AdjcorT-RBFNN for Air Quality Classification: Mitigating Multicollinearity with Real and Simulated Data

Air pollution levels have remained a significant issue worldwide despite advancements in technology, primarily due to rapid industrialization and urbanization. Among the various pollutants, PM2.5 significantly impacts air quality, posing health risks such as respiratory and cardiovascular diseases. Accurate prediction of PM2.5 levels is essential for effective air quality management. However, multicollinearity in air quality data can hinder model performance. To address this issue, this study introduces the AdjcorT-RBFNN, a two-stage feature selection method, to classify air quality in Klang, Selangor. The AdjcorT-RBFNN model selects the optimal combination of 9 feature combinations from 10 variables and outperforms the RBFNN model, which uses all 10 variables. With 7 hidden nodes and a learning rate of 0.01 for both models, AdjcorT-RBFNN achieves higher accuracy (0.62), sensitivity (0.64), specificity (0.60), precision (0.60), F1 score (0.62), and AUROC (0.62), confirming its effectiveness in classification tasks. The optimal features for predicting air quality in Klang are identified as PM2.5, PM10, relative humidity, SO2, wind direction, O3, CO, ambient temperature, and NO2. Monte Carlo simulations validate the model’s effectiveness, showing that AdjcorT-RBFNN consistently outperforms RBFNN, especially with strong negative correlations (ρ=-0.8) and larger sample sizes (N=150 and 200) further enhance classification accuracy. Compared to RBFNN, AdjcorT-RBFNN enhances class discrimination and reduces false positives, improving its reliability in detecting true classifications. These findings highlight the importance of feature selection in improving model performance, particularly in datasets with multicollinearity. Researchers, and health organizations can leverage AdjcorT-RBFNN for more accurate air quality predictions, supporting informed pollution control strategies.