To predict thermal resistivity accurately, the thermal probe testing in laboratory and data analysis were conducted to briefly investigate the effects of moisture content, dry density, mineralogy, and soil particle morphology on thermal conduction of soils. Based on the artificial neural network(ANN), the prediction models for estimating soil thermal resistivity were proposed. A comparison between models proposed and traditional empirical models was made to verify the reliability and superiority of the ANN models. The analytical results reveal that soil heat transfer performance is influenced by many factors related which would make it difficult to estimate its thermal resistivity. The proposed models based on artificial neural network make it possible to solve this problem well. The ANN individual models including two input parameters(i.e., moisture content and dry density) are suitable for a given type of soil, while the generalized model with four input parameters(i.e., moisture content, dry density, clay content, and quartz content) is not subject to the restriction of soil type. The increment in input parameter can effectively keep the accuracy of the prediction models of their calculation results. The predicted thermal resistivity values from both ANN individual models and generalized model is in good agreement with the measured ones, and the prediction performance of these two models is much better than that of the traditional empirical models. It is shown that the generalized model is preferred in estimation of thermal resistivity of various types of soils which are in obvious discrepancy in engineering properties and depositional environments.