Synergistic Estimation of Surface Particulate Matter and Ozone Pollutants to enhance accuracy and interpretability by a Deep Learning Approach

Fine particulate matter (PM 2.5) and ozone (O3) are two types of main air pollutants that attract attention and can pose significant risks to public health. To improve estimation accuracy and our understanding of them, this study presents a novel synergistic estimating deep learning model. This model integrates the convolutional neural network (CNN) with the long short-term memory (LSTM) network for the multitask estimation by a soft parameter-sharing strategy. Compared to the single-task method, our method can better consider the implicit physical-chemistry transformation between the PM2.5 and O3 to accurately estimate results. It is found that the LSTM shows higher accuracy to extract time-related features in this type of task than the extended LSTM (xLSTM), bidirectional LSTM with attention (BiLSTM-Attention), and Transformer networks. Validations demonstrate that our synergistic estimation achieves the root mean square error (RMSE) of 9.361 μ m/m3 for PM2.5 and 11.114 μ m/m3 for O3, with coefficients of determination R ²) of 0.921 and 0.942, respectively. This is better than the single estimation that R ² of PM2.5 and O3 are 0.904 and 0.935, respectively. From the perspective of interpretability, the aerosol optical depth (AOD) and NO2 data, respectively, contribute the most to estimating PM2.5 and O3. The spatial distribution of the input features has a more significant impact on the results compared to the temporal changes. However, an exception to O3 is that the correlation between NO2 from the previous three days and O3 on the current day is higher than that between the current day's NO2 and O3, revealing the lag in the transformation of precursor substances. In addition, the Shapley additive explanation value of NO2 to O3 shows an expected downward trend with increasing NO2. This correctly suggests inhibition and equilibrium of photochemical reactions caused by high concentrations. Overall, the proposed method can help to generate a high-accuracy distribution of PM2.5 and O3, and this high R ² suggests that even in the absence of physical and chemical constraints, the machine learning approach has the potential to correctly grasp the objective patterns among air pollutants.