Estimation of Surface and Depth Soil Temperature from Meteorological Data Using Machine Learning Techniques in Hyper Arid Climate

Accurate estimation of temperature at various soil depths is crucial for land-atmosphere interactions. In this study, the application of six different machine learning models including artificial neural network (ANN), decision tree (DT), cubist (CB), random forest (RF), support vector machine (SVM) and linear regression (LR) for modeling of daily soil temperature was studied at six different depths of 5, 10, 20, 30, 50 and 100 cm in Kerman. A set of accessible meteorological data including maximum and minimum temperatures, relative humidity, dew point, evapotranspiration and atmospheric pressure were used as input to the models. The degree of importance and correlational analysis was performed for the input variables based on the data of the 18-year statistical period. According to the results, the performance of all six models based on evaluation criteria (R2 >0.86, RMSE <2.8 ◦c and Bias <0.14 ◦c) was acceptable at all depths. However, RF, ANN and SVM showed very high efficiency in estimating soil temperature (R2 <0.97). Also, the DT model and then the LR model performed lower than the others. Examination of the importance of variables showed that among the input parameters, maximum and minimum temperature had the greatest effect on predicting soil temperature in all models. Finally, it can be safely acknowledged that machine learning models such as random forest, artificial neural network and support vector machine have the ability to estimate surface and depth soil temperatures in arid climates in the absence of measuring equipment. A set of meteorological data including maximum and minimum temperature, relative humidity, dew point, evapotranspiration and atmospheric pressure were used as input to the models.