Using different multimodel ensemble approaches to simulate soil moisture in a forest site with six traditional pedotransfer functions

Highlights

• Different multimodel ensemble approaches were used to simulate soil moisture contents.

• The simple model average method always had worse performance than the best single model.

• The performance of the superensemble approach was better than that of the best single model.

Abstract

Pedotransfer functions (PTFs) have routinely been used to estimate the soil hydraulic properties (SHPs) from easily measurable soil properties, such as particle-size distribution, organic matter content and bulk density. However, different PTFs often yielded different prediction results. In order to deal with the PTF selection problem, this study used multimodel ensemble approaches to simulate forest soil moisture based on the modelling results of different PTFs. A total of 300 days of observed soil moisture data at four depths (10-, 20-, 40- and 60-cm) were adopted to calibrate the Richards equation and obtain the SHPs by using the inverse option in HYDRUS-1D. Six published PTFs were selected to predict the SHPs, which were used to predict soil moisture temporal variations at these four different depths. Two multimodel ensemble methods, including the simple model average (SMA) and the multiple linear regression (MLR)-based superensemble, were used in this study. Under different selections of training periods (i.e. 50, 100 and 150 days), performances of these multimodel ensemble approaches were compared with those of the best single PTF model. The SMA always had worse performance than the best single model. However, the performances of the superensemble approach were better than those of the best single model, and even comparable to those of the calibrated soil water flow model. Results show that given the relatively long training period (>50 days), it is worthwhile to consider the superensemble method to simulate soil moisture contents in forestland.

Introduction

The forest hydrological research plays a critical role in the alternative land-use practices (e.g. restoring farmlands to forests) and the management of forests (Aranda et al., 2012). Mathematical simulation models have often been applied to hydrological processes in forest soils (Nosetto et al., 2012). Simulations of the soil moisture dynamics need soil hydraulic properties (SHPs) (i.e. soil water retention characteristics and saturated hydraulic conductivity). However, SHPs are often unavailable for large scale modelling since they are hard to be measured or calibrated with good spatial resolution. Therefore, pedotransfer functions (PTFs) have been adopted to predict SHPs from more easily measurable soil properties, such as particle-size distribution, organic matter content and bulk density (Rawls and Brakensiek, 1985). During the past thirty years, the multiple linear regression (MLR) (Saxton et al., 1986) and artificial neural networks (ANN) (Schaap et al., 1998) were routinely used to develop the PTFs from large datasets in different regions. Evaluation of PTFs is not only to compare the measured and estimated SHPs, but also to test the application of the predicted SHPs in simulating soil moisture content in the unsaturated zone. Many studies have investigated different PTFs in soil water simulation at various scales (Chirico et al., 2010, Cichota et al., 2013). Although PTFs can be well used to simulate soil water variations, the empirical nature of the MLR and ANN techniques virtually introduces the uncertainties since these PTFs were used outside of the datasets they developed. Therefore, it is still difficult to select the best PTF to estimate SHPs for site-specific water flow simulation.

A few years ago, the multimodel ensemble forecasting has been demonstrated to deal with the uncertainties in PTF selection (Guber et al., 2009). The idea of combining predictions from multiple independent models was first introduced by Bates and Granger (1969). Basic form of multimodel prediction is the simple model average (SMA) that always assigns equal weight on all the candidate models (Ajami et al., 2006). The superensemble is an improved multimodel prediction approach that combines different model forecasts during the training period using MLR (Krishnamurti et al., 2000). Many studies have found that superensemble had significantly better performance than the best single model and SMA (Krishnamurti et al., 2000, Guber et al., 2009). To our knowledge, however, few studies used the superensemble method for multimodel simulation of soil moisture content in forestland. The objectives of this study are (i) to evaluate the utilization of six traditional PTFs to predict SHPs for soil water simulation in a forest site; (ii) to compare the performances of the multimodel ensemble approaches (i.e. SMA and superensemble) with those of the best single model.