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

Abstract In this paper, focuses on the important air pollution factor like fine particulate matter (PM₂.₅ (µg/m³) and studied in the winter season, which is increase overall the worldwide and arid regions like Saudi Arabia country. Currently, the important challenges are monitoring and modeling of PM₂.₅ (µg/m³) due to infrequent ground-based observations and strong effects from natural dust storms. Hence, this paper address these limitations, this research purposes a effective framework based on the combination of multi-source remote sensing and reanalysis datasets processed by using machine learning (ML) models in the Google Earth Engine (GEE) platform, to generate fine resolution PM₂.₅ (µg/m³) maps as compare with MERRA-2 reanalysis dataset for winter-season in the Riyadh city of Saudi Arabia. The PM₂.₅ (µg/m³) is derived from MERRA-2 reanalysis datasets were combined with other variables including digital elevation model (DEM), temperature (ERA5-Land), rainfall (TerraClimate), and vegetation indices (MODIS NDVI). In this research adopted the three ML models such as Random Forest (RF), NU-supper vector regression (SVR), and Epsilon-SVR, which are trained and tested (validated) to compare their performance of prediction of PM₂.₅ (µg/m³) in the Riyadh city at Kingdom of Saudi Arabia. The effective scientific approach of this paper to details integration of climate, biophysical, and topographical variables, which have infrequently been together discovered in PM₂.₅ (µg/m³) modeling in arid region. The evaluation of performance models are computed by R 2 , RMSE, BIAS, MSE, Willmott’s Index of Agreement and Performance Index, which is helping to measure the which model best perform and various predictive strength. Finally overall the performance metrics indicated the RF model is best model to captures the detailed spatial distribution of PM₂.₅ (µg/m³), Therefore the RF model is showing the performance metrics results such as R 2 , RMSE, BIAS, MSE, WI, and PI values of 0.82, 1.76, -0.0377, 2,2583, 0.9479, and 0.4198, respectively. This model showing the minimum lower errors compared with other models. The EPSILON_SVR model is showing poor accuracy based on overall performance metrics as compared other models. To identify the urban-industrial hotspots at the Riyadh city and near by areas, where is observed PM₂.₅ (µg/m³) levels are relatively higher level. These findings highlight the significance of the combination of the data-driven approaches in the areas with limited in situ air quality information and datasets. These results could provide valuable supports and scalable solution for real-time monitoring and policy making, this model can execution implications for executing targeted air pollution control measures, finally contributing to enhanced public health and sustainable development in the desert, arid and semi-arid environments.