A comparison of machine learning-based approaches in estimating surface PM2.5 concentrations focusing on Artificial Neural Networks and high pollution events

Surface PM2.5 concentrations have significant implications for human health, necessitating accurate estimations. This study compares various machine learning models, including linear models, tree-based algorithms, and artificial neural networks (ANNs) for estimating PM2.5 concentrations using the MERRA-2 dataset from 2012 to 2023. Mutual information and Spearman cross-feature correlation scores are used during feature selections. The performance of models is evaluated using metrics including normalized Nash–Sutcliffe efficiency (NNSE), root mean standard deviation ratio (RSR), and mean percentage error (MPE). Our results show that ANNs outperform linear and tree models, particularly in estimating daily PM2.5 concentrations of 35–1000 µg/m3. ANNs improve NNSE by 119% and 46%, RSR by 40% and 24%, and MPE by 44% and 30% from linear and tree models, respectively, indicating ANN’s superior estimation performance during high pollution days. The sensitivity analysis of features that interpret the models suggests that the total extinction AOD at 550 nm and surface CO concentrations are the most important features in the Western and Eastern U.S., respectively. The findings suggest that even the simplest NNs provide better air quality estimates, especially during high pollution events, which is beneficial for long-term exposure analysis. Future research should explore more sophisticated NN architectures with spatial and temporal variations in PM2.5 to improve the model performance.