Using a Citizen-installed Network of PM2.5 Sensors to Predict Hourly PM2.5 Airborne Concentration

A growing number of scientific studies have shown that particulate matter harms the environment and endangers human health. Thus, making timely predictions about airborne particulate matter (PM) concentrations could help the general public to be better organized and avoid excessive exposure to harmful pollutants. This study analyzes the possibility of making accurate predictions about PM2.5 concentrations in ambient air. The proposed methodology is tested using the data from citizen-installed PM2.5 sensors from three locations (Serbia, North Macedonia, and Pakistan) that are relatively different in size, population (density), geographic, economic, social, and other relevant means. The data (study sample) were collected through the NASA data access viewer online platform and citizen-installed devices that sample PM2.5 concentrations (non-referent methods). Four predictive algorithms - Random Forest, XGBoost, CatBoost, and LightGBM - were employed to achieve this goal. The Sequential-Forward-Selection algorithm was used to simplify model building, contributing to the generalization of the methodology. Among the selected algorithms, CatBoost exhibited the best performance in Serbia and North Macedonia, while Random Forest performed best in Pakistan. The study conclusion is that here presented methodology is universally applicable for forecasting PM2.5 airborne concentration in the areas that are covered by citizen-installed PM2.5 sensors and are not necessarily covered by official referent sampling stations.