Independent-Target Machine Learning and a 2040 Coastal Governance Horizon for Bonny Island, Niger Delta

Operational coastal-risk decision support in the Niger Delta is constrained by three gaps: the absence of integrated socio-physical risk surfaces, the under-validation of machine-learning models against independent observational targets, and the lack of medium-horizon planning frameworks keyed to specific policy years. This study addresses all three for Bonny Island. A six-parameter Coastal Vulnerability Index was computed, and three classifiers (Random Forest, Gradient Boosting, Multilayer Perceptron) were trained on both the internally-derived CVI risk class and an independent flood-frequency target from the JRC Global Surface Water occurrence band, 1984–2021, with performance reported as Cohen's kappa. Projections under SSP5–8.5 with Niger-Delta subsidence spanned the 2030, 2040, 2050 and 2100 horizons, and exposure integrated WorldPop 2020, Google Open Buildings v3 and OpenStreetMap roads. The CVI surface places 38.3% of Bonny Island (81.80 km²) in the High and Very-High classes. Random Forest reached 0.98 internal accuracy against the CVI target but only 0.564 Cohen's kappa (macro-F1 0.747) against the independent flood-frequency target. This gap, substantial agreement, yet far below the internal score exposes the optimism of self-validated models, while the 0.564 remains competitive with international benchmarks. High-vulnerability corridors hold about 132,374 residents, 17,803 buildings and 282.5 km of road. This gap between internal and independent scores matters because most published vulnerability models report only the former, overstating reliability. For Bonny Island it yields a 2040 Planning Horizon Package across Environment, Life and Property dimensions, aligned with Sendai Framework Targets A–D and directly actionable by Bonny LGA and federal partners. Globally, the open, zero-cost pipeline is transferable to any data-poor coastal region needing defensible, independently validated risk surfaces.