The temperature prediction by a commonly used force-restore model is compared with the Oklahoma Atmospheric Surface-layer Instrumentation System Project (OASIS) soil temperature data. A significant drift is found in the deep layer temperature prediction in the first one to two days and the drift in the summer is of different sign as in winter. The neglect of the difference in seasonal-mean temperature between the skin layer and deep layer is identified as the cause and the prediction equations of the force-restore model are re-derived taking into account of the temperature change of seasonal-mean temperature with depth. The drift in the forecast temperature is thereby removed and a good agreement is found between the prediction of the improved model and the observations calibrated to the right depths. The improved equations for temperature forecast are implemented in a two-layer soil-vegetation model in the ARPS system. Additional modifications are also made to the soil moisture equations and improvements in the soil moisture prediction are also obtained.