Methane is one of the most important anthropogenic greenhouse gases with a significant impact on the Earth's radiation budget and tropospheric background ozone. Despite a well-constrained global budget, quantification of local and regional methane emissions has proven challenging. Recent advancements in airborne remote sensing instruments such as from the next-generation Airborne Visible/Infrared Imaging Spectrometer (AVIRIS-NG) provide 2-D observations of CH₄ plume column enhancements at an unprecedented resolution of 1–5 m over large geographic areas. Quantifying an emission rate from observed plumes is a critical step for understanding local emission distributions and prioritizing mitigation efforts. However, there exists no method that can predict emission rates from detected plumes in real-time without ancillary data reliably. In order to predict methane point-source emissions directly from high resolution 2-D plume images without relying on other local measurements such as background wind speeds, we trained a convolutional neural network model called MethaNet. The training data was derived from large eddy simulations of methane plumes and realistic measurement noise over agricultural, desert and urban environments. Our model has a mean absolute percentage error for predicting unseen plumes under 17%, a significant improvement from previous methods that require wind information. Using MethaNet, a validation against a natural gas controlled-release experiment agrees to within the precision error estimate. Our results support the basis for the applicability of using deep learning techniques to quantify CH₄ point sources in an automated manner over large geographical areas, not only for present and future airborne field campaigns but also for upcoming space-based observations in this decade.

甲烷是最重要的人为温室气体之一，对地球辐射收支和对流层背景臭氧有重大影响。尽管全球预算受到严格限制，但当地和区域甲烷排放的量化已被证明具有挑战性。机载遥感仪器的最新进展，例如下一代机载可见光/红外成像光谱仪 (AVIRIS-NG)，可提供 CH ₄的二维观测在大地理区域内以前所未有的 1-5 m 分辨率增强羽流柱。从观测到的羽流中量化排放率是了解当地排放分布和确定缓解工作优先级的关键步骤。然而，没有任何方法可以在没有辅助数据的情况下实时预测检测到的羽流的排放率。为了在不依赖其他局部测量值（如背景风速）的情况下直接从高分辨率 2-D 羽流图像预测甲烷点源排放，我们训练了一个称为 MethaNet 的卷积神经网络模型。训练数据来自对甲烷羽流的大型涡流模拟和农业、沙漠和城市环境中的真实测量噪声。我们的模型预测未见羽流的平均绝对百分比误差低于 17%，与之前需要风信息的方法相比有了显着改进。使用 MethaNet，对天然气控释实验的验证在精确误差估计范围内一致。我们的结果支持使用深度学习技术量化 CH 的适用性基础_{4 个}点源以自动化方式覆盖大地理区域，不仅适用于当前和未来的机载野外活动，也适用于本十年即将进行的天基观测。