Mapping U.S. Surface Methane With Artificial Neural Networks

Ground-level methane (CH4) concentrations are crucial for improving emission inventories and supporting climate and air quality management efforts. However, the limited spatial coverage of ground-based CH4 monitoring stations across the U.S. constrains detailed spatial and temporal analyses. To address this challenge, we develop artificial neural network (ANN) models to estimate daily surface CH4 concentrations across the U.S. The models use meteorological parameters from the European Centre of Medium-range Weather Forecasts Reanalysis v5 (ERA5) and CH4 column concentrations from the Sentinel-5P TROPOspheric Monitoring Instrument (TROPOMI) as inputs. Surface CH4 measurements from NOAA’s Global Monitoring Laboratory (GML) serve as the reference outputs for model training and evaluation. We implement a feed-forward backpropagation neural network with a hyperbolic tangent sigmoid activation function and evaluate three optimization algorithms, including Bayesian Regularization (BR), Levenberg–Marquardt (LM), and Scaled Conjugate Gradient (SCG), combined with two sets of input parameters. Better results are obtained using the input set that includes 2-m air temperature, soil temperature, u- and v-wind components, total precipitation, net solar radiation, and CH4 column concentrations. With this set, the LM algorithm achieves the highest accuracy (RMSE = 11.31 ppb, R = 0.84), outperforming BR (RMSE = 16.47 ppb, R = 0.79) and SCG (RMSE = 22.57 ppb, R = 0.76). These findings demonstrate the potential of combining satellite-based CH4 data and meteorological variables in machine learning models for surface CH4 estimation and provide guidance for algorithm selection in future applications. Artificial Neural Network (ANN)-Based Estimation of Surface Methane Concentrations over the U.S. This graphical abstract summarizes the application of Artificial Neural Networks (ANNs) to estimate surface methane (CH4) concentrations across the U.S., addressing gaps in sparse ground-based monitoring. ERA5 meteorological data and Sentinel-5P TROPOMI CH4 columns are integrated to capture nonlinear relationships between atmospheric drivers and CH4 distribution. Validation against surface CH4 concentrations from NOAA-GML stations shows that key parameters in Set B, including 2-m temperature, soil temperature, u- and v-wind components, total precipitation, net solar radiation, and CH4 column concentrations, used with the Levenberg–Marquardt (LM) algorithm, achieve the best accuracy (RMSE = 11.31 ppb, R = 0.84), compared with the Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) algorithms. These findings demonstrate that optimized ANN approaches, trained with satellite and reanalysis data, can reasonably reproduce surface CH4 patterns and enhance greenhouse gas monitoring in under-observed regions.