Abstract
Accurate forecasting of aquatic production is critical for sustainable fisheries management. In this study, four neural network models, namely Back Propagation (BP) neural network, BP neural networks optimized by Genetic Algorithms (GA-BP), Long Short-Term Memory neural networks (LSTM), and Radial Basis Function neural networks (RBF), are developed and compared to predict aquatic production in Zhanjiang City, China. First, key influencing factors are identified through Grey Relational Analysis (GRA), including GDP per capita, sunshine duration, and Engel coefficient. The models are trained and tested using historical production data, with performance evaluated by R² and MAE metrics. Results show that the RBF neural network achieves the highest prediction accuracy ( R ²=0.96, MAE=27725), significantly outperforming BP ( R ²=0.73), GA-BP ( R ²=0.93), and LSTM ( R ²=0.94). Sensitivity analysis is then conducted to rank the influencing factors by importance. GDP per capita is found to be the most critical factor, followed by climate-related variables (sunshine duration, temperature) and socioeconomic indicators (Engel coefficient, consumer price index). The robustness of the RBF model suggests that it can be effectively applied for regional aquatic production forecasting, supporting policymakers in resource allocation and risk mitigation. Furthermore, the factor prioritization enables aquaculture practitioners to optimize farming strategies, such as adjusting production scales based on economic and environmental trends. This study not only provides a reliable modeling framework but also highlights the key drivers affecting aquatic production, including economic, climatic, and demographic factors.