Abstract: In order to realize precision management of winter wheat, two prediction models of winter wheat yield based on soil parameters were proposed and compared. The field tests were carried out in 2008 and 2009. The variety of the experimental winter wheat was Jingdong 12, and the test area was divided into 60 zones with 5m×5m grids. The sampling point was put in the middle of the zone, and the depth of the sampling point was 5cm. Soil EC was measured by a DDB-307 EC meter, and the winter wheat yield data were provided by a CASE2366 grain harvester with GPS receiver. Gray theory were used to analyze the gray relation between soil EC value and each of other soil parameters, total nitrogen content, K+、NO3- and pH of soil. Results showed that there were high gray relation between soil EC and total nitrogen content, K+, pH of soil. Since soil organic horizons had high correlation with soil negative charge capacity, when soil had more organic horizons, there would be more soil negative ions, and the soil EC would be higher. Hence, the gray relation between K+ and EC was high. Using nitrogen fertilizer could removal caption from soil, and increase the content of K＋，Na＋，Ca2＋and Mg2＋,so that there were also high correlation between total nitrogen content and EC. The reason of high correlation between EC and soil pH was attributed to that the change of pH had influence on negative charge. After analyzing the correlation between winter wheat yield and soil EC, total nitrogen content, K+, NO3- , pH of soil in different growth period, two yield prediction algorithms of back propagation neural network (BPNN) and fuzzy least square-support vector machine (FLSSVM)were proposed. BPNN prediction model took soil EC, total nitrogen content and K+ as input and winter wheat yield as output. While FLSSVM prediction model took soil EC, nitrogen content, K+ and gray relation as input and also winter wheat yield as output. Results showed that the prediction and validation R2 of BPNN model were 0.8237 and 0.7367, respectively. Prediction R2 of FLSSVM was 0.8625, and validation R2 of FLSSVM model was 0.8003. The advantage of BPNN was fast training speed, while the disadvantage was weak generalization ability of network. FLSSVM used fuzz similar extent to fuzz input samples so that it could avoid training too much. Also because it was based on membership function, it could have several advantages such as simple structure, efficient convergence, precise forecasting, and etc. Both BPNN and FLSSVM had high accuracy prediction result and could be used in estimating yield and providing theory and technical support for precision management of crops. But the default of the FLSSVM is the portability of the model is bad, so it is still need to be improved in practical applications.