Evaluation of machine learning techniques with multiple remote sensing datasets in estimating monthly concentrations of ground-level PM2.5.

Fine particulate matter (PM2.5) has been recognized as a key air pollutant that can influence population health risk, especially during extreme cases such as wildfires. Previous studies have applied geospatial techniques such as land use regression to map the ground-level PM2.5, while some recent studies have found that Aerosol Optical Depth (AOD) derived from satellite images and machine learning techniques may be two elements that can improve spatiotemporal prediction. However, there has been a lack of studies evaluating use of different machine learning techniques with AOD datasets for mapping PM2.5, especially in areas with high spatiotemporal variability of PM2.5. In this study, we compared the performance of eight predictive algorithms with the use of multiple remote sensing datasets, including satellite-derived AOD data, for the prediction of ground-level PM2.5 concentration. Based on the results, Cubist, random forest and eXtreme Gradient Boosting were the algorithms with better performance, while Cubist was the best (CV-RMSE = 2.64 μg/m3, CV-R2 = 0.48). Variable importance analysis indicated that the predictors with the highest contributions in modelling were monthly AOD and elevation. In conclusion, appropriate selection of machine learning algorithms can improve ground-level PM2.5 estimation, especially for areas with nonlinear relationships between PM2.5 and predictors caused by complex terrain. Satellite-derived data such as AOD and land surface temperature (LST) can also be substitutes for traditional datasets retrieved from weather stations, especially for areas with sparse and uneven distribution of stations.