Towards inverse estimation of Spanish NOx emissions with TROPOMI observations using a variational autoencoder

Air pollution estimates represent key inputs in computer models for assessing air quality. They are also important in the evaluation of pollution control policies. 

In the last decade, neural networks have demonstrated exceptional ability to model complex spatiotemporal data. Meanwhile, advances in our ability to observe the earth's atmosphere using satellites have enabled the collection of high-resolution atmospheric composition data in near real-time. These developments open up opportunities to combine the predictive power of neural networks with satellite observations to deliver rapid and accurate estimates of pollutant emissions in near real-time.

Chemical weather prediction models offer insights into the forward relationship between emissions and atmospheric composition, and some studies are already suggesting that neural networks might be able to estimate with reasonable predictive skills the chemical concentrations obtained from these physics-based models. While the forward mapping is well-defined, the inverse mapping—from atmospheric composition to emissions— is not. Our objective is ultimately to exploit neural networks to predict emissions from atmospheric composition. This presents challenges, as we will show in our presentation.

We present preliminary results from our study in training a variational autoencoder, with data from a chemical weather prediction model, to invert Spanish NOx emissions. We demonstrate a workflow in which we jointly train two neural network models: one for inverse modelling of emissions and a second to regularise the predictions of the inverse model.