Deep optimization of water quality index and positive matrix factorization models for water quality evaluation and pollution source apportionment using a random forest model

Effective evaluation of water quality and accurate quantification of pollution sources are essential for the sustainable use of water resources. Although water quality index (WQI) and positive matrix factorization (PMF) models have been proven to be applicable for surface water quality assessments and pollution source apportionments, these models still have potential for further development in today's data-driven, rapidly evolving technological era. This study coupled a machine learning technique, the random forest model, with WQI and PMF models to enhance their ability to analyze water pollution issues. Monitoring data of 12 water quality indicators from six sites along the Minjiang River from 2015 to 2020 were used to build a WQI model for determining the spatiotemporal water quality characteristics. Then, coupled with the random forest model, the importance of 12 indicators relative to the WQI was assessed. The total phosphorus (TP), total nitrogen (TN), chemical oxygen demand (CODCr), dissolved oxygen (DO), and five-day biochemical oxygen demand (BOD5) were identified as the top five significant parameters influencing water quality in the region. The improved WQI model constructed based on key parameters enabled high-precision (R2 = 0.9696) water quality prediction. Furthermore, the feature importance of the indicators was used as weights to adjust the results of the PMF model, allowing for a more reasonable pollutant source apportionment and revealing potential driving factors of variations in water quality. The final contributions of pollution sources in descending order were agricultural activities (30.26%), domestic sewage (29.07%), industrial wastewater (26.25%), seasonal factors (6.45%), soil erosion (6.19%), and unidentified sources (1.78%). This study provides a new perspective for a comprehensive understanding of the water pollution characteristics of rivers, and offers valuable references for the development of targeted strategies for water quality improvement.

New insights into pollution source analysis using receptor models in the upper Yangtze river basin: Effects of land use on source identification and apportionment

To effectively control pollution and improve water quality, it is essential to accurately analyze the potential pollution sources in rivers. The study proposes a hypothesis that land use can influence the identification and apportionment of pollution sources and tested it in two areas with different types of water pollution and land use. The redundancy analysis (RDA) results showed that the response mechanisms of water quality to land use differed among regions. In both regions, the results indicated that the water quality response relationship to land use provided important objective evidence for pollution source identification, and the RDA tool optimized the procedure of source analysis for receptor models. Positive matrix decomposition (PMF) and absolute principal component score-multiple linear regression (APCS-MLR) receptor models identified five and four pollution sources along with their corresponding characteristic parameters. PMF attributed agricultural nonpoint sources (23.8%) and domestic wastewater (32.7%) as the major sources in regions 1 and 2, respectively, while APCS-MLR identified mixed sources in both regions. In terms of model performance parameters, PMF demonstrated better-fit coefficients (R²) than APCS-MLR and had a lower error rate and proportion of unidentified sources. The results show that considering the effect of land use in the source analysis can overcome the subjectivity of the receptor model and improve the accuracy of pollution source identification and apportionment. The results of the study can help managers clarify the priorities of pollution prevention and control, and provide a new methodology for water environment management in similar watersheds.