Soil temperature is an important variable in Earth sciences. The temporal and spatial variations in soil temperature are affected by numerous factors, resulting in various challenges in soil temperature prediction. For soil temperature prediction, the data-driven machine learning method is valuable and can be an important complement to physics-based process models. However, no extensive studies have been carried out on the importance of environmental factors on soil temperature. In this study, a data-driven XGBoost-LSTM method is proposed. The weights of the meteorological inputs are computed based on XGBoost, and then, the combination of meteorological inputs based on their weights is applied to obtain an optimal model by the LSTM method. An experiment is carried out at two stations in China (Changbai Mountain and Haibei). The most accurate performance for soil temperature estimation is attained, with highest values of NS = 0.932, WI = 0.983, and LMI = 0.729 and lowest values of RMSE and MAE of 2.234 and 1.716, respectively. These results show that the proposed model is generally superior to other state-of-the-art predictive models.