Advancing quantitative disaster risk assessment to strengthen climate change adaptation in Brazil: the role of data collection in the analysis of geohazard-affected areas

Abstract

Hydrological and sediment-related disasters have intensified across Brazil over the past three decades, producing severe human, economic, and territorial impacts. Climate change projections indicate further increases in extreme events, while national reports estimate a potential 40% rise in vulnerability. Therefore, this thesis aims to clarify how improved data collection and integration can strengthen quantitative disaster risk assessment and support climate-adaptive governance in Brazil. In Chapter 2, the analysis focuses on evaluating existing official risk assessment frameworks and their capacity to support effective climate-adaptive governance. A systematic assessment of the five federal surveys—the Municipal Risk Reduction Plan (PMRR), Geological Risk Survey (GRS), Susceptibility Survey (SS), Geotechnical Aptitude for Urbanization Charts (GAUC), and the Geological Hazard Survey (GHS)—revealed substantial methodological limitations. The results showed that no survey incorporates analyses of rainfall frequency, intensity, or event magnitude. Although PMRR and GRS incorporate demographic criteria, their reliance on outdated census data and highly simplified population metrics results in systematic underestimations and an incomplete socioeconomic representation of at-risk populations. Among all methodologies, GHS was the only survey to advance toward a semi-quantitative delineation of affected areas (runout), making it the most suitable survey for climate-adaptive decision-making. In Chapter 3, the role of climatic and demographic drivers in hydrological disasters in Brazil between 1991 and 2023 was quantified. Disaster occurrences rose sharply from fewer than 100 events per year in the early 1990s to more than 2,300 in recent years. Heavy rainfall events (≥100 mm/day) increased modestly (≈1.2×), while average annual precipitation remained declined. Correlation analysis and bivariate regressions combining climatic and demographic drivers showed that informal settlements and urban population were the strongest predictors of disasters frequency and severity (R2 = 0.73 and 0.66, respectively). Standardized coefficients further demonstrated that demographic indicators exert greater influence on disaster occurrence than climatic metrics, underscoring the central role of exposure and vulnerability. The results underscore the need to refine the delineation of disaster-affected areas as climatic variability and informal urban expansion continue to amplify uncertainties. In Chapter 4, a multivariate linear regression (MLR) workflow was developed and assessed for predicting landslide transport-deposit runout distance (TD-RoD) using data from the 2022 Petrópolis disaster (Rio de Janeiro). Traditional landslide size-scaling approaches—post-failure morphometries (PFM)—showed minimal predictive capacity in densely urbanized mountainous terrain (R2 = 0.06 ± 0.03). Integrating PFM predictors with pre-event geoenvironmental site condition (GSC) variables derived from remote sensing and field surveys substantially improved accuracy (R2 = 0.50 ± 0.01). However, poor generalization in the validation dataset (R2 = 0.13 ± 0.02) demonstrated that linear method was unable to fully capture the complex and context-dependent controls governing landslide runout mobility. In Chapter 5, ten non-linear machine learning algorithms (MLA) approaches were employed to overcome these constraints, yielding substantially higher predictive performance. The GSC models achieved high accuracy (training R2 = 0.79 ± 0.15; validation R2 = 0.46 ± 0.12), while combining landslide PFM with GSC attributes yielded the best results (training R2 = 0.91 ± 0.08; validation R2 = 0.55 ± 0.08). CatBoost delivered the most balanced performance across accuracy, stability, and consistency (training R2 = 0.58 ± 0.31; validation R2 = 0.36 ± 0.20). Explainable artificial intelligence (XAI) techniques, including Shapley Additive Explanations (SHAP) and Individual Conditional Expectation (ICE), clarified the internal behavior of the MLA models. These methods revealed non-linear, context-dependent response patterns governing landslide runout mobility, consistent with landslide dynamics documented in experimental, numerical, and field-based studies. Overall, these findings demonstrate that enhancing data collection is essential for improving the evaluation of geohazard-affected areas and advancing quantitative disaster risk assessment framework in Brazil. By clarifying the role and value of these datasets, this thesis offers evidence-based guidance to strengthen future climate change adaptation and disaster risk reduction strategies.