A Novel Probabilistic Graphical Model Framework for Flood Risk Assessment Across Nigerian River Basins

Comprehensive flood risk assessment across Nigeria's 24 major river basins requires a mathematical framework capable of representing the joint probabilistic relationships among hydrological, climatological, exposure, and vulnerability variables operating at multiple spatial scales from sub-basin to national level. This dissertation develops an original probabilistic graphical model framework for integrated flood risk assessment that overcomes the limitations of existing single-variable and correlation-based approaches. The theoretical foundation employs directed acyclic graph Bayesian network structures to represent causal dependencies among rainfall intensity, soil moisture, antecedent flood conditions, river routing dynamics, exposed population, structural vulnerability, and economic loss. An original theoretical contribution is a proof that the factored joint distribution over the full variable set preserves monotonicity of flood risk with respect to stochastic dominance orderings of rainfall intensity, which is a necessary condition for coherent risk-based policy ranking. The Bayesian network is parameterised using a combination of hydrological simulation model outputs from the HBV-96 model calibrated to gauged basins, remote sensing-derived flood inundation data from Sentinel-1 SAR imagery, and household vulnerability survey data from 3,200 respondents across the Niger, Benue, Anambra, and Hadejia river basins. Posterior inference using variational Bayes enables computationally tractable risk probability estimation at 30-metre spatial resolution. The framework generates spatially explicit flood risk maps that outperform current NIHSA flood zonation maps with 34 percent higher validation accuracy against documented flood event records. Keywords: Bayesian network, flood risk assessment, probabilistic graphical model, Nigerian river basins, variational inference.