Air mass factor calculation using deep neural network technique for tropospheric NO₂ retrieval from space

We performed a feasibility study on using deep neural network (DNN) techniques to calculate the air mass factor (AMF) for the satellite remote sensing of tropospheric nitrogen dioxide NO2. Satellite remote sensing in the UV and visible wavelength ranges is widely used to study the emission and distribution of tropospheric NO2, which is one of the most crucial gases in both climate change and air pollution. One of the largest sources of uncertainty in NO2 satellite retrievals is the AMF, especially when enhanced trace gas concentrations are present in the lower troposphere. Computing the AMF is a very time-consuming task that is usually performed using radiative transfer models. In general, the practical use of such models for satellite remote sensing is limited by the available computational power. We constructed two DNN models to calculate the tropospheric AMF (Trop-AMF-Net) and the altitude-dependent box-AMF (Box-AMF-Net). Trop-AMF-Net consists of five multilayer perceptrons and a linear transform. Box-AMF-Net is an encoder-decoder framework that combines Trop-AMF-Net with a long short-term memory network. We prepared two test datasets, one of which reflects the actual observed NO2 measurement pattern and the other of which assumes a uniform distribution. For the former, we assumed that the NO2 observation was performed by the Global Observing SATellite for Greenhouse Gases and Water Cycle (GOSAT-GW), which is planned for launch in 2024. For this test dataset, Trop-AMF-Net could reproduce the tropospheric AMFs with root-mean-squared percentage errors (RMSPE) of 0.121 % and 0.156 % when the model was trained using the same and uniform distributions, respectively. The RMSPEs of the box-AMFs in the troposphere predicted by Box-AMF-Net are 0.284 % for the test and training datasets when the observed pattern was used. The computation time for 10,000 samples using a combination of a central processing unit and graphics processing unit is 3.7 seconds. The RMSPE and computation time are approximately 30 times smaller and two times shorter compared to those of the commonly used look-up table and interpolation technique. Our feasibility study also highlights the importance of training the model with a dataset that is consistent with the test use.