Comparative analysis of seven machine learning algorithms and five empirical models to estimate soil thermal conductivity

Soil thermal conductivity (λ) is an important property that crucial for surface energy balance and water studies. 1602 measured soil values representing 189 soils were used to evaluate five empirical models (i.e., de Vries (1963) model (de 1963), Campbell (1985) (Campbell1985), Johansen (1975) (Johansen 1975), Côté Konrad (2005) (Côté 2005), Lu et al. (2007) (Lu 2007)) seven machine learning (ML) algorithms Decision Tree (DT), Random Forest (RF), Gradient Boosting (GBDT), Linear Regression (LR), K-Nearest Neighbors (KNN), Neural Network (NN), Gaussian Process (GP)) estimate λ. Our results demonstrated the average root mean squared error (RMSE) of ML 66% 82% on validation test sets respectively. The three best (GBDT, NN, RF) performed significantly better than 2007, 2005, 1975): 0.183 < RMSE 0.259 (W m−1 K−1) 0.293 0.320 models. For ML, we recommend GBDT, NN RF algorithms. models, use normalized 1975) over physically-based (DV1963) regression (CG1985). feature importance rankings by GBDT show moisture content bulk density are most critical factors affecting together account more 80% influence value gives consistent ranking therefore, selecting features.