Advancing Point Source Methane Emissions Quantification With UAV Sampling and Deep Learning: A Large-Eddy Simulation Study

Quantifying point-source methane emissions from atmospheric plume observations is essential for climate change mitigation but remains challenged by turbulence-induced noise. Here, we introduce a deep learning framework that integrates unmanned aerial vehicle (UAV) plume sampling with turbulence-invariant analysis to overcome this limitation. Using large-eddy simulations (LESs) spanning a wide range of atmospheric stability conditions, we show that conventional inversion methods (including the inverse Gaussian and mass balance methods) yield mean absolute percentage errors (MAPEs) of 27%–46%, primarily caused by turbulence-induced plume meandering. Furthermore, we identified the ratio of UAV speed to plume centroid speed as the dominant factor controlling this error, with lower ratios yielding lower MAPE. A denser sampling spacing and multiple consecutive flights can reduce the multi-condition mean MAPE to 25%–27%, but at the cost of approximately doubling the required time and resources. Our approach employs a U-Net model to map sparse UAV observations to time-averaged plume cross-sections, effectively reconstructing plume morphology and mitigating turbulence-induced bias. When integrated into an inverse Gaussian framework, our U-Net model reduces the MAPE from 30% to 22% in LES data, and from 38% to 29% in controlled-release experiments. This work establishes a robust, physics-informed strategy for accurate and scalable methane monitoring, highlighting the potential of deep learning to enhance greenhouse gas accountability.