Soil temperature is an important meteorological parameter which determines the rates of physical, chemical, and biological reactions in the soil. However, measured values are very sparse in space and time and often not available for a given site. In this study, two intelligent neural models including artificial neural networks (ANNs) and co-active neurofuzzy inference system (CANFIS) were used for the estimation of soil temperatures at six depths (5, 10, 20, 30, 50, and 100 cm) with minimum input data (mean air temperature). For this purpose, use was made of the 14-year meteorological data obtained for the two regions of Gorgan in northern Iran with a humid climate and Zabol in southeastern Iran with a dry climate. Comparisons of the model performances in arid and humid regions showed that both ANNs and CANFIS models performed better in arid regions. The accuracy of the soil temperature predictions by both ANNs and CANFIS models gradually decreased from the surface down to the various depths. The results also indicated the capabilities of the ANNs in predicting soil temperature in arid and humid regions.
