This study pioneers an evaluation of feature selection techniques to enhance soil temperature forecasting, addressing the critical need for accurate environmental predictions. Recognizing the limitations of deep learning models in terms of complexity and interpretability, we explored the efficacy of classical and hybrid machine learning models. Our investigation encompassed Entropy Theory, Stability Selection, and the Gamma Test across multiple datasets. The standout method, Stability Selection, when integrated with the innovative SS_MLP_AdaBelief model, demonstrated significant predictive accuracy, underscoring the importance of evapotranspiration and minimum temperature as key variables. Despite the N-Beats model's limitations, our comparative analysis, visualized through Taylor diagrams, emphasizes the necessity for precise feature selection and the synergistic application of variables and models. This research not only advances the field of soil temperature prediction but also offers valuable insights for future applications, highlighting the potential of methodical feature selection and model integration in overcoming the challenges of traditional deep learning approaches.

Figure: schematic flow chart of MLP-AdaBelief.

Figure: schematic flow chart of MLP-Ranger-AdaBelief.