Application of Deep Neural Networks in Top-Down NOx Emissions Estimation

This study presents an application of deep neural networks in the estimation of NOx emission using space-borne and surface observations. While top-down constraint of NOx emissions using satellite observations has been a popular way to complement bottom-up emissions estimation, its actual application is limited due to the nonlinear characteristics of the observed responses to the change of emissions. To consider these chemical nonlinearity and physical uncertainties, especially from local transports, we developed an emission inverse modeling framework to utilize a machine learning approach trained with pseudo-observational data sets based on a perfect model assumption. The perfect model assumes that the modeled world is true. Based on this assumption, we build pseudo- observations and pseudo- satellite that mimic true observations of real world. Pseudo observations are used to train the system to learn the emission-to-concentration relation using the deep neutral networks method in the Google TensorFlow library. Three sets of twelve years’ CMAQ simulations over East Asia (27-km domain) and South Korea (9-km domain), using MICS-Asia, INTEX-B, CREATE (for Asia) and CAPSS (for South Korea) emission inventories, are used for the training. Results demonstrate that, after training, the system is able to regenerate NOx emissions of independent period with accuracy. We further discuss the uncertainty and efficiency in using pseudo- observations, and limitations in actual application to real world observations.