Comparative Analysis of Machine Learning and Deep Learning Approaches for Hourly Soil Temperature Estimation At Multiple Depths from Meteorological Data in A Semi-Arid Region of Burkina Faso

Accurate estimation of soil temperature is essential for understanding land–atmosphere interactions and managing agricultural systems, par-‎particularly in semi-arid regions. This study evaluates and compares the performance of multiple modeling approaches: machine learning (Random Forest, XGBoost, LightGBM, SVR, GPR, Decision Tree), and deep learning (MLP-ANN, LSTM, ConvLSTM, Hybrid CNN–‎LSTM) to predict hourly soil temperature at depths of 10, 20, 30, and 40 cm in a semi-arid land from Burkina Faso. Using high-frequency ‎meteorological data collected, the models were assessed based on their accuracy, temporal generalization, and variable importance. Results ‎show that among ML models, Random Forest demonstrates strong performance up to 30 cm. Deep learning models, particularly ConvLSTM and Hybrid CNN–LSTM, outperform all others across depths, capturing the spatio-temporal variability of subsurface temperature ‎with high fidelity. SHAP analysis reveals that air temperature (Ta) and solar radiation (RAD) dominate at the surface, while dew point tem-‎temperature (DPT) and vapor pressure deficit (VPD) gain importance in deeper layers. Variables such as rainfall, wind direction, and relative ‎humidity show minimal contribution and can be excluded without sacrificing performance‎.