Landslides drivers and thresholds: a comparative study of Brazilian municipalities
This master dissertation aims to evaluate the predictive potential of precipitation and temperature variables for landslide occurrence, with an initial application in the municipality of Santos, São Paulo (Brazil), followed by a comparison with other municipalities in Brazil. While there remains a need for locally adapted early warning approaches, this study addresses a gap in the literature by testing predictive models calibrated to the conditions of municipalities in developing countries. Logistic Regression and XGBoost models were initially applied using lagged precipitation and temperature as predictors. Monotonic constraints were included in the XGBoost model to reflect domain knowledge. Both models were evaluated using nested cross-validation with hyperparameter tuning. Partial dependence plots were used to interpret variable influence. Results indicate that rainfall on the day of the event is the most significant predictor, while higher temperatures are associated with reduced landslide risk. A Generative Adversarial Network (GAN) will be explored in future stages. Environmental and socio-environmental variables, including slope, soil, geology, vegetation, land use, and the presence of informal settlements, will be incorporated to contextualize meteorological triggers and to support future comparisons across municipalities.