Wavelet neural networks and gene expression programming models to predict short-term soil temperature at different depths

Abstract

Soil temperature (ST), as one of the critical meteorological parameters, has great effects on many underground soil ecological processes. Due to the fact that accurate measuring of ST is costly because of launching field equipment, evolving predictive models to approximate ST is of great importance. Therefore, achieving accurate, reliable and easily attainable predictions of daily ST values is the main objective of the current research. To that end, the usefulness of three data-driven procedures containing artificial neural networks (ANN), wavelet neural networks (WNN) and gene expression programming (GEP) were examined for the estimation of ST at different soil depths at Tabriz synoptic station, north-west of Iran. In conformity with the correlation coefficients among ST and meteorological parameters, it was found that air temperature, Sunshine hours and radiation had the most and unquestionable effects on ST prediction at all considered depths. For evaluating the performance of these approaches, four different statistical error measures were used: coefficient of correlation (CC), mean absolute error (MAE), root mean squared error (RMSE) and Akaike's information criterion (AIC). Moreover, Taylor diagrams were employed for assessing the similarity between the observed and predicted ST values. Results revealed that the WNN in all considered depths had the best performance in ST prediction, but with increasing soil depth, the effect of meteorological parameters and estimation accuracy were reduced rapidly. As a conclusion, the lower values of RMSE and higher values of CC proved the effectiveness of WNN for predicting ST at the studied depths.