Statistical Learning for Integrating Environmental and Economic Dimensions of Coastal and Marine Resilience

Coastal and marine ecosystems are complex socio-environmental systems where physical processes, ecological dynamics, and human interventions interact across multiple spatial and temporal scales. Understanding and managing these systems requires the integration of several data sources and the application of flexible, data-driven statistical frameworks. This thesis brings together a collection of studies that develop and apply modern statistical learning methods to analyse marine processes, biological resources, and environmental investments in the Italian and Mediterranean contexts. The first part of the thesis focuses on the physical components of the marine environment. It begins with the development of a bivariate finite-mixture model based on Generalised Additive Models for Location, Scale, and Shape (GAMLSS) to jointly analyse meteorological tides and wave heights. This approach allows for the identification of latent sea-state regimes and reveals how wind, atmospheric pressure, and fetch interact to shape sea conditions. The following study introduces a zero-inflated hidden semi-Markov model with covariate-dependent sojourn parameters to model the temporal evolution of flooding events in the Venice Lagoon, providing a probabilistic framework for hazard assessment. The third paper extends the analysis to biological systems, applying a Bayesian spatio-temporal model to explore variability in sardine landings across the Mediterranean. By integrating fishery-dependent data within a hierarchical framework, it quantifies spatial heterogeneity and temporal trends in a key fishery resource. The second part of the thesis turns to the socio-economic dimension of coastal and marine ecosystem management, addressing the crucial link between environmental risk and public investment. A harmonised and georeferenced dataset of Italian public investments for soil and coastal defence is constructed, enabling detailed analysis of spatial and temporal funding patterns. Building on this resource, the R package PublicWorksFinanceIT is developed to automate data retrieval, integration, and visualisation, promoting transparency and reproducibility in the use of open government data. Finally, a Bayesian spatio-temporal model is applied to quantify the relationships between public investments, environmental drivers, and demographic pressures. The results highlight spatial disparities, delayed responses to environmental covariates, and the influence of local socio-economic factors on funding allocation. Overall, the studies presented in this thesis illustrate how statistical learning can bridge the gap between physical, ecological, and policy-oriented analyses of coastal and marine systems. By combining mixture models, hidden processes, and hierarchical Bayesian frameworks with open and reproducible data infrastructures, the research advances both methodological innovation and practical understanding of how societies monitor, model, and manage their marine and coastal environments. The integrated perspective developed here supports the transition toward data-informed decision-making for sustainable and resilient coastal governance.