Study on the model transfer method for quantitative detection of water quality indicators in surface water

Surface water is a core resource for maintaining ecological balance and supporting the sustainable development of the national economy, and its quality monitoring is crucial to environmental safety. Ultraviolet–visible (UV–Vis) spectroscopy has become an ideal choice for online water quality monitoring because it requires no chemical reagents and can rapidly and simultaneously detect multiple parameters such as total organic carbon (TOC), nitrate nitrogen, and nitrite nitrogen. However, in practical applications, differences in spectral responses between different instruments often lead to model failure, necessitating repeated modeling for each instrument, which significantly increases costs. To address this issue, this paper proposes a sample selection method for the transfer set based on joint Euclidean distance and cosine information (SPEC) and employs robust regression based on univariate correction for model transfer. Specifically, SPEC ensures that the selected samples contain richer spectral information and enhances their representativeness by fusing spectral intensity and shape features. Meanwhile, robust regression based on univariate correction treats the differences in single-wavelength responses between instruments as random disturbances, incorporates them into calculations as a regularization term, and obtains the transformation matrix by solving robust regression residuals, thereby achieving accurate calibration of the secondary instrument spectra to the primary instrument spectra. Experimental results with real surface water samples show that this scheme effectively realizes model transfer between primary and secondary instruments: compared with traditional transfer set selection methods and model transfer methods, the prediction errors in the detection of TOC, nitrate nitrogen, and nitrite nitrogen are significantly reduced, and the model stability and accuracy are better. This method is particularly suitable for cross-regional surface water monitoring scenarios with multi-instrument collaboration, as it can support model transfer for multi-parameter simultaneous detection under the condition of limited samples. This study provides a new idea for cross-instrument model transfer in spectral analysis and helps promote the large-scale application of UV–Vis spectroscopy in multi-instrument collaborative water quality monitoring.