Integrating physics-based modeling and deep learning for high-resolution vertical chlorophyll-a predictions in the ocean.

Accurate prediction of the vertical distribution of ocean chlorophyll-a concentration is essential for understanding marine ecosystem structure and carbon cycling. Traditional observation methods, including biogeochemical Argo floats and passive remote sensing, are limited by low vertical resolution and cloud interference. Purely physical models rely on idealized assumptions, while deep learning models require large datasets, making both approaches less robust in complex environments. This study proposes a hybrid framework that combines an Active-Passive Fusion Water Column Optical Profile model with deep learning architectures, including Transformer, long short-term memory, and convolutional neural network. A Bayesian Model Combination approach is further incorporated to enhance predictive performance. Results indicate that chlorophyll-a concentration profiles generated by the Active-Passive Fusion Water Column Optical Profile model align well with water optical properties and validation data, with mean absolute errors (MAE) ranging from 0.023 to 0.207 mg/m3, providing accurate and physically interpretable training data. Deep learning models achieve high predictive performance, with coefficients of determination reaching up to 0.97, effectively capturing the vertical variability of chlorophyll-a concentration. The Bayesian Model Combination strategy further improves predictions across oceanic regions. In subregion a of the Indian Ocean, MAE decreased from 0.0326 mg/m3 for a single model to 0.0289 mg/m3, demonstrating the robustness and reliability of the proposed framework in complex marine conditions.