Applying Gaussian Process Machine Learning and Modern Probabilistic Programming to Satellite Data to Infer CO₂...

Area of Science:

• Earth and Atmospheric Sciences
• Machine Learning
• Environmental Monitoring

Background:

• Satellite observations are crucial for understanding CO2 distribution, but current models struggle with spatial-temporal data correlations.
• Accurate estimation of spatial length scales in atmospheric inverse models remains a challenge.

Purpose of the Study:

• To develop an advanced inference model using Gaussian Process (GP) machine learning (ML) to process spatiotemporal covariance in satellite CO2 data.
• To estimate hyperparameters like covariance length scales and improve atmospheric inversions of emissions.
• To integrate GP ML with probabilistic programming languages (PPLs) and the GEOS-Chem model for enhanced CO2 emission analysis.

Main Methods:

• Developed a GP ML inversion system utilizing modern PPLs and the GEOS-Chem chemical transport model.
• Simulated atmospheric CO2 concentrations using Orbiting Carbon Observatory-2/3 (OCO-2/3) data for July 2020 within a supervised learning framework.
• Employed sector-specific emission scaling factors as hidden predictors for the GP model.

Main Results:

• The GP model, accelerated by GPU-enabled PPLs, successfully retrieved true emission scaling factors.
• The system effectively inferred hidden noise levels within the satellite CO2 concentration data.
• Demonstrated the model's capability to handle complex covariance structures in spatiotemporal data.

Conclusions:

• The proposed GP ML approach enhances the analysis of satellite CO2 data by accurately processing spatiotemporal correlations.
• This method offers a robust framework for atmospheric inversions, improving the estimation of CO2 emissions.
• The approach is scalable for larger regions and complex datasets, advancing the analysis of OCO-2/3 and similar satellite missions.