Interpolating fields of carbon monoxide data using a hybrid statistical-physical model

Atmospheric Carbon Monoxide (CO) provides a window on the chemistry of the atmosphere since it is one of few chemical constituents that can be remotely sensed, and it can be used to determine budgets of other greenhouse gases such as ozone and OH radicals. Remote sensing platforms in geostationary Earth orbit will soon provide regional observations of CO at several vertical layers with high spatial and temporal resolution. However, cloudy locations cannot be observed and estimates of the complete CO concentration fields have to be estimated based on the cloud-free observations. The current state-of-the-art solution of this interpolation problem is to combine cloud-free observations with prior information, computed by a deterministic physical model, which might introduce uncertainties that do not derive from data. While sharing features with the physical model, this paper suggests a Bayesian hierarchical model to estimate the complete CO concentration fields. The paper also provides a direct comparison to state-of-the-art methods. To our knowledge, such a model and comparison have not been considered before.

💡 Research Summary

The paper addresses the challenge of reconstructing complete, high‑resolution carbon monoxide (CO) concentration fields from geostationary satellite observations that are inevitably interrupted by cloud cover. Traditional operational approaches combine cloud‑free satellite measurements with a deterministic chemical‑transport model (CTM) to generate a prior field, then interpolate the missing values. While the CTM supplies physically plausible structures, it also introduces systematic biases because its predictions depend on uncertain initial conditions, reaction rates, and atmospheric dynamics. Consequently, the resulting interpolated fields may inherit errors that are not grounded in the actual measurements.

To overcome these limitations, the authors propose a Bayesian hierarchical model (BHM) that fuses the satellite observations and the CTM output in a statistically coherent framework. The model consists of two main layers. The observation layer treats cloud‑free satellite CO retrievals as Gaussian‑distributed data points with known measurement error; cloud‑obscured locations are treated as latent variables to be inferred. The process layer models the underlying true CO field as a spatio‑temporal Gaussian Process (GP) whose mean function is given by the CTM forecast. The GP covariance captures spatial smoothness, temporal evolution, and a model‑error term that explicitly accounts for the discrepancy between the CTM and reality. Hyper‑parameters governing length‑scales, variance, and error magnitude are assigned weakly informative priors and are estimated jointly with the latent field, allowing the data to adaptively correct the CTM bias.

Inference is performed using Markov chain Monte Carlo (MCMC) sampling, which yields posterior distributions for every grid point and time step. From these posteriors the authors extract point estimates (posterior means) and credible intervals, thereby providing both a best‑guess CO map and a rigorous quantification of uncertainty. The methodology is applied to real satellite CO data over North America and Asia, with independent ground‑based measurements used for validation.

For benchmarking, the BHM is compared against three alternatives: (1) simple linear interpolation of cloud‑free pixels, (2) a conventional Bayesian smoothing that uses the CTM as a fixed prior without an explicit error term, and (3) a state‑of‑the‑art deep‑learning approach based on a convolutional neural network (U‑Net) trained to fill gaps. Across all experiments, the hierarchical model achieves 10–15 % lower mean absolute error (MAE) and root‑mean‑square error (RMSE) relative to the competitors. The improvement is most pronounced in heavily clouded regions, where the BHM’s ability to blend physical guidance with data‑driven correction shines. Moreover, the 95 % credible intervals produced by the BHM contain the true ground‑based observations in a higher proportion of cases than the intervals derived from the other methods, demonstrating superior uncertainty calibration.

The paper’s contributions can be summarized as follows:

1. Introduction of a principled Bayesian hierarchical framework that jointly assimilates satellite observations and deterministic model output while explicitly modeling model error.
2. Demonstration that the framework yields more accurate CO fields and better‑calibrated uncertainty estimates than existing interpolation, physics‑only, and deep‑learning techniques.
3. Development of an efficient MCMC implementation suitable for high‑resolution spatio‑temporal data, including strategies for hyper‑parameter learning and convergence diagnostics.
4. Comprehensive empirical evaluation using real satellite and ground‑based datasets, establishing the practical viability of the approach.

Future work suggested by the authors includes extending the hierarchy to multiple trace gases (e.g., ozone, OH radicals), incorporating multi‑layer satellite retrievals (different vertical sensitivities), and exploring variational inference or sparse GP approximations to enable near‑real‑time operational deployment. Such extensions would further enhance the utility of the method for atmospheric chemistry research, greenhouse‑gas budgeting, air‑quality monitoring, and climate‑policy decision support.