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

Area of Science:

• Oceanography
• Marine Biogeochemistry
• Artificial Intelligence in Environmental Science

Background:

• Accurate prediction of ocean chlorophyll-a vertical distribution is crucial for marine ecosystem structure and carbon cycling.
• Traditional methods like Argo floats and remote sensing have limitations in vertical resolution and cloud interference.
• Existing physical and deep learning models struggle with robustness in complex marine environments due to idealized assumptions or data requirements.

Purpose of the Study:

• To develop and validate a hybrid framework for predicting ocean chlorophyll-a vertical distribution.
• To integrate an Active-Passive Fusion Water Column Optical Profile model with deep learning architectures.
• To enhance predictive accuracy using a Bayesian Model Combination approach.

Main Methods:

• Implementation of an Active-Passive Fusion Water Column Optical Profile model.
• Application of deep learning architectures: Transformer, long short-term memory (LSTM), and convolutional neural network (CNN).
• Incorporation of a Bayesian Model Combination strategy to integrate multiple model predictions.

Main Results:

• The Active-Passive Fusion model produced chlorophyll-a profiles consistent with optical properties and validation data (MAE: 0.023–0.207 mg/m³).
• Deep learning models demonstrated high predictive power (R² up to 0.97), effectively capturing vertical chlorophyll-a variability.
• Bayesian Model Combination improved predictions, reducing MAE in the Indian Ocean from 0.0326 to 0.0289 mg/m³.

Conclusions:

• The proposed hybrid framework offers a robust and reliable method for predicting ocean chlorophyll-a vertical distribution.
• The integration of optical profiling, deep learning, and Bayesian methods overcomes limitations of traditional approaches.
• This framework provides accurate, physically interpretable data essential for marine ecosystem and carbon cycle research.