Estimating parameters for the Kostiakov-Lewis infiltration model from soil physical properties

Abstract

Infiltration modeling is an important tool to describe the process of water infiltration in the soil. However, direct measurements of the parameters of infiltration models are usually time-consuming and laborious. The present study proposed an effective method to estimate parameters of the Kostiakov-Lewis model (a classical infiltration model) from soil physical properties (SPPs). Parameters k, α, and f0 of the Kostiakov-Lewis infiltration models were measured in 240 double-ring field experiments in Shanxi Province, China. SPPs at the corresponding experimental points were measured at the topsoil layer (TL, 0–20 cm) and the top-subsoil layer (TSL, 0–20 and 20–40 cm). The Kennard-Stone (KS) sampling method and principal component analysis (PCA) were used for dividing training samples and extracting principal components (PCs) of SPPs, respectively. Partial least squares (PLS), back-propagation neural networks (BPNNs), and a support vector machine (SVM) were used to establish models for estimating k, α, and f0 with the SPPs of TL and TSL as the input variables (IV). The differences in soil density (BD), texture, and moisture content (θv) were found in topsoil and subsoil, but loading distributions of SPPs on PCs present different degrees of correlation. Moreover, SVM produced the most accurate estimation among these three methods for using the SPP of TL and TSL as inputs. The highest accuracy for k estimations was obtained by SVM using the SPP of TL as IV; R and RMSE in the model test process were 0.78 and 0.3 cm min−1, respectively. However, using SPP of TSL as IV obtained the highest accuracy for both α and f0 estimations with the SVM method (R values were 0.71 and 0.82, respectively, and RMSE values were 0.03 and 0.018 cm min−1) in the model testing. The SVM method with SPPs as inputs is an effective and practical method for estimating the parameters of the Kostiakov-Lewis infiltration model.