Stationary Susceptibility and Dynamic Landslide Hazard Forecasting System for Georgia Using ERA5, Fuzzy Logic, and Hybrid Machine Learning

Georgia, due to its high mountains, growth of population, intensive land use, and the number of vulnerable infrastructures (roads, large engineering constructions, etc.), as well as climate change, belongs to the one of most vulnerable regions in the world, as it is prone to hazardous geological processes such as landslides (LSs). As the statistical data on LSs in Georgia are of low quality, due to political turmoils in the first decades of XXI-th century, a three-stage methodology was applied for assessing LS susceptibility. On the first stage, we perform the initial statistical assessment of LS susceptibility, using 11 weighted objective factors analysis. Then the machine learning (ML) methods were used for obtaining the final map of LS susceptibility. In conditions of Georgia, the intensive and prolonged precipitation is one of the main triggering factors for activation of LS processes. Therefore, in this study we also use hourly precipitation data, meteorological station observations, a historical landslide catalogue and hybrid machine learning approaches to construct spatial-temporal LS hazard maps under 24 h, 3-day, 5-day, and 30-day cumulative precipitation conditions. Predictive performance was evaluated retrospectively using a balanced sample of 309 landslide events and 309 susceptibility-matched controls, including a temporal train–test split. The results characterize discrimination under this sampling design; performance at the natural frequency of landslides requires independent evaluation. The results obtained show that the acceleration of global warming in recent decades affects the assessment of LS susceptibility/hazard.