A boosting algorithms approach to thermal and physico-chemical properties of contaminated sandy loam soils
摘要
To effectively plan, manage and use soil resources, there is a need to evaluate its critical properties such as thermal conductivity and physico-chemical properties. These properties play a major role in agriculture, land surface studies, surface energy balance and geothermal energy research, however obtaining them takes several time and efforts. While there has been growing interest in rapid and precise measurement of soil thermal conductivity (λs) and specific heat capacity (Cs) using machine learning algorithms, creating a universal model applicable across diverse soil types remains challenging due to data limitations. Furthermore, this limitation has restricted the accuracy of parametric models. To address this data gap, Conditional Generative Adversarial Networks were used to augment sandy loam soils measurements such as specific soil physico-chemical characteristics, oil and grease concentration, and thermophysical parameters from two representative car mechanic villages (MV) on two distinct geological formations (Basement Complex and Sedimentary formations) in Ogun State, Nigeria. A comparative analysis was then conducted among ten Ensemble machine learning models such as Adaptive boosting (Adaboost), Gradient boosting (GBDT), Light gradient boosting (Lightgbm), Categoricla boosting (Catboost), and Extreme gradient boosting (XGboost) amongst others. The multilayer perception and random forest outperformed other models in simulating both λs and Cs, demonstrating high accuracy and generalization. Thermal diffusivity (TD) and specific heat capacity emerged as the most influential parameters for predicting λs and for specific heat capacity (Cs), porosity and pH. However, thermal admittance, thermal resistivity, pH, oil and grease content—were also critical factors. This study offers valuable data on soil thermal properties and provides a parameterization framework for land surface process research.