Seasonal to multiannual marine ecosystem prediction using a deep learning approach

Marine biogeochemistry governs the flux of climate-active gases at the ocean-atmosphere interface, influencing diverse climate feedbacks. Despite advances in Earth System Models (ESMs) for climate-ecosystem predictions, challenges persist in the initialization and validation with biogeochemical observation data. In this study, Convolutional Neural Network (CNN)-based models predict chlorophyll concentrations in a productive large coastal area. The model was trained and validated using Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model ensemble datasets and physical–biogeochemical reanalysis data from a data assimilative ESM run. Through sensitivity tests on the model structure and input data, the CNN-based model demonstrates physical interpretability consistent with previous studies. Our optimized model adeptly reproduces annual observational chlorophyll variations in coastal regions where dynamic models face challenges, demonstrating comparable prediction skill to dynamic models in seasonal prediction by capturing large-scale climate variabilities. These findings highlight the importance of combining dynamic models and deep learning approaches, offering the potential for more accurate and comprehensive predictions of marine ecosystems.