Inverting Spanish NOx emissions using a neural network

Air pollutant emissions represent key input information for modeling atmospheric chemical composition or evaluating air pollution control policies. Over the last decades, different approaches have been proposed for estimating emissions. These approaches include collecting activity data combined with emission factors to build bottom-up emission inventories, developing appropriate spatial and temporal disaggregation proxies applied to national or regional total estimates to construct top-down emission inventories, and creating complex data assimilation workflows around chemistry-transport models combined with observations to perform air pollution emission inverse modeling.

In the last decade, deep neural networks (DNNs) have demonstrated exceptional ability to model complex spatiotemporal data. Meanwhile, advances in Earth observation systems, such as the Tropospheric Monitoring Instrument (TROPOMI), have enabled the collection of high-resolution atmospheric composition data in near real-time. These developments open up opportunities to integrate the predictive power of DNNs with satellite observations to deliver rapid and accurate estimates of pollutant emissions in near real-time.

Deterministic CTMs offer insights into the forward relationship between emissions and atmospheric composition, and some studies are already suggesting that DNN 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 inherently ill-posed. Our objective is ultimately to exploit DNNs for doing air pollutant emission inverse modeling, without using the traditional data assimilation approach. This presents challenges, requiring the application of regularization techniques to address the ambiguity raised by this ill-posed problem.

Here, we will present the preliminary results of our study on training regularized DNNs for inverting the NOx emissions in Spain , utilizing training data derived from the MONARCH air quality model. We will take advantage of the flexibility offered by these models to create different training datasets and assess the performance of our models across different data scenarios.