Two-Stage Feature Selection for PM2.5 Prediction: adjcor-TRBFNN Approach to Address Multicollinearity in Air Quality Data in Urban Area

PM2.5 is a dangerous pollutant that has been shown to significantly affect human health due to its small size ($\leq 2.5$ micrometers). Urban areas often experience higher PM2.5 levels due to high traffic volume, industrial activities, and population density. Accurate prediction of PM2.5 levels not only enhances the body of knowledge on air quality modeling but also provides governments with critical insights to implement timely interventions, reduce public health risks, and guide policymaking for air quality management. Thus, this study aims to classify the next day’s air quality based on PM2.5 levels in Banting, an urban area in Selangor, Malaysia, using a machine learning approach. Radial Basis Function Neural Network (RBFNN) is a neural network method widely used for prediction and classification analysis due to it can handle complex nonlinear relationship in the dataset. However, RBFNN does not address multicollinearity in air quality data. To address this, we used the adjcorT feature selection method to identify the significant features affecting next-day PM2.5 levels, while considering the multicollinearity issues in air quality data. Thus, a two-stage feature selection method, adjcorT-RBFNN, was implemented in this study to enhance the classification of nextday PM2.5 levels. The research utilized hourly secondary data from the Department of Environment Malaysia (DOE), spanning 2018 to 2022. The findings indicate that PM10, PM2.5, CO, NO2, humidity, SO2, wind speed, and wind direction are significant features for classifying next-day PM2.5 levels in Banting. Moreover, the proposed adjcorT-RBFNN technique outperformed the RBFNN method in terms of accuracy, specificity, precision, F1 score, and AUROC, with scores of $0.761,0.752,0.752,0.761$, and 0.761, respectively. Hence, the proposed two-stage feature selection method, adjcorT-RBFNN, can assist future researchers in improving PM2.5 level predictions in other urban areas.