Reconstructing 6-hourly PM 2.5 datasets from 1960 to 1 2020 in China 2

16 Fine particulate matter (PM 2.5 ) has altered radiation balance on earth and raised environmental and 17 health risks for decades, but only been monitored widely since 2013 in China. Historical long-term 18 PM 2.5 records with high temporal resolution are essential but lacking for both research and 19 environmental management. Here, we reconstruct a site-based PM 2.5 dataset at 6-hour intervals from 20 1960 to 2020 that combines long-term visibility, conventional meteorological observations, 21 emissions, and elevation. The PM 2.5 concentration at each site is estimated based on an advanced 22 machine learning model, LightGBM, that takes advantage of spatial features from 20 surrounding 23 meteorological stations. Our model's performance is comparable or even better than those of 24 previous studies in by-year cross validation (CV) (R 2 =0.7) and spatial CV (R 2 =0.76), and more 25 advantageous in long-term records and high temporal resolution. This model also reconstructs a 26 0.25°×0.25°, 6-hourly, gridded PM 2.5 dataset by incorporating spatial features. The results show 27 PM 2.5 pollution worsens gradually or maintains before 2010 from an interdecadal scale but mitigates 28 in the following decade. Although the turning points vary in different regions, PM 2.5 mass 29 concentrations in key regions decreased significantly after 2013 due to clean air actions. In particular, 30 the annual average value of PM 2.5 in 2020 is nearly at the lowest value in history since 1960. These 31 two PM 2.5 datasets (publicly available at https://doi.org/10.5281/zenodo.6372847) provide 32 spatiotemporal variations at high resolution, which lay the foundation of research studies associated 33 with air pollution, climate change, and atmospheric chemical reanalysis. based on 674 publicly available 101 meteorological stations. Gui et al. (2020) constructed a virtual daily PM 2.5 network at 1180 102 meteorological sites between 2017-2018. Our previous research also shows that the visibility-based 103 machine learning model that takes advantage of spatial features has great potential in reconstructing 104 historical PM 2.5 datasets with long-term records and high temporal resolution (Zhong et al., 2021). 105 In this study, we reconstruct a site-based PM 2.5 dataset at 6-hour intervals from 1960 to 2020 based 106 on long-term visibility and conventional meteorological observations from ~2450 national stations, 107 together with emissions and elevation. The PM 2.5 concentration at each site is estimated based on a 108 Light Gradient Boosting Machine (LightGBM) model that takes advantage of spatial features from 109 20 surrounding meteorological stations. By incorporating spatial features, this model also 110 reconstructs a 0.25°×0.25°, 6-hourly, gridded PM 2.5 dataset. These two PM 2.5 datasets provide 111 spatiotemporal variations at high resolution, which constitute the basis for research studies 112 associated with air pollution, climate change, and atmospheric chemical reanalysis.