Quantifying methane point-source emissions with hyper-spectral imagery and the deep learning model

Methane (CH4), a potent greenhouse gas with 28–36 times the global warming potential of CO2 over a 100-year period, is critical to climate mitigation, yet its point-source quantification remains challenging due to spatial-temporal variability, surface interference, and reliance on auxiliary data (e.g., wind speed). This study proposes a deep learning framework integrated with high-fidelity hyperspectral simulation to address these limitations. Synthetic Airborne Visible/Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) data were generated by coupling the Weather Research and Forecasting model in Large-Eddy Simulation mode (WRF-LES) (for plume dynamics) and MODTRAN (for radiative transfer), covering eight surface types, four albedo levels, and emission rates of 500–2000 kg/h. A multitask deep learning model (GasNet) was developed, including UNet for plume mask detection, ResNet for methane concentration retrieval, and EfficientNet-V2L for emission rate estimation. Additionally, a Dynamically Regularized Adaptive Matched Filter (DRAMF) was integrated to enhance robustness against surface interference. Results show GasNet achieves a root-mean-square error (RMSE) of 2101 kg/h, mean absolute percentage error (MAPE) of 10.05%, and Pearson R of 95.70% for emission quantification—outperforming traditional methods like Integrated Mass Enhancement (IME) by 1287 kg/h (RMSE reduction) and 3.92% (MAPE reduction). DRAMF improves the true positive rate (TP) to 81.4% in low-albedo scenarios, 128% higher than the albedo-corrected reweighted-L1-matched filter (ACRWL1MF). Validation on real PRISMA and Sentinel-2 data confirms the model’s ability to detect plumes with emission rates of 1112–7615 kg/h. This framework provides an automated, wind-speed-independent solution for high-precision methane point-source monitoring, supporting global climate mitigation efforts.