Non-contact hyperspectral monitoring of urban wastewater quality: Optimization of model calibration and performance

Monitoring pollution in urban drainage systems (UDS) is challenging due to their inaccessibility and harsh conditions. Reflectance spectrophotometry has emerged as a promising non-contact technique, but prior studies have largely focused on synthetic wastewaters or single sites. In this work, we evaluate visible and near-infrared hyperspectral imaging for monitoring of turbidity, dissolved organic carbon (DOC), and ammonium nitrogen (NH₄-N) in raw urban wastewater. We collected samples from five locations, including a stormwater sewer, two foul sewers, and two combined sewers. We used partial least squares regression models to predict pollutants from hyperspectral data. For model calibration, we explored three approaches: local models trained with five to thirty samples from a single site, global models trained on data from all sites but one, and hybrid models combining global models with two to twenty site-specific samples to enhance accuracy. Model performance is in general best for foul sewers, followed by the stormwater sewer, while it is worst for more variable combined sewers. Local calibrations with thirty training samples perform best, with cross-validated median errors of 6.5 % for turbidity, 14.1 % for DOC, and 22.3 % for NH₄-N. Global models perform satisfactorily only for turbidity, with a median error of 11.8 %. We also showed that model performance for turbidity can be explained by its correlation with reflected light intensity, while performance for DOC and NH₄-N can mainly be explained by their correlations with turbidity. Overall, these findings demonstrate the potential of (hyper)spectral imaging for low-maintenance, non-contact monitoring of key pollutants in urban drainage systems.