Half-Century Monthly Mean PM2.5 over Global Land from Visibility Observations

Abstract Fine particulate matter (PM 2.5 ) significantly impacts the climate, environment, and human health. PM 2.5 data are obtained through direct ground-based measurements, indirect estimations, and modeling reanalysis. However, historical PM 2.5 data are limited by the short duration of ground-based observations and satellite-based estimations, and the constraints of assimilation data scarcity in reanalysis, especially before 2000. Leveraging the strong correlation between PM 2.5 and visibility and the extensive, long-term visibility observations, this study reconstructs a half-century (1973–2022) 0.25° monthly PM 2.5 data over global land using a machine learning model and geographically weighted regression. The reconstruction integrates PM 2.5 observations (∼5000 sites), visibility data (∼12 000 stations), and auxiliary datasets. Machine learning model evaluations show excellent performance and robust predictions for site-scale PM 2.5 data ( R = 0.97 for cross validation; R = 0.85 and 0.92 for historical and future scenarios). Space–time consistency examinations demonstrate strong agreement with independent observations and satellite-based datasets. The gridded PM 2.5 data by the interpolation model perform well against ground-truth data ( R = 0.92), and error analysis indicates small, stable errors (1.95–2.32 μ g m −3 ), although confidence is lower in unobserved regions such as deserts. During 1973 and 2022, spatial heterogeneity indicates high levels in northern and central Africa and Asia, moderate levels in Central and South America, eastern Europe, southern Africa, and Australia, and low levels in North America and western Europe. Regional trends highlight a decline in Asia due to pollution controls after 2010, an increase in Western North America linked to anthropogenic emissions post-2015, and a sustained growth in Central America likely driven by urbanization. It fills historical gaps in PM 2.5 records and provides an essential foundation for climate, environment, and human health studies. Significance Statement The purpose of this study is to provide a half-century global PM 2.5 dataset (1973–2022). This work addresses the substantial lack of long-term PM 2.5 observations, especially before 2000, using a machine learning approach combined with spatial interpolation based on the widespread and historical availability of visibility data. The resulting dataset presents consistent global coverage and enables robust assessments of long-term PM 2.5 trends and their implications for climate, environmental change, and public health. Notably, the dataset reveals key regional patterns, including recent declines in PM 2.5 concentrations in parts of Asia attributable to strengthened air pollution controls.