Comparison of two receptor models PCA-MLR and PMF for source identification and apportionment of pollution carried by runoff from catchment and sub-watershed areas with mixed land cover in South Korea

Abstract

The application and comparison of receptor modeling techniques based on ambient air quality and particulate matter increasingly being studied. However, less is known about the comparison of receptor modeling techniques using spatial runoff quality data to identify and quantify the stormwater runoff pollution. This study compared the performance of principal component analysis-multiple linear regressions (PCA-MLR) and positive matrix factorization (PMF) models on stormwater runoff data collected from a small catchment (Site 1) with urban development activity and a sub-watershed outlet (Site 2). In both sites, the PCA-MLR model identified three pollution sources, whereas PMF identified five with a detailed source mechanism including two additional sources. Furthermore, the spatial land-use land-cover (LULC) analysis results indicate that the Site 1 exhibited a rapid conversion of the native area into a built-up area over the monitoring period compared to Site 2. Based on the modeling results, domestic wastewater and soil erosion were the major source of pollution at Site 1 and Site 2, respectively. The performance evaluation statistics including Nash coefficient (0.86–0.99), % error (<−14 to 2), and coefficient of determination (R2 ≤ 0.99) showed better performance for the PMF model than the PCA-MLR model. Overall, the PMF receptor modeling approach was found to be more robust for the current study sites with different land use types. The findings of this study could provide a basis for further application of these receptor models and their comparison using spatial-temporal ionic and sediment related runoff monitoring data. Also, the models from this research could be combined with other receptor models on runoff quality data (e.g. CMB or UNMIX) to explore and inter-compare the outcomes, and to determine how the model results are affected by modifications to input data and model parameters. Therefore, further research is required to precisely assess the accuracy of both receptor models.