Machine Education Approach for Generating Accurate NO2 and PM2.5 Dense Pollution Maps in Israel

Avitay Geltman, Barak Fishbain

Abstract: Dense pollution maps are essential for understanding and reducing air pollution. However, pollution measurements are often sparse. Common methods to address this gap are chemistry transport models (CTMs) and air pollution interpolation models. Nonetheless, many of these models have poor performances and intrinsic systematic biases. Here, a machine education approach, which integrates a CTM and in situ measurements with Artificial Intelligence (AI) techniques, is proposed for generating dense pollution maps with enhanced accuracy. Specifically, the CHIMERE CTM is combined with a Neural Network, eXtreme Gradient Boosting (XGBoost), or Random Forest AI methods. The results show that the educated machine models significantly improved predictions of air pollution levels compared to the standalone CTM. The machine education model combining a neural network with CHIMERE performed best among all models, followed closely by the educated model with XGBoost, while the Random Forest model trailed. Relative to CHIMERE, the proposed models achieved reductions of up to 51.34% and 50.54% in the root mean square error (RMSE) and mean absolute error (MAE), respectively, for NO2. For PM2.5, the models demonstrated relative reductions of up to 40.08% and 36.54%, respectively. CHIMERE exhibited an inherent systematic underestimation bias, characterized by mean fractional bias (MFB) values of 0.509 and 0.725 for NO2 and PM2.5, respectively. Our models successfully eliminated this bias. Furthermore, promising dense air pollution maps were generated using our models.