A novel method based on multi-source spectral characteristics of the combination is proposed for chemical oxygen demand detection. First, the ultraviolet and near infrared spectrum of the actual water samples are collected respectively. After pretreatment of the spectrum data, the features of the spectrum are extracted by the nonnegative matrix factorization algorithm for training after normalization. Particle swarm and least squares support vector machines algorithm are applied to predicting chemical oxygen demand of the validation set of water samples. The effect of spectrum’s base number on the predicted results is discussed. The experimental results show that the best base number of the ultraviolet spectrum is 5, the best base number of the near infrared spectrum is 2; The validation set correlation coefficient of the prediction model is 0.999 8, and the root mean square error of prediction is 3.26 mg·L-1. Experimental results demonstrate that the nonnegative matrix factorization algorithm is more suitable for feature extraction of spectral data, and the least squares support vector machines algorithm as a quantitative model correction method of the actual water samples can get good prediction accuracy with different feature extraction methods(principal component analysis, independent component analysis), spectroscopic methods(ultraviolet spectrum method, near infrared spectrum method) and different combination pattern (data direct combination, combining data first, then feature extraction) respectively.