CH4Vision: Machine Learning Estimation of Methane Flux with GaoFen-5 Hyperspectral Imagery

Accurate detection and quantification of point-source methane (CH₄) emissions are essential for methane budget assessment and effective environmental management.The widely used integrated mass enhancement (IME) method suffers from substantial uncertainties due to its simplified linear assumptions linking methane mass enhancements to wind fields. To address this limitation, we proposed CH4Vision, a machine learning framework designed to estimate methane fluxes from satellite observations by leveraging plume morphology features. To train and validate the model, we generated a series of synthetic CH₄ plumes, each corresponding to a distinct emission rate, using large-eddy simulations. These simulated plumes were embedded into shortwave infrared hyperspectral scenes acquired by the Advanced Hyperspectral Imager onboard China’s GaoFen-5A/B satellites. Methane enhancements were then retrieved to identify plumes, from which a comprehensive set of plume-morphology metrics was extracted for emission flux modeling. Subsequently, we trained a random forest regression model to relate these morphological features and along-wind speed to the known emission rates. Validation indicates that CH4Vision outperforms IME, achieving a 3% to 9% improvement in R² and a 14% to 36.5% reduction in prediction error. Importantly, CH4Vision exhibits greater robustness to errors in methane enhancement retrievals, wind speed simulations, and plume segmentation. Additional validation against ground-based controlled-release experiments in Arizona, USA, demonstrated that predictions remained within ±100 kg/h of the true flux. When applied to Shanxi Province, China, CH4Vision identified nearly 150 coal-related sources emitting over 3,000 kg/h, whereas IME underestimated total emissions by 15.5% compared with our framework. Overall, these results establish CH4Vision as a robust and scalable solution for global methane emission quantification, with marked potential for application to other trace gases.