Physics constraint Deep Learning based radiative transfer model.

Area of Science:

• Atmospheric Science
• Oceanography
• Remote Sensing

Background:

• Deep learning (DL) models like TensorFlow, Keras, and PyTorch are successful in forward modeling applications.
• Existing DL radiative transfer models for clear-sky ocean conditions struggle to accurately derive physical properties such as the Jacobian.
• The Jacobian is crucial for calculating radiance sensitivities to geophysical parameters, essential for satellite data assimilation and environmental data retrieval.

Purpose of the Study:

• To develop a physics-constrained deep learning radiative transfer model for clear-sky ocean conditions.
• To address the challenge of accurately predicting physical properties, specifically the Jacobian, from DL forward models.
• To ensure the derived DL model retains correct physical principles.

Main Methods:

• Implemented a physics constraint during the deep learning training process for the radiative transfer model.
• Utilized open-source deep learning libraries (e.g., TensorFlow, Keras, PyTorch).
• Focused on clear-sky conditions over oceans.

Main Results:

• The physics-constrained DL model successfully captures radiance sensitivities (Jacobian).
• The model overcomes the issue of multiple solutions fitting forward model results during training.
• Accurate prediction of physical properties required for satellite radiance assimilation is achieved.

Conclusions:

• Introducing a physics constraint into DL training enables the derivation of physically consistent radiative transfer models.
• The developed DL model accurately captures Jacobian, improving its utility for satellite data assimilation and geophysical parameter retrieval.
• This approach enhances the reliability and applicability of deep learning in radiative transfer modeling for Earth observation.