The identification of water quality is an important prerequisite for accurate spectroscopic detection of water quality parameters. Aiming at the problem of large redundancy of spectral data collected by direct spectrum water quality detection system, the principal component analysis is used to eliminate the correlation of information indexes, and the spectral data is reduced and the feature information is extracted. The UV-Vis spectra of water from a chemical plant and a stream were collected. The discriminant model was established by using the method of principal component analysis and Fisher discriminant. First, 12 sets of water samples were used as training samples and 6 groups as test samples. Then, the discriminant ability of the model was demonstrated and tested, and compared with the traditional Fisher discriminant model. Finally, The experimental results show that the joint Fisher discriminant model can effectively eliminate the influence of information redundancy. Compared with the traditional Fisher discriminant model, it has the advantages of high classification precision, zero return error rate and short calculation time. The calculation time is reduced from 0.6733 s to 0.6012 s. This method provides an efficient means for practical application of direct spectrum method to determine the water quality.