Intuitively tuned elastic bias correction of atmospheric inversion using Gaussian process prior: Application to accidental radioactive emissions

Precise estimation of atmospheric pollutant releases is crucial for assessing the impact of environmental accidents. Atmospheric inversion typically relies on a linear model with a source-receptor sensitivity (SRS) matrix, which may contain significant errors or even completely fail to capture the real magnitude of the event. We propose a correction of the SRS matrix formulated as slight shifts in the observation locations, effectively warping the sensitivity field. To constrain these shifts and ensure data-driven corrections, we model them using a Gaussian process prior. This prior not only enforces smoothness and sparsity, but also enables posterior prediction of shifts at previously unseen locations. This key feature provides a mechanism for hyper-parameter tuning: the predicted shift field can be visualized on a map and assessed by an expert. We present a user-friendly framework that combines a Bayesian inversion model with correction and a tuning algorithm based on L-curve-like plots and the maps of predicted shifts. The proposed method is demonstrated on three case studies: the ETEX-I experiment, the emissions during the 2020 Chernobyl wildfires, and the release in 2017.