According to the missing data problem in the ground temperature observation, as well as national ordinary meteorological observing Station has no deep ground temperature observation service, realtime ground temperature deduction models and deep ground temperature deduction models (40～160 cm ground temperature deduction model & 320 cm ground temperature deduction model) by the Back Propagation(BP) neural network is proposed. The former can be used to fill in the missing data of ground temperature observation data, and the latter can be used to estimate the deep ground temperature data in the area without no deep ground temperature observation. The BP neural network is trained by using a small number of samples and tested with all the data of the sample station. The neural network parameters are adjusted repeatedly, and the models with good error performance are selected. And then the output error of the ground temperature model is tested by using the contrast station data. The accuracy rate of realtime ground temperature deduction model in the sample station is 77.705% as well as in the contrast station is 66.168%. More than 72% of the output error of the 40～160 cm ground temperature deduction model is less than 0.5 ℃, and more than 83% of the output error of the 320 cm ground temperature deduction model is less than 1℃. The experimental results show that the temperature deduction model established by this method has high precision and practicability.