As an important indicator of soil characteristics, soil temperature has a great impact on agricultural production, building energy savings and for shallow geothermal applications. Data-driven models have been developed and achieved good accuracy in monthly soil temperatures or daily soil temperatures predictions of a single site taking air temperature, solar radiant and time as inputs. Models’ accuracy obviously dropped if they are applied for predicting daily soil temperatures for various climates on a continental scale. We proposed a new data-driven model based on the support vector machine (SVM). The new model considers daily soil temperature variations as superposition of annual average ground temperatures predictions (long-term climates impact) and daily ground temperature amplitude predictions (short-term climates impact). Annual average soil temperature are determined by air temperature, solar radiant, wind speed and relative humidity; daily soil temperature amplitudes by air temperature amplitudes, solar radiant and day of year. For daily soil temperature predictions at 16 sites located in arid or dry summer climates, warm climates and snow climates in United States, the new model's mean absolute error is 1.26 °C and root mean square error is 1.66 °C. Meanwhile, traditional SVM model's mean absolute error is 2.20 °C and root mean square error is 2.91 °C.