Temporal and spatial patterns of PM2.5 in Mainz, Germany: insights from satellite observations and machine learning

Abstract

Fine particulate matter (PM2.5) remains one of the most harmful air pollutants affecting human health and urban environments. Although ground-based monitoring stations provide accurate measurements, their limited spatial coverage constrains detailed assessments of local air quality. This study investigates the temporal and spatial distribution of satellite-derived PM2.5 concentrations in Mainz, Germany, between 2017 and 2022 using statistical and machine learning approaches. Monthly and annual PM2.5 data from the Atmospheric Composition Analysis Group were combined with meteorological variables obtained from the WorldClim database. Temporal trends were evaluated using linear regression analyses of monthly and seasonal averages, while spatial patterns were modelled using Thin Plate Spline (TPS) interpolation and Random Forest (RF) regression. The results demonstrate a consistent decline in annual PM2.5 concentrations throughout the study period, with the strongest reductions observed during winter. Summer and autumn exhibited relatively stable concentration levels. Spatial modelling revealed persistent pollution hotspots in the southern part of Mainz, whereas lower concentrations occurred in the northern and peripheral areas. TPS produced smooth and interpretable spatial surfaces suitable for visualization, while RF captured additional spatial variability by incorporating meteorological predictors. However, the RF model was limited by the availability of climate data and the exclusion of anthropogenic variables such as traffic intensity and industrial emissions. The findings demonstrate that integrating satellite observations with geospatial modelling provides valuable insights into urban air pollution dynamics and supports environmental monitoring where ground-based measurements are sparse. The proposed workflow offers a practical framework for regional air quality assessment and future spatial prediction studies.