Soil temperature (ST) is an essential catchment property strongly influenced by air temperature (Ta). ST is also the key factor in sustainable agricultural developments, so researchers are still motivated to develop robust machine learning (ML) models to predict ST more reliably. Four different ML models, utilizing the standalone algorithms (i.e., artificial neural networks: ‘ANN’ and co-active neuro-fuzzy inference systems: ‘CANFIS’) and complementary algorithms (i.e., wavelet transformation combined with ANN: ‘WANN’ and wavelet transformation combined with CANFIS: ‘WCANFIS’) were developed to predict the ST at six meteorological stations incorporating a wide range of climatic features to improve the overall performance. The study has utilized data over the period 2000–2010, collected at 12 locations in Iran. In the first phase of this research, the effects of climate variability on the changes in ST at different depths (i.e., 5, 10, 20, 30, 50 and 100 cm) were explored using air temperature as the exploratory and ST as the response variable. Assessing the performance of the predictive models used in ST prediction, the results indicated good predictive capability of the WCANFIS model, thus, advocating its potential utility in ST prediction problems, especially over diverse climatic regions. This study has also ascertained that the minimum and the maximum predictive errors were encountered at a depth of about 20 cm and 100 cm, respectively. The assessment of climatic features based on air temperature datasets on the performance of the models indicated the highest efficacy demonstrated by the ANN model for the case A–C–W climate type (i.e., a moist climate regime: Arid, temperature regime in winter: Cool, and temperature regime in summer: Warm), in comparison with the PH–C–W climate type (moist regime: Per-humid) for the other best ML models (i.e., WANN, WCANFIS and CANFIS). The order of the model accuracies based on the root mean square error (RMSE) can be ranked wi
