According to the noise and overlapping characteristics of agricultural irrigation water quality monitoring data for the comprehensive evaluation may bring about the boundary fuzzy problem. This paper proposes an improved Genetic Algorithm (GA) to avoid premature convergence, the global optimal solution of the function of the Projection Pursuit (PP) function is used as the comprehensive evaluation weight coefficient, and then based on Fuzzy Support Vector Machine (FSVM) for comprehensive evaluation. Finally, considering the resolution of the general discrete evaluation grade is low, this paper presents the credibility of the comprehensive evaluation level as a supplement for improve and refine the evaluation results. The results show that the discriminant degree of the comprehensive evaluation grade of the sample is larger than the critical value by using the evaluation model proposed in this paper. The classification result is consistent with the evaluation grade, and the water quality evaluation accuracy is achieved. It shows that the model is reasonable and feasible, and the information of comprehensive evaluation is complete and the accuracy is high.