To conduct green, rapid and accurate detection of organic pollutants in water, the current paper proposes a detection method of Chemical Oxygen Demand (COD) based on fluorescence multi-spectral fusion. The experimental samples consist of 53 actual water samples including inshore seawater and surface water. The physicochemical valuesof COD of the samples are obtained by standard chemical methods. A fluorescence spectrophotometer is used to collect the three-dimensional fluorescence spectra of the samples, and the spectral data are processed and modeled. The three-dimensional fluorescence spectrum is spread at the excitation wavelength in the excitation wavelength range of 200～330 nm and the emission wavelength range of 250～500 nm, with excitation wavelength interval being 5 nm, and the emission wavelength interval 2 nm. With the ant colony optimization-interval partial least squares (ACO-iPLS) as the feature extraction algorithm and the least squares support vector machine algorithm optimized by particle swarm optimization (PSO-LSSVM) as the modeling method, the prediction model of fluorescence emission spectral data at a single excitation wavelength, the fluorescence multi-spectral data-level fusion (Low-Level Data Fusion, LLDF) model and the fluorescence multi-spectral feature-level fusion (Mid-Level Data Fusion, MLDF) model are built respectively, and the prediction effects of various models are compared. The results show that there exist some differences in the prediction effect of the models for the fluorescence emission spectrum data at different excitation wavelengths. The prediction model of the fluorescence emission spectrum data at the excitation wavelength of 265 nm is optimal, with determinant coefficient (R２ｐ) and the root mean square error in prediction (RMSEP) of the calibration set being 0.990 1 and 1.198 6 mg·L-1 respectively. For fluorescence multi-spectral data-level fusion models, fluorescence emission spectra at excitation wavelengths of 235, 265, and 290 nm (abbreviated as: LLDF-PSO-LSSVM) have the best prediction effect, with the results of R2p and RMSEP being 0.992 2 and 1.055 1 mg·L-1 respectively. For fluorescence multi-spectral feature-level fusion models, fluorescence emission spectra at excitation wavelengths of 265, 290, and 305 nm (abbreviated as: MLDF-PSO-LSSVM) have the best prediction effect, with the R2ｐbeing 0.998 2 and the RMSEP being 0.534 2 mg·L-1. A comprehensive comparison of various modeling results shows that the model of MLDF-PSO-LSSVM has the best performance, indicating that the multi-spectral feature-level fusion model based on fluorescence emission spectrum data is more accurate and more effective for predicting COD of water quality.