Uncertainty assessment in satellite-based greenhouse gas emissions estimates using emulated atmospheric transport

TL;DR

The paper tackles the challenge of quantifying transport-model uncertainty in top-down greenhouse gas flux estimation. It introduces an ensemble-based, graph neural network emulator (GATES) of a Lagrangian Particle Dispersion Model to rapidly generate transport footprints and methane mole fractions, enabling uncertainty-aware inferences at continental scales. Demonstrated on GOSAT methane data for Brazil in 2016, the approach achieves ~1000× speed-up relative to NAME LPDM while preserving major footprint structures, and shows that ensemble spread correlates with emulation error, highlighting spatial and temporal patterns of low-confidence predictions. This method offers a scalable path toward robust satellite-based GHG monitoring and could be integrated into Bayesian inversion frameworks to improve the reliability of top-down flux estimates across regions and species.

Abstract

Monitoring greenhouse gas emissions and evaluating national inventories require efficient, scalable, and reliable inference methods. Top-down approaches, combined with recent advances in satellite observations, provide new opportunities to evaluate emissions at continental and global scales. However, transport models used in these methods remain a key source of uncertainty: they are computationally expensive to run at scale, and their uncertainty is difficult to characterise. Artificial intelligence offers a dual opportunity to accelerate transport simulations and to quantify their associated uncertainty. We present an ensemble-based pipeline for estimating atmospheric transport "footprints", greenhouse gas mole fraction measurements, and their uncertainties using a graph neural network emulator of a Lagrangian Particle Dispersion Model (LPDM). The approach is demonstrated with GOSAT (Greenhouse Gases Observing Satellite) observations for Brazil in 2016. The emulator achieved a ~1000x speed-up over the NAME LPDM, while reproducing large-scale footprint structures. Ensembles were calculated to quantify absolute and relative uncertainty, revealing spatial correlations with prediction error. The results show that ensemble spread highlights low-confidence spatial and temporal predictions for both atmospheric transport footprints and methane mole fractions. While demonstrated here for an LPDM emulator, the approach could be applied more generally to atmospheric transport models, supporting uncertainty-aware greenhouse gas inversion systems and improving the robustness of satellite-based emissions monitoring. With further development, ensemble-based emulators could also help explore systematic LPDM errors, offering a computationally efficient pathway towards a more comprehensive uncertainty budget in greenhouse gas flux estimates.

Figures (6)

• Figure 1: Top row: predicted atmospheric transport footprints generated by the four GATES (LPDM emulator) models for the same randomly selected time point from the test set. Bottom row: A comparison against the true LPDM footprint (e) with the GATES emulator ensemble mean prediction (f), normalised mean absolute error (NMAE) between the two (g), and coefficient of variation (CV) of predictions (h).
• Figure 2: Mean performance during GATES LPDM emulator model development across standard machine learning metrics. Panels b-f present performance on the test set. NMAE = Normalised mean absolute error. MSE = Mean squared error. IoU = Intersection over union. R$^2$ = coefficient of determination. Error bars represent the standard deviation across the four trained models.
• Figure 3: Left: temporal coefficient of variation of the mean prediction over the entire test set. Right: wind rose for surface-level winds, centred around the release point per footprint of the test set.
• Figure 4: Spatial maps of mean footprint sensitivities across South America. a) LPDM-generated, b) GATES-predicted, c) standard deviation of predictions, and d) coefficient of variation across the four models.
• Figure 5: Relative uncertainty for predicted mole fractions across the first 200 time points of the test set, compared against mole fractions using LPDM footprints. Both are calculated using the bottom-up flux field.