Explainable and spatial dependence deep learning model for satellite-based O3 monitoring in China

Environmental exposure to surface ozone (O3) has become a major public health concern. To accurately estimate the spatial-coverage O3 from sparse ground-truth data, we here propose a two-stage, deep learning model, “the explainable and spatial dependence deep learning model (ExDLM)”, which combines convolutional neural networks (CNN), deep neural network (DNN), and integrated gradients (IG). Compared to individual CNN and DNN, our model showed higher accuracy and exhibited the highest R2 of 0.78 and the lowest RMSE of 18.35 μg/m3. The estimated O3 was 66.19 ± 33.87 μg/m3 as compared to the 69.51 ± 39.38 μg/m3 calculated using the ground-truth data. Using ExDLM, we interpreted the contribution of nearby cities to O3 in Beijing during extreme weather (dust storms) and clean days. During dust storms, the surrounding dust cells had negative IG scores, ranging from −1.43 to −0.01, indicating that these areas inhibited the formation of O3 in Beijing. Conversely, in clean days, especially during summer when O3 pollution is often extreme, the surrounding cells had positive scores, indicating that these areas enhanced O3 formation. Nearby cities had the highest scores, ranging from 0.05 to 0.11. Using the proposed model, we were able to assess O3 dynamics in Beijing, with greater temporal and spatial accuracy than that achieved by current models. The ExDLM also allows for finer-scale analysis of O3 pollution, even under dust storms conditions, which traditionally limit model accuracy, as well as great spatial interpretability.