Modelling soil temperature at multiple depths in Saurashtra region (Junagadh) of Gujarat using machine learning and shapely approach

Forecasting soil temperature (ST) at multiple depths is crucial for understanding meteorological processes, enhancing agricultural resilience, and assessing ecological and environmental risks. Data driven model represents an alternative tool to the conventional measurement of ST e.g. soil thermometer. To develop the ML model, weekly ST and relevant meteorological variables for the city of Junagadh (Saurashtra region) are collected for the period of 2010–2023. A thorough feature analysis was performed to select the most promising feature using Pearson correlation coefficient and shapely approach. The model was developed using different combinations of input parameters (M1–M7) and trained using different machine learning algorithms. This research aims to evaluate four different machine learning approaches namely, Random Forest (RF), Gradient Boosting Regression (GBR), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM), to predict the soil temperature at 5 cm, 10 cm and 20 cm depth. The result of this study showed that by choosing the optimum input parameter, there is no significant impact on accuracy of model. The best performance was obtained for Model 7 f(TDB, TMax, TMin, Evapo) model at the 10-cm soil depth, as it provided the greatest correlation coefficient (r = 0.9967) and the lowest value for root mean square error (RMSE = 0.3410 °C) and percent bias (PBIAS = − 0.0115). The result showed that model performance differences are often statistically significant, especially at shallower depths (ST5, ST10), but less so at ST20. In the current study, besides evaluating the potential of four machine learning models, the interpretation of the machine learning algorithm for soil temperature prediction was explored using SHapley Additive exPlanations (SHAP). The study used an explainable artificial intelligence (XAI) approach to provide novel interpretation and insights to elucidate model formulation and relative predictor importance.