Source Reconstruction of Atmospheric Releases by Bayesian Inference and the Backward Atmospheric Dispersion Model: An Application to ETEX-I Data

Abstract

Source term reconstruction methods attempt to calculate the most likely source parameters of an atmospheric release given measurements, including both location and release amount. However, source term reconstruction is vulnerable to uncertainties. In this paper, a method combining Bayesian inference with the backward atmospheric dispersion model is developed for robust source term reconstruction. The backward model is used to quantify the relationship between the source and measurements and to reduce the search range of the Bayesian inference. A Markov chain Monte Carlo method is used to sample from the multidimensional parameter space of the source term. The source location and release rate are estimated simultaneously, and the posterior probability distribution is produced by applying Bayes’ theorem. The proposed method is applied to a set of real concentration data from the ETEX-I experiment. The results demonstrate that the source location is estimated to be −2.86° ± 1.01°E, 48.25° ± 0.33°N, and the release rate is estimated to be 20.16 ± 3.56 kg/h. The true source location is correctly estimated to be within a one standard deviation interval, and the release rate is correctly determined to be within a three standard deviation interval.

Generalized spatiotemporally-decoupled framework for reconstructing the source of non-constant atmospheric radionuclide releases

“…This assumes that the modelobservation discrepancies follow a certain statistical distribution (i.e. the likelihood of Bayesian methods), with the normal (Eslinger and Schrom, 2016;Guo et al, 2009;Keats et al, 2007Keats et al, , 2010Rajaona et al, 2015;Xue et al, 2017a, b;Yee, 2017;Yee et al, 2008;Zhao et al, 2021) and log-normal (Chow et al, 2008;Dumont Le Brazidec et al, 2020;KIM et al, 2011;Monache et al, 2008;Saunier et al, 2019;Senocak, 2010;Senocak et al, 2008) distributions being two popular choices. Other candidate distributions include Cauchy, log-Cauchy, and T3-10, which have been compared with normal and log-normal distributions in reconstructing the source parameters of the Prairie Grass field experiment (Wang et al, 2017).…”

“…Deterministic assumptions mainly involve entropy (Krysta and Bocquet, 2007;Bocquet, 2005b, a) and a constant release rate (Kovalets et al, 2020(Kovalets et al, , 2018Efthimiou et al, 2018Efthimiou et al, , 2017Tomas et al, 2021;Andronopoulos and Kovalets, 2021;Ma et al, 2018). Compared with entropy, the constant-release assumption is more popular and is embedded in many Bayesian methods (Yee et al, 2008;Eslinger and Schrom, 2016;Meutter and Hoffman, 2020;Zhao et al, 2021;De Meutter et al, 2021), substantially reducing the dimension of the solution space to 5 or 6 (i.e. the two or three source location coordinates, the start and end time of the release, and the total release).…”