the 2014 Mount Polley mine tailing dam failure, and the 2013 Bingham Canyon copper mine landslide. The results show that remote sensing indexes can successfully detect mining failure. In summary, this dissertation demonstrates that new approaches that leverage Artificial Intelligence (AI) and remote-sensing data can be valuable for ground instability detection following natural hazards and can set the stage for fully-autonomous infrastructure monitoring in an expedited and efficient manner.