In the water quality measurement based on UV-Visible spectroscopy, the spectral signal is easily disturbed by system noise and the scattering of suspended solids, and there are information redundancy, multicollinearity and other characteristics, resulting in a large deviation in the selection of characteristic wavelength in the COD measurement of water quality. Therefore, this paper proposes an optimization algorithm of water COD detection characteristic wavelength based on embedded particle swarm genetic (EPSO_GA) algorithm to improve the accuracy of wavelength selection. In order to verify the feasibility of the optimization algorithm for detecting characteristic wavelength, spectral data of water samples from a university pond, domestic sewage and drainage ditch were collected, and EPSO_GA algorithm was used to select characteristic wavelength from the pre-processed spectral data. EPSO_GA algorithm adopts the real coding method to realize unified coding of particle swarm optimization (PSO) algorithm and genetic optimization (GA) algorithm. The operations of selection, crossover, and mutation of the GA algorithm are embedded when the particles are updated in the PSO algorithm, which improves the limitations of these two algorithms in spectral wavelength feature selection. The characteristic wavelength selected by EPSO_GA algorithm was combined with the partial least square method (PLS) to construct the water COD prediction model of EPSO_GA_PLS, and compared with the traditional PSO algorithm and the GA algorithm, the PSO_PLS, GA_PLS established by the characteristic wavelength and the PLS water quality COD prediction model constructed by the full spectrum are compared. The results showed that compared with PSO_PLS, GA_PLS and the PLS water quality COD prediction model constructed by full-spectrum, EPSO_GA improves the precocious and slow convergence speed of PSO and GA in spectral characteristic wavelength selection, and reduces the complexity of constructing PLS water quality COD prediction model in the whole spectrum, the prediction accuracy of the model is improved. In the EPSO_GA_PLS water quality COD prediction model established based on EPSO_GA algorithm, the root-mean-square error decreased to 0.212 3, and the prediction accuracy increased to 0.999 3, which can quickly and quantitatively detect water quality COD, provides a better prediction model for COD measurement by UV-Visible spectroscopy.