Integrating Mesoscale Eddies Into NARX Neural Networks for Improved Chlorophyll-A Forecasting From Remote Sensing Data in the South China Sea

Marine ecosystem forecasting necessitates advanced mining of multiscale spatiotemporal patterns from satellite data, particularly in resolving nonlinear couplings between surface dynamics and biogeochemical processes. This study integrates mesoscale eddy features—derived from the META3.2 altimetry dataset—into a Nonlinear AutoRegressive with Exogenous Inputs (NARX) neural network to improve chlorophyll-a (Chl-a) forecasting in the deep South China Sea. Leveraging 11 years (2010–2020) of remote sensing data (SST, SSH, wind stress curl, Chl-a), we developed a hybrid data mining framework that synergizes surface parameter sequences with eddy attributes (amplitude, speed, lifetime). The NARX model reduced 7–15-day Chl-a forecasting errors by 26–38% compared to eddy-excluded baselines (p < 0.05), with peak performance in cyclonic eddy zones where nutrient upwelling dominates. By mining over 7 million data points, we demonstrate that eddy data incorporation enhances prediction accuracy through physics-informed feature engineering, outperforming conventional satellite-only approaches. This work establishes a paradigm for multivariate remote sensing data synthesis, bridging oceanographic process understanding with AI-driven forecasting in data-sparse deep-sea environments.