PREDICTIVE MODELLING OF MARINE SENSOR DATA FOR SUSTAINABLE FISHERIES MANAGEMENT IN THE GULF OF GUINEA

Overfishing and environmental change are squeezing the fisheries of the Gulf of Guinea, and the people who depend on them, harder every year. This study asks a practical question: can data streaming off marine sensors be turned into forecasts good enough to actually guide management decisions? Drawing on sixty months of in-situ and satellite-derived observations (sea-surface temperature, salinity, chlorophyll-a, dissolved oxygen and turbidity) collected across the coastal waters of Nigeria, Ghana and Côte d'Ivoire, we built and compared four predictive models for catch-per-unit-effort (CPUE): support vector regression, random forest, gradient-boosted trees (XGBoost) and a long short-term memory (LSTM) neural network. The LSTM came out ahead (R² = 0.91, RMSE = 1.62 tonnes), with XGBoost close behind and far cheaper to train. Chlorophyll-a concentration and sea-surface temperature dominated the feature-importance rankings, which lines up with what marine ecologists would expect about primary productivity driving fish abundance. We argue that these models are accurate enough to support near-real-time, adaptive management, and we set out what it would take to deploy them responsibly in a data-scarce, multi-jurisdiction setting