Spatiotemporal Analysis and Forecasting of PM2.5 Air Pollution in Azerbaijan Using SARIMA and Prophet Models

Fine particulate matter (PM2.5) remains a major environmental and public health concern worldwide. This study investigates the temporal and spatial dynamics of PM2.5 concentrations across Azerbaijan between 2019 and 2024 and develops forecasts using two transparent time-series models. Monthly PM2.5 fields were derived exclusively from ECMWF CAMS reanalysis products, while Sentinel-5P TROPOMI observations were used only for qualitative spatial interpretation and contextual comparison. All datasets were processed in Google Earth Engine and analyzed in R. Temporal dynamics were modelled using SARIMA and Prophet, while spatial patterns were characterized through stratified random-point sampling followed by inverse distance weighting (IDW) interpolation. The results reveal a pronounced seasonal cycle in which PM2.5 concentrations peak during the warm season rather than in winter. The highest concentrations consistently occurred in June and September, with the absolute maximum reaching 28.9 µg/m³ in June 2019, whereas the lowest values were systematically recorded in January and February. At the interannual scale, 2019 emerged as the most polluted year, followed by a substantial decline during 2020–2023, partly associated with COVID-19–related mobility restrictions, before concentrations increased again during the second half of 2024. Spatial analysis identified persistent hotspots in the central and south-central lowland districts, while the eastern coastal zone—particularly the Absheron Peninsula and greater Baku—experienced intense but episodic pollution events. Getis-Ord hotspot analysis and a highly significant positive Moran’s I statistic (0.922, p < 0.001) confirmed strong and non-random spatial clustering of PM2.5 concentrations. Across most locations and months, Prophet produced lower forecasting errors than SARIMA during the 2024 validation period (MAE = 0.180 µg/m³; RMSE = 0.202 µg/m³). Prophet also reproduced the location and intensity of seasonal pollution maxima more successfully, whereas SARIMA showed systematic spatial displacement of hotspots during several critical months. However, the trend shift observed during late 2024 and the absence of ground-based calibration data indicate that forecasting PM2.5 under non-stationary environmental conditions remains challenging. Overall, the proposed framework—combining freely available reanalysis products, simple spatial interpolation, and interpretable forecasting models—provides a practical and low-cost approach for routine air-quality monitoring, hotspot identification, and policy support in data-scarce regions.