The research of soil classification and soil grade evaluation is often based on fuzzy theory. So, the traditional method has an inevitable problem about weight matrix which given by some experts, and the final result can be influenced by artificial factors. The essentials of fuzzy synthetically judge is to handle the data of high dimension. That is to reducing the dimension number. The weight matrix in fuzzy theory is corresponding to low dimension projection value of each index. But we can′ t define whether the weight matrix given by experts is the best projection value or not. So, the authors apply a new technique of falling dimension named projection pursuit to soil study, through using the improved real coding based accelerating genetic algorithm to optimize the projection direction. Thus, it can transfer multi-dimension data into one dimension data, through searching for the optimum projection direction to realize the soil classification and its grade evaluation. The method can avoid the artificial disturbance, and acquire preferably effect. Thus, the paper provides a new method to the research of soil classification and grade evaluation.