Soil temperature is an important factor in biogeochemical processes. On-site monitoring of soil temperature is limited in spatio-temporal scale as compared to air temperature data inventories due to various management difficulties. Therefore, empirical models were developed by taking 30-year long-term (1985-2014) air and soil temperature data for prediction of soil temperatures at three depths (5, 15, 30 cm) in morning (0636 Indian standard time) and afternoon (1336 Indian standard time) for alluvial soils in lower Indo-Gangetic plain. At 5 cm depth, power and exponential regression models were best fitted for daily data in morning and afternoon, respectively, but it was reverse at 15 cm. However, at 30 cm, exponential models were best fitted for both the times. Regression analysis revealed that in morning for all three depths and in afternoon for 30 cm depth, soil temperatures (daily, weekly, and monthly) could be predicted more efficiently with the help of corresponding mean air temperature than that of maximum and minimum. However, in afternoon, prediction of soil temperature at 5 and 15 cm depths were more precised for all the time intervals when maximum air temperature was used, except for weekly soil temperature at 15 cm, where the use of mean air temperature gave better prediction.