Soil temperature directly influences the organic components and chemical composition of the soil, significantly affecting agricultural practices. The 50 cm depth is particularly critical for wheat, a strategic crop. This study aims to accurately predict daily soil temperature at a depth of 50 cm using minimal meteorological input parameters in the semi-arid regions of Urmia, Mako, and Takab, Iran (2001-2022). The models employed include hybrid multivariate adaptive regression spline with empirical mode decomposition (EMD-MARS) and ensemble empirical mode decomposition (EEMD-MARS), as well as support vector machine combined with gradient boosted tree (SVM-GBT) and stochastic gradient descent (SVM-SGD). The results indicate that the SVM-SGD model achieved the highest accuracy at Takab station (RMSE = 0.781 degrees C), while the SVM-GBT model performed best for Urmia (RMSE = 0.581 degrees C) and Mako (RMSE = 0.834 degrees C). The EEMD-MARS model also demonstrated strong performance, with RMSE values of 0.629, 0.835, and 0.957 degrees C for Urmia, Takab, and Mako stations, respectively. Combined models consistently outperformed hybrid models across all stations. These findings highlight the potential of the SVM-SGD and SVM-GBT models for precise soil temperature prediction, providing valuable insights for improving irrigation scheduling and crop management in semi-arid regions.