Noise‐Assisted Multivariate Empirical Mode Decomposition of Saturated Hydraulic Conductivity along a South‐North Transect across the Loess Plateau of China

Abstract

Core Ideas K s and eight associated factors were decomposed into different scales using NA‐MEMD. Small‐scale variations of K s were dominated by soil properties, especially bulk density. Large‐scale K s variations were mainly controlled by topographic and climatic factors. Summing up MLR estimates of K s scale components surpassed MLR estimation of K s before NA‐MEMD. Saturated hydraulic conductivity ( K s ) usually varies at multiple scales in space, as affected by different soil and environmental processes operating at diverse scales. Identifying spatial process relationships can be challenging due to overlapping of underlying processes at different scales. The objective of this study was to evaluate noise‐assisted multivariate empirical mode decomposition (NA‐MEMD) for characterizing K s variability and depicting its scale‐dependent relationships with different soil properties and environmental factors. At an interval of 10 km along an 860‐km south‐north transect across the Loess Plateau of China, K s , bulk density, soil organic carbon content, sand and clay contents at three depths of 0 to 10, 10 to 20 and 20 to 40 cm were investigated as well as four environmental factors of elevation, slope gradient, annual precipitation and temperature. Decomposed into different intrinsic mode functions (IMFs) and residues by NA‐MEMD, K s at all depths were found to vary at the smallest scale of 29 km mainly, corresponding to IMF1s, which manifested 33.0 to 48.1% of the total K s variance. The small‐scale variations of K s reflected not only in IMF1s but also in IMF2s and IMF3s were dominated by soil properties especially bulk density, and the large‐scale variations corresponding to IMF4s and IMF5s were controlled by environmental factors in general. For each depth, K s at the scale of investigation was estimated by adding all the IMFs and residue derived from the factor components at equivalent scales using multiple linear regression (MLR). Such K s estimations after NA‐MEMD evidently outperformed the MLR before NA‐MEMD by explaining additional 9.4 to 18.7% of the total K s variance, but underperformed the artificial neural network and state‐space approach also implemented on undecomposed spatial series of K s and its underlying factors. NA‐MEMD serves as a useful tool for K s characterization and its incorporation with nonlinear functions or spatial interactions with impact factors is suggested for the estimation of K s and other soil processes.