MEQNet: Deep Learning for Methane Point Source Emission Quantification from Sentinel-2 Observations

Mitigating methane emissions is critical for addressing global warming, and accurate point-source quantification is crucial for identifying super-emitters and targeted mitigation. Sentinel-2's fine spatial resolution, global coverage and open accessibility make it highly promising for large-scale monitoring. Current quantification methods, whether retrieval-based physical models or AI models trained on simulated data, are computationally costly, make restrictive assumptions, generalize poorly to real imagery, and are sensitive to surface reflectance, leading to unreliable emission estimates. We propose an end-to-end \textit{Methane Emission Quantification Network (MEQNet)} that directly estimates methane column enhancements and emission rates from bi-temporal Sentinel-2 imagery and auxiliary wind data. MEQNet builds a direct mapping from Sentinel-2 reflectance to emissions and rates, further distinguishing dynamic plumes from background interference by exploiting methane-sensitive spectral bands and modeling spectral-temporal differences. Integrating 10-meter wind vectors further enables physically consistent rate estimation by accounting for plume transport dynamics. To enhance model generalizability, we construct a dataset using real Sentinel-2 observations with emissions from hyperspectral measurements and inventories, covering diverse surface and emission types. Experimental results demonstrate that MEQNet enables scalable, rapid, and accurate methane emission quantification across complex surfaces.