Physics-Constrained Bayesian Neural Networks for Aerosol Retrieval From Hyperspectral Satellite Measurements With Integrated Uncertainty Quantification

This is the authors' abstract. We don't add key points for this paper.

This study introduces an innovative operational Bayesian neural network framework for high-precision joint retrieval of aerosol optical depth (AOD) and layer height (ALH) with physically-consistent uncertainty decomposition from TROPOMI hyperspectral measurements. Unlike conventional approaches, three different full-physics Bayesian neural network architectures (implemented via Bayes-by-Backprop, Dropout, and Batch Norm techniques) are developed to simultaneously estimate target parameters and their heteroscedastic aleatoric uncertainties while preserving radiative transfer constraints. Epistemic uncertainties are quantified via Monte Carlo sampling of stochastic forward propagation, enabling systematic separation of data-driven vs. model-driven uncertainties. A comprehensive validation demonstrates: (1) Synthetic experiments show epistemic uncertainties strongly correlate with retrieval errors, particularly for observing geometries outside the training data distribution, outperforming aleatoric estimates; (2) Analyses using TROPOMI measurements demonstrate that the framework delivers comparable accuracy to operational products while providing unique uncertainty diagnostics. The framework’s computational efficiency combined with its probabilistic outputs establishes a new paradigm for characterizing aerosol properties from satellite measurements, particularly valuable for climate and air quality applications.