Sensitivity of a bayesian source-term estimation model to spatiotemporal sensor resolution

Abstract

Abstract Source-term estimation (STE) methods attempt to calculate the most-likely source characteristics of an atmospheric release given concentration observations. The quality of the STE depends partially on the time and space scales of the observations, sensor locations, and release parameters. We quantify the sensitivity of a probabilistic STE algorithm that was previously validated using observational data collected during a controlled tracer release experiment to the spatiotemporal resolution of observing networks. We have also made many improvements to the STE algorithm, which extends applicability to coarser-resolution observational datasets. Improvements include the use of a fully-connected deep neural network model emulator with dynamically optimized architecture and better goodness-of-fit (GoF) metrics to measure the discrepancy between model and observational data. The GoF metrics, which are more robust and resilient than previous metrics, are the Spearman rank-based correlation coefficient and a variation of the binary f 1 classification score adapted for regression problems. Using synthetically generated observations over East Asia, the skill of the improved algorithm is quantified over a broad range of sensor configurations and release scenarios. The evaluation is broken into three experiments. First, a validation study shows that the proposed GoF metrics perform more reliably than other possible candidates. Next, data-denial techniques are applied to a single release scenario over the Korean peninsula, where the skill of the inversion is shown to be highly sensitive to the number and location of deployed sensors but less sensitive to temporal resolution. Finally, the STE algorithm is tested for many release locations throughout the geographic model domain, where the STE algorithm performs well for all but a few cases. The results indicate that the STE algorithm provides informative source-parameter posterior probability distributions utilizing data collected by sparse sensor networks; however, the skill of the STE algorithm improves significantly for higher resolution networks.