Deep Learning Accelerates CO2 Retrieval from Satellite Data with Better Uncertainty

Monitoring atmospheric carbon dioxide (CO2) from space is critical for understanding the global carbon cycle. NASA's Orbiting Carbon Observatory-2 (OCO-2) collects high-resolution spectra to estimate CO2 levels, but current retrieval algorithms are computationally intensive and often fail to properly quantify uncertainties. A new deep learning framework has been developed to address these limitations. It is trained and validated using a high-fidelity simulation dataset that incorporates realistic forward model errors, crucial for robust uncertainty quantification. The architecture employs a multi-branch neural network to encode spectral bands. The framework estimates posterior distributions of CO2 using scalable uncertainty quantification methods: Laplace approximations and normalizing flows. This approach offers several advantages: orders of magnitude faster inference for real-time data processing, robustness to model errors, superior predictive accuracy, better-calibrated uncertainty estimates, and the ability to model complex, non-Gaussian posterior distributions. These advancements pave the way for next-generation operational processing systems.