ANN-ENHANCED ARIMA MODELS FOR SST-BASED SEASONAL FISH-CATCH FORECASTING AND DECISION-SUPPORT IN BENGKULU WATERS, INDONESIA

This study presents a comparative evaluation of hybrid ARIMA-family forecasting models combined with Artificial Neural Networks (ANNs) for seasonal fish-catch prediction in Bengkulu waters, Indonesia, while assessing the role of sea surface temperature (SST) as an environmental predictor. Monthly SST data from NASA’s Giovanni portal and pelagic fish-catch records collected between January 2017 and June 2025 were used to develop ARIMA, ARIMAX, SARIMA, and SARIMAX models, whose residuals were subsequently modeled using Feedforward Neural Networks (FFNN) and Long Short-Term Memory (LSTM) networks to capture nonlinear temporal dependencies. Among the evaluated models, the hybrid ARIMA–LSTM achieved the highest forecasting accuracy on the available dataset, with an RMSE of 76.779 and a MAPE of 19.223%, whereas hybrid models that explicitly incorporate SST as a linear exogenous predictor showed lower predictive performance. These findings suggest that although SST remains an ecologically important environmental driver of pelagic fisheries, its predictive contribution may be better captured by nonlinear, lag-dependent relationships embedded in historical fish-catch observations rather than by contemporaneous linear exogenous modeling. Overall, this study provides empirical evidence for selecting appropriate hybrid forecasting models under practical fisheries data conditions and highlights their potential application as analytical components within fisheries Decision Support Systems (DSS) for adaptive fisheries management.