Predicting annual PM2.5 in mainland China from 2014 to 2020 using multi temporal satellite product: An improved deep learning approach with spatial generalization ability

Abstract

The estimation of long-term fine particulate matter (PM2.5) concentrations and trend assessments play a critical role in preventing health risks to the human body. Deep learning methods based on time series have been highly accurate in predicting PM2.5. However, most time series models lack the ability for spatial generalization because they cannot combine the analysis at different spatial scales. In addition, temporal long-term trend analysis have not been reported in most of the studies. In this study, to reveal the spatiotemporal variability and trends of PM2.5, an improved deep learning framework named the SpatioTemporal Enhanced Neural Network (STENN) is developed for estimating PM2.5 concentrations with a spatial resolution of 1 km. Based on the bidirectional long short-term memory (LSTM) structure and attention mechanisms, the model provides a geographic-data-driven approach to incorporate the impact of the spatial heterogeneity and time dependence of PM2.5, which demonstrates that it has robust spatiotemporal transferable power with an R2 of 0.89 produced by cross validation (CV). High-resolution (1 km) and high-quality annual PM2.5 products for mainland China from 2014 to 2020 were constructed. In comparison with the current 1-km PM2.5 products, our framework demonstrates better stability in different regions, especially in terms of the high-value estimations and spatial continuity. The spatiotemporal PM2.5 distributions were also analyzed based on the time-series products. After the implementation of various control policies for atmospheric pollution, a declining trend of PM2.5 concentrations was observed in 88.79% of China between 2014 and 2020, with a mean decrease rate of 3.35 μg m−3 yr−1. This result indicates that the control policies of the Chinese government were effective in reducing PM2.5 concentrations. The PM2.5 concentrations in China reveal an exponential temporal trend, from a rapid decline to a gradual slowdown and a stable phase. To realize the vision of the Beautiful China Initiative, a regionally targeted policy for air pollution management is required. This study provides valuable implications for a more detailed analysis of the spatiotemporal variations in PM2.5 at small and medium spatial scales by developing an improved deep learning approach with a spatial generalization ability and integration with multi-temporal satellite products.