Exploring vegetation-driven microclimatic effects on soil temperature dynamics in tropical climates through machine learning and explainable AI

Soil temperature (ST) is a crucial parameter in tropical environments, influencing microbial activity, nutrient cycling, and root growth. Accurate and cost-effective prediction of ST is essential for understanding soil health and supporting resilient ecosystems. This study investigates ST dynamics across four different urban tropical microclimates within a 150-square-meter area over a four-month period, utilizing a dataset comprising 5,856 observations collected in a tropical setting. Advanced machine learning modelling, including Random Forest Regressor (RFR), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Multilayer Perceptron (MLP), coupled with Explainable Artificial Intelligence (XAI) was employed. Results reveal that the interplay of solar azimuth and vegetation cover governs ST variations. The XGBoost outperformed all other machine learning models, exhibiting the most accurate predictions and resulting root mean square error (RMSE) values of 0.298 ± 0.008ºC for modelling ST at 10 cm depth, 0.117 ± 0.006ºC at 30 cm and 0.064 ± 0.002ºC at 50 cm. XAI analysis highlighted air temperature as the dominant predictor of ST at 10 cm, while deeper layers were influenced by temperature of the overlaying soil layer, followed by solar radiation and soil water content. These findings emphasize the potential of integrating machine learning (ML) and XAI for explicit and reliable ST prediction and advancing plant growth.