This study proposes a method for detecting chemical oxygen demand (COD) in water using fluorescence emission spectroscopy. The spectral data were preprocessed using multiple techniques, including multiplicative scatter correction (MSC), first-order derivative, standard normal variate transformation (SNV), min-max normalization, and Savitzky-Golay smoothing. Key feature bands were selected using backward interval partial least squares (BiPLS) and synergy interval partial least squares (SiPLS). A prediction model was then developed using partial least squares (PLS) to improve spectral processing efficiency and prediction accuracy. Experimental results demonstrated that Savitzky-Golay smoothing provided the best preprocessing performance, while BiPLS showed superior selectivity for feature extraction. At an excitation wavelength (Ex) of 310 nm, the PLS model, optimized by combining Savitzky-Golay smoothing and BiPLS feature extraction, achieved optimal performance, with the validation set correlation coefficient (rp) of 0.9191, the root mean square error of prediction (RMSEP) of 3.3488 mg/L, and the prediction Bias of -0.2835 mg/L. This method offers a practical approach for rapid COD detection in water quality assessment.