Soil temperature (Ts) plays a key role in physical, biological and chemical processes in terrestrial ecosystems. Accurate estimation of Ts at various soil depths is crucial for land-atmosphere interactions. This study investigated the applicability of four different machine learning models, extreme learning machine (ELM), generalized regression neural networks (GRNN), backpropagation neural networks (BPNN) and random forests (RF), for modeling half-hourly Ts at four different depths of 2 cm, 5 cm, 10 cm, and 20 cm on the Loess Plateau of China. A field experiment was conducted to measure half-hourly Ts and meteorological variables. Air temperature, wind speed, relative humidity, solar radiation, and vapor pressure deficit were used as inputs to train the models for estimation of half-hourly Ts. The results showed ELM, GRNN, BPNN and RF models provided desirable performance in modeling half-hourly Ts at all depths, with root mean square error values ranging 2.26–2.95, 2.36–3.10, 2.32–3.04 and 2.31–3.00 °C, mean absolute error values ranging 1.76–2.26, 1.83–2.31, 1.80–2.32 and 1.79–2.26 °C, Nash–Sutcliffe coefficient values ranging 0.856–0.930, 0.841–0.924, 0.847–0.927 and 0.850–0.927, and concordance correlation coefficient values ranging 0.925–0.965, 0.925–0.963, 0.928–0.963, and 0.924–0.961 for the ELM, GRNN, BPNN, and RF models, respectively. There was a statistically significant agreement (P < 0.001) between the measured and modeled values at both half-hour and daily timescales, and the box plots showed the distributional differences between the measured and modeled values were small. Generally, the ELM model had slightly better performance with much better computation speed than GRNN, BPNN as well as RF models at half-hourly timescales, thus the ELM model was highly recommended to estimate Ts at different soil depths.