It is extremely important to obtain accurate information of soil moisture and understand the dynamic change pattern of soil moisture. A support vector machine algorithm（SVM） was introduced for fitting analysis and regression training optimization of ultrasonic velocity-soil moisture under four different treatments of alternating wet and dry， and a prediction model of the soil moisture based on ultrasonic velocity was constructed. The prediction model was used to estimate the water content of soil in farmland tillage layer with different moisture requirements. The results showed that the average relative error of verifying soil moisture in the field with the model constructed was about 1.5% compared with the traditional drying method. It is indicated that the prediction model for ultrasonic velocity-soil moisture based on SVM model can effectively describe the soil moisture in the area studied. It will provide reference for utilizing ultrasonic characteristics to achieve continuous monitoring of soil moisture in farmland.