Monthly soil temperature modelling across different depths using machine learning algorithms in northeast India

Soil temperature is one of the most crucial factors for agricultural systems, which affects various hydrogeological processes in soil. Soil temperature is a frequently used variable in machine learning (ML) approaches for modelling those intricate processes. In regions with limited data, such as northeast India, soil temperature is a meteorological variable that is rarely available. The present study aims at soil temperature modelling in three different soil depths (5, 15 and 30 cm), using eight long-term meteorological variables, for example, minimum, mean and maximum air temperatures, rainfall, sunshine hours, minimum and maximum relative humidity, and number of rainy days over the Tochlai region of Jorhat, Assam (northeast India). To accomplish this goal, five suitable ML techniques, for example, Random Forest Regression (RFR), Support Vector Regression (SVR), Boosted Regression Trees (BRT), Classification And Regression Tree (CART), and Bayesian Neural Network (BNN) were employed. Results indicate that the RFR (R² = 0.983; NSE = 0.983) and the BNN (R² = 0.981; NSE = 0.981) outperformed other models in 5 cm soil temperature prediction. In the case of 15 cm soil temperature prediction, the BNN (R² = 0.988; NSE = 0.988) and the RFR (R² = 0.987; NSE = 0.987) performed better than other models. However, the BNN (R² = 0.975; NSE = 0.975) and CART (R² = 0.970; NSE = 0.970) accurately predicted 30 cm soil temperatures in the study region. Based on the analysis for determining the variable importance, air temperature (mean temperature, followed by minimum and maximum temperatures) was found to be the most influential variable for predicting soil temperature. The findings of the study will be beneficial in formulating future research directions, like soil temperature-based crop modelling in data-limited regions and beyond, especially in the absence of multi-depth soil temperature information.