A machine learning framework for assessing systemic societal vulnerability to coastal flooding

This study develops a machine-learning framework for comparing and ranking household-level vulnerability to coastal flooding in the Lekki Peninsula, Nigeria. The research employed a multi-source dataset, including household surveys, key informant interviews, field observations and secondary data. The expectation-maximisation (EM) algorithm identifies latent vulnerability clusters, while support vector machines (SVMs) model nonlinear relationships to predict household membership. A probability-weighted index and confidence threshold are integrated to quantify classification reliability. The EM-SVM model achieved 98% accuracy on the test dataset, with near-perfect separability, as indicated by an area under the curve of 0.99. Spatial analysis showed that 7.4% of households were highly vulnerable, 68.9% moderately vulnerable and 23.5% had low vulnerability.

Five vulnerability profiles emerged; two exhibited relatively high education and, in one case, higher income, but remained moderately vulnerable due to dependence on flood-exposed healthcare and water infrastructure and housing types susceptible to flooding. One profile showed low vulnerability due to lower exposure of healthcare infrastructure and higher self-perceived coping capacity, indicating the role of critical infrastructure reliability and psychological preparedness. The findings reinforce the understanding that vulnerability is systemic and intersectional rather than determined by independent socio-economic attributes. Through this framework, the vulnerability approach reduces bias and offers a transparent, data-driven alternative to conventional approaches, thereby reinforcing the evidence base for equitable flood-risk planning and more targeted adaptation interventions.