Estimating Tropospheric Methane from Cross-track Infrared Sounder (CrIS) Spectra using a Machine Learning Method

The Cross-track Infrared Sounder (CrIS) on the Joint Polar Satellite System (JPSS) satellites is a Fourier transform spectrometer, providing sounding (temperature and humidity) and trace gas the of the atmosphere with 2221 spectral channels along three infrared bands. This paper presents a Machine Learning (ML) method to retrieve the methane in the middle troposphere from CrIS spectra. The spectral signature of a target trace is a complex function of the methane and the other parameters, such as the state of the atmosphere (atmospheric temperature and humidity profiles) and surface (skin temperature and emissivity), interfering trace gas, and the viewing geometry. Different from traditional physical retrieval methods, the main idea of the ML-based approach is to use a neural network (NN) to approximate the complex inverse function that maps the target trace gas concentration to the CrIS radiances measured based on the training datasets.

As the first step, the authors utilize the Thermodynamic Initial Guess Retrieval (TIGR) dataset coupling with a three-dimensional chemical-transport model outputs (e.g., CAMS global greenhouse gas reanalysis) to build the training datasets through radiative transfer model calculations. The training dataset will cover a large range not only of the concentration of the target trace gas but also of the auxiliary parameters on the state of the atmosphere and surface. Competitive networks such as generative adversarial networks (GAN) in which multiple networks (of varying structure) compete will be trained to finally determine model parameters (weights for each node) based on the training datasets. Finally, we will use the real CrIS spectral as inputs to estimate the tropospheric methene from the trained neural network models. The retrieved methane will be compared with the existing sounding products from a physical method (e.g., NUCPAS ) as well as those measured from ground and aircraft to further evaluate the proposed ML-based retrieval method.