Estimating Soil Temperature from Metrological Data Using Extreme Learning Machine, Artificial Neural Network and Multiple Linear Regression Models

Soil temperature (Ts) is a key factor controlling the soil physical, chemical and biological properties and processes and consequently affects agricultural crop productions. The objective of this study was to estimate Ts from meteorological data using different machine learning methods. For this purpose, meteorological data and soil temperature at different depths (5, 10, 20, 30, 50 and 100 cm from the soil surface) for 25 years (1994-2018) were collected from 17 synoptic stations in Khuzestan province, Iran. Air temperature, wind speed, relative humidity, evaporation, precipitation, and vapor pressure were used as inputs to train the models. Multiple Linear Regression (MLR), Multilayer Perceptron Neural Network (MLPNN) and Extreme Learning Machine (ELM) models were used to predict soil temperature from metrological data. The results indicated that all models predicted temperature of the top layer (0-30 cm) better than the ones in sublayers. on the other hand, by increasing soil depth the accuracy of the models diminished; so that, the best and worst Ts predictions were belong to 5 cm and 100 cm depth, respectively. The results revealed that MLR, MLPNN and ELM models provided desirable performance in modeling Ts at all depths, with R2 values ranging 0.700-0.864, 0.967-0.997, and 0.967-99, RMSE values ranging 2.557–2.873, 0.034–0.072, and 0.028–0.078 °C, and MAE values ranging 1.398–1.529, 0.023–0.063, and 0.023–0.065 °C, respectively. Overall, the results showed that MLPNN and ELM models had approximately similar performance and better accuracy than MLR model. However, because of the high computational speed of the ELM model, it is recommended to use MLPNN model for estimation of soil profile Ts.