CARS_SPA optimized UV-Vis spectroscopy for rapid and robust COD prediction in water samples

Ultraviolet-visible (UV-Vis) absorption spectroscopy has been widely adopted in water quality monitoring due to its non-destructive, rapid, and pollution-free characteristics. However, high-dimensional spectral data are often contaminated by system noise and scattering from suspended solids, which adversely affects the accuracy and robustness of chemical oxygen demand (COD) prediction models. To address this issue, this study proposes a hybrid feature wavelength selection method-Competitive Adaptive Reweighted Sampling combined with Successive Projections Algorithm (CARS_SPA)-which performs two-stage screening to effectively eliminate redundant and irrelevant variables. Using real water sample spectra preprocessed by Savitzky-Golay smoothing, the CARS_SPA-LSSVM model significantly outperforms full-spectrum and single-algorithm approaches: on the validation set, it reduces root mean square error (RMSE) by 28.31%, 15.68%, and 14.91% compared to full-spectrum, CARS, and SPA, respectively, achieving a coefficient of determination (R²) of 0.9588. Further optimization of LSSVM hyperparameters using the Dung Beetle Optimizer (DBO) yields an improved RMSE of 0.1525 mg·L^-1 and R² of 0.9664. The results demonstrate that the proposed CARS-SPA-DBO-LSSVM framework achieves high prediction accuracy and robustness with minimal feature dimensionality, offering a promising solution for rapid and reliable COD monitoring in practical water quality applications.