Construction and analysis of atmospheric visibility and fog-haze datasets in China (2001−2023) based on machine learning models

Research on atmospheric visibility and haze events is vital for human health and sustainable societal development. This study aims to construct long-term daily-scale visibility and fog-haze datasets for China using machine learning models, analyze their spatiotemporal evolution characteristics, identify key influencing factors of visibility. The results show that the LightGBM model achieves high accuracy (R 2 = 0.79, RMSE = 4.03 km), representing a 7.04 % improvement in R 2 and 0.6 km reduction in RMSE compared to the baseline model, while maintaining robust stability across both seasonal (R 2 = 0.77–0.79, RMSE = 3.83–4.18 km) and spatial scales (R 2 = 0.58–0.76, RMSE <4 km). Atmospheric visibility showed an initial decline followed by an upward trend from 2001 to 2023, reaching its lowest point in 2015 (20.77 km) and recovering to 21.26 km by 2023. Visibility was lowest in winter (20.33 km) and highest in summer (22.32 km), with low-value areas concentrated in eastern and central city clusters. Overall, a declining trend was observed, with an annual average decrease of 0.03–0.08 km. From 2000 to 2013, the annual average visibility decreased by 0.09–0.13 km, with the decline area accounting for 22.56 %–55.38 %. From 2013 to 2023, the annual average visibility increased by 0.03–0.13 km, with the increase concentrated in the central and eastern regions (area ratio of 12.7 %–39.1 %), while the western regions continued to show a downward trend (area ratio of 8 %–35.3 %). SO₂, relative humidity, and surface pressure are the most important variables affecting atmospheric visibility, with contributions of 17.43 %, 13.29 %, and 10.58 %, respectively. The thresholds for these variables are 11.4 μg/m 3 , 74.9 %, and 998 hPa. The frequency of haze in China increased from 9.97 days in 2001 to 20.3 days in 2013 and then decreased to 11.03 days by 2023. Overall, there is an average annual decrease of 0.5 days, with the spatial pattern showing a decline in the eastern and central regions, and an increase in the western region. Specifically, light haze and alert haze decreased by 0.75 and 0.2 days per year, respectively, while mild haze increased by 0.32 days per year. The area without haze has continuously decreased by 11.89 %, while the area with low-frequency haze has increased by 12.91 %. The area with high-frequency haze first increased and then decreased, showing an overall reduction of 1.03 %, mainly concentrated in city clusters in the eastern and central regions. From 2001 to 2013, the area of high-frequency haze increased by 5.68 %, while it decreased by 6.71 % from 2013 to 2023. The research results provide important references for a deeper understanding of atmospheric environmental changes, while also offering data support and scientific basis for atmospheric environment research and effective management. • A machine learning model constructed China’s seamless daily visibility and fog-haze dataset. • Spatiotemporal evolution of long-term visibility and haze frequency was revealed. • SO₂, relative humidity, and surface pressure were key drivers of atmospheric visibility.