A robust UV–Vis spectral method for seawater COD detection based on principal component turbidity correction and ensemble variable selection

Reliable assessment of chemical oxygen demand (COD) in coastal seawater is essential for monitoring organic pollution and protecting marine ecosystems. However, turbidity-induced spectral interference and high-dimensional redundancy in UV–Vis spectra hinder the accuracy of conventional modeling methods. This study presents an integrated approach that combines a novel Principal Component Disturbance Feedback Turbidity Compensation (PC-FDTC Hybrid) with a Stacking-based Variable Selection (SBVS) strategy to enhance COD prediction. The PC-FDTC Hybrid introduces the first principal component as a feedback correction term to improve spectral baseline stability in the 200–400 nm range. Meanwhile, the SBVS strategy integrates CARS, MC-UVE, and Random Frog algorithms via meta-learning to identify robust and informative wavelengths. A total of 646 seawater samples from Jiaozhou Bay, China, were analyzed using a self-developed UV–Vis multi-parameter sensor system. Experimental results show that the proposed SBVS-PLSR model, combined with Savitzky–Golay smoothing and PC-FDTC compensation, achieved superior prediction accuracy (R² = 0.993, RMSEP = 0.340), significantly outperforming traditional full-spectrum and single-method models. Overall, the proposed methodology enhances model robustness, reduces spectral complexity by over 96%, and provides a compact, accurate, and field-deployable solution for real-time COD monitoring in turbid marine environments.