Intelligent monitoring and management of urban drainage systems through machine-learning-driven fusion of EEM and UV-Vis.

Reagent-free dissolved organic carbon (DOC) monitoring can support integrated management of urban drainage systems (UDS). This study developed a fusion spectrum (FS) framework integrating excitation-emission matrix fluorescence, ultraviolet-visible (UV-Vis) spectroscopy, and machine learning (ML) for DOC prediction across sewers, WWTPs, rivers, and seas. Among eight ML algorithms, XGBoost showed the best overall performance and was used to construct the unified XG-FS-UDS model. The unified model achieved an overall R2 of 0.96. On individual matrices, R2 values were 0.89, 0.96, 0.66, and 0.92 for WWTP, sewer, river, and sea samples, respectively, comparable to but not exceeding matrix-specific models. Under leave-one-matrix-out validation, R2 values were 0.88, 0.93, 0.60, and 0.84, while chronological splitting yielded 0.84, 0.90, 0.53, and 0.84, indicating matrix-dependent generalization and reduced performance for river samples. The mean-per-matrix baseline performed substantially worse, confirming that predictions relied on within-matrix spectral variation rather than matrix-level DOC means. DOC composition shifted from protein-like substances in sewers and WWTPs toward fulvic- and humic-like substances in rivers and seas. Shapley additive explanations analysis identified PARAFAC-derived components 1 and 2 and the fourth UV-Vis principal component 4 as major contributors. Overall, the framework demonstrates the feasibility of a single interpretable model for multi-matrix DOC monitoring with a modest performance trade-off relative to dedicated models.