Random Forest Model Accurately Predicts Winter PM2.5 Across Riyadh

A machine learning framework developed using the Google Earth Engine platform accurately mapped atmospheric fine particulate matter (PM2.5) concentrations across Riyadh, Saudi Arabia, during the winter season, with a Random Forest model outperforming other tested algorithms. In the study, researcher Chaitanya Baliram Pande addressed the challenges of evaluating PM2.5 levels in arid regions where ground-based air quality monitoring stations are infrequent and natural dust storms exert strong environmental effects.

To overcome the sparsity of ground observations, the framework integrated multi-source remote sensing and reanalysis datasets within the Google Earth Engine platform. Atmospheric PM2.5 data derived from the MERRA-2 reanalysis dataset were combined with topographical, climatic, and biophysical variables. Specifically, the framework incorporated digital elevation model (DEM) data, temperature readings from ERA5-Land, rainfall estimates from TerraClimate, and vegetation indices from MODIS NDVI.

The study evaluated three distinct machine learning algorithms trained and validated to predict winter PM2.5 spatial distributions across Riyadh. These models included Random Forest (RF), NU-support vector regression (NU-SVR), and Epsilon-support vector regression (Epsilon-SVR). This setup allowed for a comparative assessment of predictive performance by integrating climate, biophysical, and elevation factors that have rarely been evaluated concurrently in arid zone air pollution modeling.

Predictive performance across the candidate models was evaluated using multiple statistical criteria, including the coefficient of determination (R2), root mean square error (RMSE), mean squared error (MSE), bias, Willmott’s Index of Agreement (WI), and Performance Index (PI). The Random Forest model achieved the highest predictive accuracy and exhibited the lowest prediction errors among the evaluated models, yielding an R2 of 0.82, an RMSE of 1.76, an MSE of 2.2583, a bias of -0.0377, a Willmott’s Index of 0.9479, and a Performance Index of 0.4198. Conversely, the Epsilon-SVR model showed the poorest accuracy across the metric evaluations.

The fine-resolution PM2.5 mapping generated by the Random Forest model captured detailed spatial distributions of particulate matter across the study area. This spatial modeling enabled the identification of urban-industrial hotspots located within Riyadh and its adjacent vicinities, where PM2.5 concentrations were observed at relatively elevated levels compared to surrounding areas.

The findings demonstrate the utility of combined remote sensing and machine learning frameworks for atmospheric monitoring in regions with limited in situ air quality information and datasets. The research indicates that this scalable data-driven workflow offers tools for real-time monitoring, informed policymaking, and targeted pollution mitigation strategies intended to support public health and sustainable development across arid, semi-arid, and desert environments.

Subject of Research: Machine learning prediction of winter PM2.5 air pollution in Riyadh

Article Title: Prediction modeling of spatial distribution for atmospheric PM2.5 (µg/m³) level in winter season of Saudi Arabia city using machine learning models and GEE approachs

Article References: Pande, C. B. (2026). Prediction modeling of spatial distribution for atmospheric PM2.5 (µg/m³) level in winter season of Saudi Arabia city using machine learning models and GEE approachs. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09659-x

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09659-x

Keywords: PM2.5, air pollution, machine learning, Random Forest, Riyadh, Google Earth Engine, remote sensing, Saudi Arabia, Prediction, modeling, spatial, distribution