A spatiotemporally separated framework for reconstructing the sources of atmospheric radionuclide releases

Abstract

Determining the source location and release rate are critical tasks when assessing the environmental consequences of atmospheric radionuclide releases, but they remain challenging because of the huge multi-dimensional solution space. We propose a spatiotemporally separated two-step framework that reduces the dimension of the solution space in each step and improves the source reconstruction accuracy. The separation process applies a temporal sliding-window average filter to the observations, thereby reducing the influence of temporal variations in the release rate on the observations and ensuring that the features of the filtered data are dominated by the source location. A machine-learning model is trained to link these features to the source location, enabling independent source-location estimations. The release rate is then determined using the projected alternating minimization with L1 norm and total variation regularization algorithm. This method is validated against the local-scale SCK CEN (Belgian Nuclear Research Centre) 41Ar field experiment and the first release of the continental-scale European Tracer Experiment, for which the lowest source-location errors are 4.52 m and 5.19 km, respectively. This presents higher accuracy and a smaller uncertainty range than the correlation-based and Bayesian methods when estimating the source location. The temporal variations in release rates are accurately reconstructed, and the mean relative errors in the total release are 65.09 % and 72.14 % lower than the Bayesian method for the SCK CEN experiment and the European Tracer Experiment, respectively. A sensitivity study demonstrates the robustness of the proposed method to different hyperparameters. With an appropriate site layout, low error levels can be achieved from only a single observation site or under meteorological errors.