Development of air–soil temperature model using computational intelligence paradigms: artificial neural network versus multiple linear regression

Abstract

A multilayer perceptron artificial neural network (MLPNN) model is proposed to estimate the soil temperature from air temperature. The results obtained using MLPNN are compared with those obtained using the multiple linear regression (MLR) using data measured at hourly time step. Firstly, the two models were developed and compared using only air temperature as input variable, and results obtained reveals that soil temperature can be estimated reliably with acceptable accuracy and the MLPNN was slightly more accurate than the MLR model. During the testing phase, the coefficient of correlation R, RMSE and MAE were calculated as 0.834, 7.476 °C, and 5.757 °C for the MLPNN model, while the MLR model provided an R, RMSE and MAE of 0.827, 7.646 °C and 6.039 °C. Secondly, in order to improve the accuracy of the models, the periodicity represented by the hours of the day number, the day of the month, and the month number, were included as input variable in addition to the air temperature. Results obtained demonstrated that the performances of the MLPNN model were significantly improved and the performances of the MLR model were marginally increased. Results demonstrated that MLPNN model is an excellent alternative to the direct measurement of soil temperature.