Estimating parameters of empirical infiltration models from the global dataset using machine learning

It is beneficial to develop pedotransfer relationships to estimate infiltration equation coefficients in site-specific conditions from readily available data. No systematic studies have been published concerning the relationships between the accuracy of the infiltration equation and the accuracy of the predicted coefficients in this equation. The objective of this work was to test the hypothesis that, for the same infiltration data, the accuracy of pedotransfer predictions for coefficients in an infiltration equation is greater for the infiltration equation that performs better. The hypothesis was tested using the commonly employed Horton and Mezencev (modified Kostiakov) infiltration equations with data from the Soil Water Infiltration Global database. The random forest machine learning algorithm was used to develop the pedotransfer model. The Horton and the Mezencev models performed better with 928 and 758 datasets, respectively. The accuracy of the estimates of the infiltration equation coefficients did not differ substantially between the estimates obtained from all data and from the data where the infiltration equation had lower root-mean-squared error values. The root-mean-squared error values of the pedotransfer estimates decreased by 2 to 25% when only datasets with the same infiltration measurement method were considered. The development of predictive pedotransfer equations with the data obtained from the same infiltration measurement method is recommended.