Estimation of soil temperatures with machine learning algorithms—Giresun and Bayburt stations in Turkey

Since soil temperature (ST) is one of the most critical determinants affecting the soil’s physical and chemical properties, the studies on soil temperature estimation increase with the widespread use of deep learning and machine learning algorithms. This study estimates soil temperature at four depths for Giresun and Bayburt stations in Turkey employing the Bayesian Tuned Gaussian Process Regression (BT-GPR), Bayesian Tuned Support Vector Regression (BT-SVR), and Long Short Term Memory (LSTM) models. The stations were selected from semiarid (Bayburt station) and very humid (Giresun station) climates to compare the models’ performance and measure their applicability in different climate classes. Common meteorological indicators were determined as input parameters in the developed models, and a five-and-a-half-year daily dataset was used for all models. This paper represents a novel scheme to optimize the hyperparameters of kernel functions for GPR and SVR models using the Bayesian optimization method to expand predictive efficiency. The developed GPR and SVR models’ outputs are compared with LSTM via three statistical metrics comprising the root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R2). The results show that the BT-GPR model has a superior estimation ability than other developed models for the two stations. The daily ST estimation with the highest accuracy was obtained at a 5-cm depth using BT-GPR at Giresun station (RMSE = 0.0439, R2 = 0.9535, MAE = 0.0344 in the testing phase) and Bayburt station (RMSE = 0.0525, R2 = 0.9438, MAE = 0.0412 in the testing phase). These outcomes provide helpful benchmarking guidance for future soil temperature investigation at various depths across the selected regions.