Land use regression modelling of long term PM2.5 exposure using regulatory monitoring data and machine learning in Mashhad Iran

Accurate assessment of long-term exposure to fine particulate matter (PM 2.5 ) remains challenging in many Middle Eastern cities because of sparse monitoring networks and complex emission sources. This study developed a machine learning-based land use regression (LUR) framework using a Gradient Boosting Machine (GBM) to estimate annual mean PM 2.5 concentrations across Mashhad, the second most populous city in Iran. Six-year average PM 2.5 concentrations (2019–2025) were obtained from 21 fixed-site regulatory monitoring stations. A total of 108 geospatial predictor variables were generated within a geographic information system, including topographic characteristics (elevation and slope), population density, land use attributes, vegetation indices (NDVI), transportation-related variables, and proximity indicators. Predictor variables were calculated using circular buffer radii ranging from 250 to 2000 m to identify the optimal spatial scale of influence.Ten statistical and machine-learning algorithms were evaluated using leave-one-out cross-validation (LOOCV) to assess out-of-sample predictive performance. Linear and regularized regression models showed relatively limited predictive ability (LOOCV R 2 = 0.517–0.562), whereas non-linear and ensemble approaches demonstrated improved performance. Among all evaluated algorithms, the Gradient Boosting Machine achieved the best predictive performance, with a leave-one-out cross-validated coefficient of determination (R 2 ) of 0.824 and a root mean square error (RMSE) of 2.347 µg m −3 .The final model consistently selected predictor variables within a 1000 m buffer radius, highlighting the importance of neighborhood-scale processes in shaping long-term PM 2.5 exposure. Elevation, primary road length, highway length, and population density emerged as the most influential predictors, reflecting the combined effects of topography and traffic-related anthropogenic activity. These findings demonstrate that machine learning-based land use regression models can provide reliable spatial estimates of long-term PM 2.5 exposure in urban environments with limited monitoring infrastructure and may support future epidemiological studies and evidence-based air quality management.