Abstract
Methane is a potent but short-lived greenhouse gas, and reducing its emissions offers one of the fastest pathways to slow near-term warming. In New Zealand, where livestock agriculture accounts for most national emissions, monitoring diffuse methane sources remains challenging. Satellite instruments such as Sentinel-5P provide broad spatial coverage, but persistent cloud cover, complex terrain, and weak atmospheric signals raise uncertainty about their effectiveness. Atmospheric inversions can integrate satellite data with chemical transport models to estimate surface fluxes, yet their high computational cost limits operational use.
This thesis evaluates whether satellite observations, inversion modelling, and artificial intelligence can together support a robust national methane-monitoring system. It presents the first comprehensive assessment of Sentinel-5P’s capability over New Zealand, develops a high-resolution Bayesian inversion tailored to agricultural landscapes, and introduces a deep-learning emulator that reproduces inversion outputs with significantly reduced computational demand. The results demonstrate that meaningful flux information can be extracted even in weak-signal environments, and that deep-learning models can approximate inversion-like behaviour when trained carefully on physically consistent datasets. Together, these advances establish a scalable framework for national methane accounting and point toward a new generation of agile, data-driven greenhouse gas monitoring systems.