Abstract: The study aims to build LM-BP network structure (5-M-1) model by the BP neural network method, so as to achieve the division of soil nutrient levels and provide reliable basis for optimum soil nutrient management. Levenberg-Marquardt (LM) training algorithm was used to construct a three-layer network model (input, hidden, and output layer), which was used for soil nutrient evaluation. Criteria for soil nutrient evaluation were used as model samples to train and test BP neural network, and comprehensive evaluation of soil nutrient levels in Shexian County was made. The results showed that the output predicted value by LM-BP network structure coincided with the actual reference value. Comprehensive evaluation result of soil nutrient levels in Shexian County which based on gray correlation model and principal component analysis was basically the same with the simulation result of BP neural network. LM-BP network structure achieved favorable predicted results in its application to the grading of soil nutrient levels, which could provide a solid foundation for the application of intelligence algorithm in agriculture.