Louis, Miguel Jean. A multi-model ensemble approach for estimating saturated hydraulic conductivity of soils. Retrieved from https://doi.org/doi:10.7282/t3-gxc1-hw91
DescriptionSaturated hydraulic conductivity (Ksat) is an essential parameter for designing irrigation and drainage systems and a required input for hydrological models. Many models using a variety of input properties are available to predict Ksat, but predictions are notoriously poor. In this study, 42 existing models grouped as particle size-, water retention-, and image-based models were used to predict Ksat. Based on their performance, a subset of these models was combined through Bayesian Model Averaging (BMA) to generate multi-model ensembles of Ksat. The BMA technique results in a weighted average of probability density functions that is centered on the individual bias-corrected models. The weights produced by this method are equivalent to posterior probabilities of the models generating the predictions and reflect the models’ relative contributions to the final prediction. This method was applied on a dataset of 170 core samples collected as part of an international project (SOILSPACE) where soil physical parameters (sand, silt, clay, and organic matter contents and bulk density), soil hydraulic properties (Ksat, and water retention data) and 3D images were measured in each of the cores. The dataset was split into training (65%) and testing (35%) subsets. The former subset was used to generate eighteen multi-model combinations from BMA, among which five were selected as the best multi-model combinations. The image-based models accounted for more than 95% of the models constituting the multi-model combinations derived from BMA. This finding emphasizes the importance of the parameters derived from image processing in predicting Ksat. The coefficient of determination (R2) obtained for the best model member was 0.32, whereas for all selected multi-model combinations and the final BMA predictive model, the R2 values were greater than 0.70 and 0.60 for the training and testing subsets, respectively. Overall, predictions from BMA models were better than the best individual performing soil hydraulic model for the entire dataset.