Multivariate mixture density network for real-time monitoring of non-point source pollution in small watershed based...

Monitoring of non-point source pollution (NPSP) in small watersheds suffers from low monitoring frequency and sparse spatial coverage, limiting both our understanding and effective management of NPSP. UV-Visible (UV-Vis) spectrometry, known for its rapid response and cost-effectiveness, offers a promising solution to address this monitoring scarcity. However, key water quality parameters of NPSP, total phosphorus (TP) and total nitrogen (TN) (e.g., ammonia), are inherently challenging to retrieve from UV-Vis spectrum, because TP is optically-nonactive parameter and TN contains optically-nonactive parameter (e.g., ammonia). Moreover, uncertainty evaluation of retrieved water quality relies heavily on concurrent field measurements, lacking effective method for real-time uncertainty assessment. To address these challenges, we constructed a UV-Vis spectrometer network comprising 20 spectrometers and implemented a four-month in-situ NPSP monitoring in Pizhou City, China. Multivariate mixture density network (multi-MDN) was developed to model the covariation relationship among NPSP-related water quality conditioned on the given spectrum, enabling the joint retrieval of TN, TP, and COD_Mn (chemical oxygen demand). By leveraging this covariation relationship among NPSP-related water quality, the proposed multi-MDN effectively improved generalization capability for both optically-active and optically-nonactive constituents, with R² of 0.92, 0.84 and 0.90 for TN, TP and COD_Mn. We further proposed per-estimation uncertainty evaluation method for real-time assessment of NPSP monitoring and designed several experiments to examine its utility. Per-estimation uncertainty effectively identified large retrieval error, detected spectrometer anomalies (e.g., insufficient probe submergence, biofouling), and implied model update. Using the real-time in-situ observation from UV-Visible spectrometer network, we revealed the spatiotemporal variation of NPSP in rural and urban rivers of Pizhou. This study provides an accurate, robust and cost-effective solution to improve spatiotemporal coverage of NPSP monitoring, supporting informed management in small watersheds.