Efficient Computational Statistical Approaches for Local and Regional Flux Inversions

<p dir="ltr">Flux inversion is a statistical procedure through which sources and sinks of a gas are estimated from observations of atmospheric composition, in the form of gas mole fraction. It often involves running a Lagrangian particle dispersion model (LPDM) to generate sensitivities between observations and fluxes over a spatial domain. This can be computationally prohibitive for large areas or long data records. In this thesis, we explore several ways in which the associated computational time can be improved. For flux inversions over regional scales, we develop a novel spatiotemporal emulator for LPDM sensitivities, built using a convolutional variational autoencoder (CVAE), and a spatio-temporal Gaussian Process emulator. We show that our CVAE-based emulator outperforms the more traditional emulator built using empirical orthogonal functions, and that it can be used with different LPDMs. For local scale flux inversions, we explore the use of a surrogate model in place of the full-blown LPDM, namely, the Gaussian plume dispersion model. We propose integrating the Gaussian plume model within a fully Bayesian framework, to account for uncertainty in the measurements, the meteorological data, and the atmospheric model itself. Using data collected from a controlled-release experiment, we show that the inversion framework is robust to instrument type and meteorological conditions. The worst median emission-rate estimate we obtain from all the inversions is within 36% of the true value, while the worst posterior 95% credible interval has a limit within 11% of the true value. We conclude that the use of emulators or surrogate models in place of the full-blown LPDM can reliably reduce the computing time needed for high-resolution flux inversions.</p>