Abstract

This study investigates the application of Support Vector Machines (SVM) to classify major air pollution sources in Bengaluru, India, by integrating routinely collected air quality, meteorological, and land-use data from 2021 to 2023. The main objective is to assess whether commonly available datasets can accurately distinguish between vehicular, industrial, domestic, and biomass burning sources. Air pollutant concentrations (PM₂.₅, PM₁₀, NO₂, SO₂, CO, O₃) were combined with meteorological parameters, satellite-derived land-cover indices (NDVI, NDBI), and urban activity datasets to develop feature vectors for classification. Data preprocessing ensured quality control, synchronisation, and normalisation, while principal component analysis reduced dimensionality. An SVM with a radial basis function kernel was trained and evaluated using stratified cross-validation, with model stability improved through auxiliary Support Vector Regression (SVR) for temporal smoothing. The classifier achieved an overall accuracy of 70% (Cohen's kappa: 0.59), with best performance for biomass burning (F1-score: 0.78) and industrial emissions (F1-score: 0.68), and moderate success in differentiating vehicular (F1-score: 0.63) and domestic (F1-score: 0.64) sources. Predictor importance analysis revealed that road density, wind-adjusted pollutant concentrations, and land-cover indices were most influential. Spatial and temporal validation demonstrated consistency with external ground-truth activities. The findings suggest that SVM, supplemented by routine datasets, provides a robust, cost-effective alternative to traditional source apportionment for urban air quality management, with potential for real-time application in rapidly growing cities.