Estimating Soil Temperature at Various Depths in Bangladesh: A Comparative Analysis of Advanced Machine Learning Tree-Based Models

Soil temperature at various depths is vital for hydrology, ecology, agriculture, and engineering, influenced by weather conditions and physical factors. This research utilizes machine learning models with historical data and temporal windowing to predict soil temperature at different depths, effectively addressing nonlinear complexities in environmental studies and meteorology. Utilizing evaluation criteria such as the Root Mean Square Error, Nash-Sutcliffe coefficient, and Pearson Correlation coefficient, it is advisable to employ all four data-based models (Decision Trees, Random Forest, Gradient Boosting Trees, and a hybrid DT-GBT) for accurate soil temperature estimation. The findings reveal that the RF model performed better than others in accurately estimating soil temperature at depths of 10 cm, 20 cm, and 30 cm. Notably, the RF model, followed by the hybrid DT-GBT, GBT, and DT methods, consistently demonstrated accurate predictions across various depths. Overall, the RF model exhibited slightly better performance with significantly faster computation speed compared to other models, making it highly recommended for accurate soil temperature estimation across different depths. These models provide cost-effective alternatives to on-site measurements, delivering valuable benefits to advantages to agriculture. Their accuracy and reliability enhance decision-making processes, thereby improving overall agricultural practices.