Predicting the mobile water content of vineyard soils in New South Wales, Australia

Agricultural Water Management 148 (2015) 34–42

Contents lists available at ScienceDirect

Agricultural Water Management journal homepage: www.elsevier.com/locate/agwat

Predicting the mobile water content of vineyard soils in New South Wales, Australia Jonathan E. Holland a,∗,1 , Asim Biswas b,2 a b

Wagga Wagga Agricultural Institute, NSW Department of Primary Industries, Wagga Wagga 2650, NSW, Australia Department of Natural Resource Sciences, McGill University, 21111 Lakeshore Road, Ste-Anne-de-Bellevue, Montréal, QC, Canada H9X 3V9

a r t i c l e

i n f o

Article history: Received 26 March 2014 Accepted 10 September 2014 Keywords: Pedotransfer functions Solute transport Nutrient leaching Mobile water fraction Tracer

a b s t r a c t Better understanding of the relationship between soil properties and soil function is required to minimise nutrient losses from agriculture and protect the environment. There is a need to predict the solute movement through horticultural soils because of the intensive management practices. Thus, the mobile water content ( m ), the active fraction of soil water content engaged in solute transport, is a suitable soil property to investigate further. Accurate measurement of such solute transport properties in the field are costly, labour intensive and time consuming but there are opportunities to establish predictive relationships.  m was measured together with other basic soil properties to test established predictive relationships (known as pedotransfer functions, PTFs) and to develop new PTFs. The field measurements were taken on a diverse range of vineyard soils across New South Wales (NSW), Australia. Poor predictions were found with available PTFs for  m and the mobile water fraction f (= m / fm ; where  fm = volumetric water content at which  m was measured). Backward stepwise multiple regression analysis produced better PTF model predictions than the multiple linear regression analysis for new PTFs that were calculated. Differences in the analysis methods showed a trade-off between the prediction capacity and the number of predictor variables in each PTF model. A prediction accuracy of between 80 to 90% was found with 3 predictor variables in the PTFs for  m and f. Both the PTFs developed were in strong agreement with the measured properties (minimum R2 = 0.82). For the  m PTF, the % clay content (varied from 11 to 59) was the strongest predictor variable while bulk density (ranged from 1.2 to 1.51 g cm−3 ) contributed the smallest. The PTF for f was similar to  m except it was the % soil organic carbon which had the smallest contribution. These relationships are useful to predict  m and f from easily measured soil physical properties of vineyard soils in NSW, but further testing on a wider range of soils is required. Crown Copyright © 2014 Published by Elsevier B.V. All rights reserved.

1. Introduction Understanding the movement of soil nutrients is important for agricultural production and environmental health. Estimating the availability and loss of nutrients is a key task for agronomists in research and commercial fields. This requires a determination of the loss of nutrients in order to make robust fertilizer recommendations (Chen et al., 2008). The loss of nutrients from agricultural land, especially nitrogen and phosphorus e.g. nitrate leaching from pastures (Ridley et al., 2001), remains an issue of concern for fresh

∗ Corresponding author. Tel.: +61 2 6938 1948; fax: +61 2 6938 1809. E-mail addresses: [email protected] (J.E. Holland), [email protected] (A. Biswas). 1 Formerly: National Wine and Grape Industry Centre, Charles Sturt University, Wagga Wagga 2678, NSW, Australia. 2 Formerly: PO Box 1666, CSIRO Land and Water, Canberra 2601, ACT, Australia. http://dx.doi.org/10.1016/j.agwat.2014.09.018 0378-3774/Crown Copyright © 2014 Published by Elsevier B.V. All rights reserved.

water quality globally (Drewry et al., 2006). This can be more acute in countries like Australia due to its climatic conditions. The rate at which nutrients move from the soil surface through the profile to ground water is governed by the inherent soil properties, especially those related to soil structure. Where land is used for agricultural crops the soil structure formation and stability is largely controlled by the management practices. Less intensive management practices such as reduced cultivation in perennial crops allows the formation of soil structure. Additionally, the irrigation of perennial crops is more common in Australia than for annual crops (Bryan et al., 2009). Because of the importance of irrigation for viticulture in Australia (Stevens et al., 2011) and internationally (López-Urrea et al., 2012; Williams, 2012), this study was undertaken on vineyard soils to investigate the relationship between basic soil properties (such as texture, bulk density) and the mobile water content ( m ). The relationship can help us better understand and predict the movement of solutes through perennial crop growing soils. More broadly, there is a lack of relationships across the world for the

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

prediction of solute transport parameters in perennial crops that are applicable at the field-scale. Further research is required in understanding the relationship between basic soil physical and key soil hydraulic properties for soils with this land use type. This type of information has major implications on irrigation and surface water management in agricultural systems and subsurface water quality management. The movement of water and dissolved nutrients in soil has been described in simplified terms as the flow through 2 regions of pore space i.e. connected pores and isolated micropores (Philip, 1968). The connected pores allow convective transport of water and nutrients and are defined as the mobile region ( m ); while the isolated micropores are poorly connected or stagnant and are known as the immobile region ( im ) (van Genuchten and Weirenga, 1976). The  m is the active fraction of soil water content engaged in solute transport and has been described as the rapid-mobile porosity (Legout et al., 2009) or the water-conducting macroporosity (van Tol et al., 2012). Various field (in-situ) and laboratory (columns) experiments have been conducted to estimate the mobile and immobile regions. For example, Clothier et al. (1992) estimated  m by measuring the tracer concentration during a controlled unsaturated flow while White et al. (1986) estimated the  m from solute breakthrough data. Lee and Casey (2005) used shallow measurements of  m to predict field-scale solute transport of an irrigated soil; under the conditions observed the  m compared well with a more complex and data intensive method with the transfer function model. For a given soil  m / was reported nearly constant over a range of unsaturated potential heads (−20 mm to −150 mm) measured at the initial soil water content of ∼0.3 cm3 cm−3 (Clothier et al., 1995). However,  m and the soil properties controlling  m are known to vary across different soil types (Clothier et al., 1995; Okom et al., 2000; Oliver and Smettem, 2003; Vogeler et al., 2006). Because of the potential efficiency in saving time and reducing the number of experimental measurements needed, this study has focused on better understanding the  m variable and the most important soil properties which influence  m . Developing predictive relationships (i.e. pedotransfer functions, PTFs) between soil properties has proven successful for soil hydraulic properties e.g. hydraulic conductivity (Minasny and McBratney, 2000; Paydar and Ringrose-Voase, 2003) and the soil water characteristic (Vervoort et al., 2006). Much less attention has been given to solute transport parameters. Nevertheless, interest in  m and other solute transport parameters has led to studies that investigate relationships with basic soil properties such as soil texture or soil organic carbon. This involves collecting quantitative data on these soil properties to explore relationships that can be used for predictive purposes. For this study the application is to aid the understanding on movement of nutrients in vineyard soils. PTFs have been developed for solute transport parameters such as the diffusion function (Dm ) using easily measured soil properties (Goncalves et al., 2001; Shaw et al., 2000). The mobile water fraction f (= m /;  is equivalent to  fm in this study, therefore  =  fm =  m +  im which is the measured water content at the applied potential) was found to vary with soil texture across a range of different soils in Victoria, Australia (Okom et al., 2000). The authors of this study also suggested a predictive relationship between f and other soil properties which was developed from  m PTFs provided earlier by Okom (1998). A review of solute transport parameters by Minasny and Perfect (2004) found that the data is limited and the methods used to generate data is not consistent. Despite this Minasny and Perfect (2004) were able to develop a PTF to predict  m using data from several studies. The authors used 28 different soils with clay content between 10 and 40%. These PTFs (Minasny and Perfect, 2004; Okom, 1998) have not yet been independently tested and it is not known how widely they can be applied or how they perform. In addition to robust testing, this

35

study will investigate opportunities to improve accuracy for PTFs of  m . This study aims to improve understanding of basic soil properties that control and influence the movement of soil nutrients and solutes. This will be achieved by focusing on the  m and exploring relationships between basic soil properties and solute transport and/or soil hydraulic properties. Consequently the specific objectives of the present study were to (i) measure the mobile water content ( m ), mobile fraction (f) and unsaturated hydraulic conductivity (K) together with a range of basic soil properties; (ii) test established PTFs and; (iii) develop new PTFs for  m and f for perennial crop growing soils. The new PTFs were developed from field measurements on a selection of vineyard soils across New South Wales (NSW), Australia. Vineyard soils and the management practices adopted in growing grapes can be considered similar to soils for some other perennial crops.

2. Materials and methods 2.1. Study sites and field measurements Measurements were completed on soils of 10 different vineyards located across NSW, Australia (Fig. 1). These soils covered a wide range of soil textural classes from sandy loam through to heavy clay and are classified as Chromosols to Dermosols (according to Australian Soil Classification) (Isbell, 2002) which is equivalent to Lixisols and Luvisol, respectively (IUSS Working Group WRB, 2006); representing the most important wine growing regions in NSW. A tension infiltrometer (Soil Measurement Systems, Tucson, AZ, USA) was used to supply a tracer solution (0.01 M KBr) to the soil surface following the method of Okom et al. (2000). This enabled the determination of the K and  m at the same point (Fig. 1). The soil was sampled for a selection of key properties (including soil texture, soil water content, soil organic carbon content and bulk density) both close by and beneath the position of the tension infiltrometer disc (Fig. 1). These measurements were taken during the period from February to May 2010. Between 4 and 8 tension infiltrometer measurements were taken along the vine row at each location (within 10 m linear distance), while other soil properties were measured on samples taken close by (<0.15 m from the disc edge) (Fig. 1). The sampling and measurements were taken in the area where fertilizers and irrigation water is commonly applied. Each vineyard chosen was drip irrigated and managed commercially.

2.2. Hydraulic conductivity A tension infiltrometer (200 mm diameter) was used to measure the in-situ K. The measurement sites selected were level and free from surface protruding stones or signs of recent disturbance. Wherever necessary, the grass or other vegetation covers were removed with care without disturbing the soil surface. A 0.01 M KBr solution was used in the infiltrometer to determine the  m from the same measurements as the K. The measurements were performed at a single potential (−20 mm) because for a particular soil the f is independent of the potential applied, but vary across soils (Okom et al., 2000). Following White et al. (1992) the early-time infiltration data was used to determine the sorptivity (S), while the steady state infiltration data enabled the calculation of K. The slope of first 300 s of infiltration data (I v. t1/2 ) was used to determine S with: I = St 1/2

(1)

36

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

Fig. 1. Sampling locations and the sampling scheme used at each location for determining the unsaturated hydraulic conductivity(K), mobile water content ( m ), initial ( i ) and final volumetric soil water content ( fm ), bulk density (BD), soil organic C (SOC) and particle size distribution.

where I = cumulative infiltration (cm) and t = time (s). For the K (mm h−1 ) the steady state infiltration rate (Q) was determined from the following equation: K=

Q 4bS 2 − 2 r r

ICS-2500 system). The  m was calculated with the following equation (Clothier et al., 1992): m = fm

(2)

here Q (mm h−1 ) is the steady state infiltration rate, (which usually took at least an hour for this to be reached);  =  fm – i , where  fm was the final volumetric water content at the supply surface (at which mobile water content was determined) and  i was the initial volumetric water content; r is the radius of the infiltration disc and b (0.5 ≤ b ≤ /4) is a parameter which depends only on the shape of the soil water diffusivity function and is set at 0.55 for this study following White and Sully (1987). 2.3. Mobile water content ( m ) The  m was determined following the method of Okom et al. (2000). At five points a push-tube corer sampled the area beneath the infiltration disc with the following depth increments: 0–3, 3–6; 6–9 and 9–12 cm. The volumetric water content ( fm ) and the Br concentration (C) were determined on these core samples. The  i and  fm were determined by oven drying until constant weight (usually 48 h at 105 ◦ C) and the material was then passed through a 2 mm sieve. On these samples the C was measured from an aqueous soil extract solution (1:1 ratio) using ion chromatography (Dionex

C C0

(3)

The C was determined for each depth increment and summed to give a total C (mg Br cm−2 ) for all depths, while C0 (mg Br cm−2 ) was the concentration of the applied Br tracer solution. The average of the five core samples was used to determine the C and the  fm (Fig. 1). The exchange between the immobile and mobile domains of soil water was not considered to be significant over the short time scale of the measurements in this study (Clothier et al., 1995; Oliver and Smettem, 2003). 2.4. Predictor variables In order to develop new predictive relationships and also to test published PTFs a range of additional basic soil properties were measured. These soil properties were chosen because previous studies (Minasny and Perfect, 2004; Okom et al., 2000) have shown their utility for the prediction of  m . The soil bulk density (BD) (g cm−3 ) was determined using the core method (McKenzie et al., 2002). The initial (i.e. before the tracer was supplied) volumetric soil water content ( i ; cm3 cm−3 ) was determined by sampling close to the infiltrometer disc (Fig. 1). The same soil sample was used to determine the texture by hydrometer method (Gee and Or, 2002) and the soil organic carbon (SOC) (%) following Heanes (1984) method.

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

Soil particle size classes were defined according to the International system (White, 2006) where clay is <2 ␮m, silt is 2–20 ␮m and sand is 20–2000 ␮m.

a model is used to represent the process that generated data) and is a measure of the relative quality of a statistical model (PTF). The AIC provides a means for model selection and was calculated as follows (Burnham and Anderson, 2002):

2.5. Pedotransfer function development and statistical analysis Multiple linear regression (MLR) analysis was carried out in SigmaPlot (Statsoft Inc.) to develop predictive relationships (i.e. PTFs) between  m and f and 4 measured soil properties or predictor variables (sand, clay, BD and SOC). Silt was also another measured variable, but was excluded as it is not independent of sand or clay and this avoids the issue of multicollinearity. The co-linearity between the predictor variables was examined using the variance inflation factor (VIF) and the tolerance (=1/VIF) (Hair et al., 1995). The normality of the predictor variables as well as the measured values of the variables being predicted ( m and f) were examined using skewness and kurtosis values and the Shapiro–Wilk test. Our analysis suggested these variables had a normal distribution, thus transformation of the data was not required. There was no separate dataset available to test the predictive relationship. Instead, we used cross validation to estimate how accurately the predictive relationship would perform in practice (Picard and Cook, 1984). The leave-one-out cross validation method was employed, where a single observation from the original dataset was used as the validation data and the remaining nine observations were used as the training data. A training dataset was used to develop the predictive relationship and the validation dataset was used to test the relationship independently, where each observation was used once as validation data. It is important to quantify the contribution of predictor variables in a PTF but often not done. Quantification of contribution can indicate the importance of each predictor in the PTF. This can also help in deciding on the number of predictors in the PTF without compromising the prediction capability. We have quantified the contribution of each predictor variable (i.e. soil property) in the PTF. This was achieved using a backward stepwise multiple regression (BSMR). In the BSMR, all predictor variables (sand, clay, BD and SOC) were first included in the regression (same as MLR) and then the soil properties were excluded one by one based on their contribution and importance in the PTF. A variable was excluded from the PTF if the F value (5.1) or the P value (0.05) for that variable exceeded a critical level. At each exclusion step, the coefficient of determination (R2 ) value, standard error of estimate (SEE) and the change in R2 value were also calculated. The BSMR analysis concluded with a final set of soil properties that described the strongest predictive relationship with each of the variables contributing significantly (P < 0.05). There were 4 predictor variables used in the PTFs for  m (clay, sand, SOC and BD) and f (clay, sand, BD and SOC). Systematic analysis of the residuals did not identify any significant trends or relationships. The Breusch–Pagan test (Breusch and Pagan, 1979) was used to identify homoscedasticity by testing the dependence of the variance of residuals on the values of independent variables. The presence of homoscedasticity was confirmed from the P value of the test result. The relationships developed through MLR and BSMR were compared to examine the contribution of each variable to the predictive relationship and to enable the strongest predictive relationship to be developed with the available data. Statistical evaluation of the PTFs was undertaken using the root mean squared error (RMSE) as this is recognised as a measure of PTF accuracy (McBratney et al., 2011). The 95% confidence interval was determined for the relationship between the predicted and measured values for the variables of interest. Furthermore, the Akaike Information Criterion (AIC) was calculated for each PTF (using MLR and BSMR analysis). The AIC uses Kullback–Leibler information or the information entropy (relative estimate of information lost when

37

AIC = n × log

 SS  n

+ 2 × (K + 1)

(4)

where n = sample size, SS = sum of squared errors, and K = number of predictor variables in the model. The PTF with the lowest AIC is the ‘best’ PTF model amongst those considered.

3. Results and discussion 3.1. Exploratory data analysis The  m values ranged from 0.04 for the Orange soil to 0.12 for the Murrumbateman soil. A similar magnitude of variation was found for the mobile fraction (f) values which had a median of 0.34 over all of the soils. It is difficult to make direct comparisons between the  m values measured in this study with those reported elsewhere due to the differences in depth (and hence soil volume) and measurement techniques. Nevertheless, it is noted that a wide range of  m values have been reported in the literature for different soils e.g. from 0.05 cm3 cm−3 (Okom, 1998), to 0.28 cm3 cm−3 (Clothier et al., 1995) and up to 0.8 cm3 cm−3 (Alletto et al., 2006). In addition to depth, such variations to  m can also be partially explained due to tillage (Alletto et al., 2011) or management practices (Vogeler et al., 2006) which alter the soil physical properties including structure. The use of broad geological (parent material or degree of weathering) type descriptors was considered, but found not to be consistent with the predictive approach outlined above (Section 2.5). Furthermore, the PTFs presented in this paper are based upon soils formed on several different parent materials including granite, trachyte, a sandstone/siltstone complex and alluvial material. O’Connell and Ryan (2002) suggest that the future application of soil PTFs should be restricted to areas with a similar parent material compared with where they are developed. The most important properties of the soils (for the surface horizon 0–12 cm depth) at each sampling location are given in Table 1; descriptive statistics for these soil properties are provided (Table A1 in the Appendix). The range in the soil property values from the ten vineyard locations highlight the distinct differences between the soils that were sampled. In this study there were no major differences to report between the sites for either tillage or other management practices that would influence the  m directly or any of the major predictor variables. The sites differed widely in soil texture. The clay content varied from 11 to 59% while the sand content ranged from a minimum of 18 to a maximum of 82%. The BD values were relatively uniform and the coefficient of variation (CV) was low (<10%). The SOC had greater variability and an overall median of 1.4%. The K (at −20 mm potential) and S varied more than any of the other soil properties; for instance the difference in K between the Young (185 mm h−1 ) and Mathoura (19 mm h−1 ) soil was very close to an order of magnitude. The soils containing the most sand content had the greatest K values. The S did not vary as much as the K (CV% 58 compared to 113) across all of the soils. The correlation of both variables (S and K) was weak with other soil properties (including  m ). The high variability of K between soils with different texture has been previously noted and consequently texture specific PTFs (e.g. for sandy soils) has been suggested (Minasny and McBratney, 2000). Due to the small number of soils in this study attempts to develop a PTF for K was not statistically significant.

38

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

Table 1 The mean valuesa for the bulk density (BD), organic carbon (SOC %), clay (%), sand (%), unsaturated hydraulic conductivity (K at −20 mm potential), sorptivity (S), mobile water content ( m ), volumetric soil water content () and the mobile water fraction (f) for the surface soil (0–12 cm depth) for each soil type. Location

Clay (%)

Sand (%)

BD (g cm−3 )

Porosity (%)

SOC (%)

K (mm h−1 )

S (mm h−0.5 )

 m (cm3 cm−3 )

 (cm3 cm−3 )

f

Hunter valley 2 Murrumbateman Cowra Gundagai Dareton Mudgee Mathoura Hunter valley 1 Young Orange

60 19 13 21 13 37 29 38 19 20

18 61 76 70 83 45 47 47 70 55

1.20 1.38 1.46 1.37 1.45 1.22 1.51 1.32 1.42 1.22

55 48 45 48 45 54 43 50 46 54

1.70 1.92 0.93 1.01 0.74 1.99 1.85 1.37 1.33 2.71

60 31 143 36 177 76 19 15 185 122

10 4 16 7 8 15 5 4 16 12

0.12 0.12 0.11 0.10 0.10 0.10 0.07 0.05 0.05 0.04

0.39 0.25 0.26 0.20 0.18 0.27 0.19 0.32 0.22 0.35

0.29 0.44 0.38 0.42 0.47 0.33 0.36 0.16 0.21 0.12

a Additional basic statistics for these data including the no. of observations, minimum, maximum, median, standard deviation and the coefficient of variation are given in the Appendix (Table A1).

Table 2 Established pedotransfer functions (PTFs) for mobile water content ( m ) and mobile water fraction (f) with the original R2 values and R2 values using our measured data. PTF no.

Prediction variable

Regression relationship

R2 (original)

R2 (our data)

Equation (reference)

1 2 3 4a 5 6

m f f m m m

0.232(BD/1.3)−0.023 + 0.0021 × Silt (BD/1.3)−0.205 − 0.0089 × Clay 1 − 0.0221Clay(1–BD/2.65) 0.2549(BD/1.3)−0.2954 + 0.002 × Silt − 0.008 × Clay −0.1193 + 0.92 ×  − 0.00115 × Sand −0.1712 + 1.3148 ×  − 0.00496 × Clay2

0.67 0.65 0.75 0.67 0.44 0.47

0.08 0.02 0.00 0.15 0.02 0.25

Equ. 7.5 (Okom, 1998) Equ. 7.6 (Okom, 1998) White et al. (1999) Equ. 48 (Minasny and Perfect, 2004) Equ. 54 (Minasny and Perfect, 2004) Equ. 56 (Minasny and Perfect, 2004)

a

This PTF was given by Minasny and Perfect (2004) as being developed by Okom (1998), but we could not find it in Okom (1998).

3.2. Relationship between basic soil properties and  m —Testing established pedotransfer functions (PTFs) We tested 6 established PTFs (Table 2) of either  m or f with the measured soil properties in this study (Table 1). These PTFs showed poor predictive capability which was confirmed from the low R2 values (Table 2). For example, Okom (1998) provided PTFs (no. 1 and 2 in Table 2) for  m and f predicted values (ranging from 0.36 to 0.78 for  m and 0.49 to 0.88 for f) that were on average 3 times greater than the measured values (Table 1). The PTF (no. 3 in Table 2) for f by White et al. (1999) resulted in large over predictions for 9 out of 10 soil types and were 1.4 to >6 times greater than the measured values. Likewise PTF no. 4 in Table 2 over predicted  m between 3 and >10 times. The 2 PTFs (no. 5 and 6 in Table 2) developed by Minasny and Perfect (2004) include the  parameter which is questionable. The  parameter should not be used as it is related to  m (i.e. there is co-linearity between  and  m ). Further  is not a stable soil property and is associated with significant spatial and temporal variation (Gomez et al., 2009; Timm et al., 2011) and dependent on the surface potential or flux. This clearly showed that the established predictive relationships performed poorly when applied to our data. These PTFs either produced a large over prediction or spurious values of  m e.g. applying our data to PTF no. 6 in Table 2 resulted in negative values for m . 3.3. New PTFs for  m and f New PTFs were developed for predicting  m and f from the measured soil properties using the MLR and BSMR analysis (Table 3). For  m , the PTF using MLR (including all predictor variables) produced an R2 value of 0.88, while following BSMR (including the most significant predictor variables) had an R2 value of 0.82. The RMSE values (0.0106 and 0.0127, respectively) were small which indicate acceptable accuracy for the new predictive relationships. Additional detailed statistics of PTFs and the parameters are provided (Table A2 in the Appendix). Most values were close to the 1:1

line except for 3 soils, which were outside of the 95% confidence band (Fig. 2a). All soil properties (clay, sand, SOC, and BD) were used in the MLR to develop the predictive relationship. The inclusion/exclusion of one soil property in the predictive relationship and its tradeoff with the predictive power of the relationship was examined using the BSMR. The variables in the regression relationship column in Table 3 are listed according to the contribution (from most to least) of each predictor variable. Hence the clay content had the greatest contribution in establishing the PTF for  m while BD has the smallest (using MLR analysis). This may be due to the stronger positive correlation of  m with clay (r = 0.46) compared to the other soil properties. Moreover, the stronger influence of clay content on the predictive relationship may be due to the higher stability of soil structure providing well-networked pore spaces (Marshall et al., 1996). The wide range of clay content (Table A1) in the samples may also better explain the clay response variable in comparison to narrow spread of the BD values. Therefore, if we remove the BD from the predictor variables in developing regression relationship (i.e. BSMR) the overall R2 value is reduced by only 6% (Table 3). The PTF determined using the BSMR performed better compared with the MLR analysis (as indicated by the AIC values) (Table 3). The AIC penalizes the number of parameters in PTF model (Burnham and Anderson, 2002). Thus, despite the prediction capability (R2 = 0.82) of the 3-parameter BSMR-based PTF, the lower AIC value (Table 3) showed it was better compared with the 4-parameter MLR based PTF (R2 = 0.88). This may be because AIC deals with the trade-off between the goodness of fit and the complexity of the PTFs. For the f PTF, the R2 value was 0.89 and 0.88 for MLR-based and BSMR-based PTFs, respectively, with small and similar RMSE (Table 3). The predicted versus measured values were very close to the 1:1 line except for 2 soils that were outside of the 95% confidence band (Fig. 2b). The SOC had the smallest contribution in developing the predictive relationship and if we remove the SOC from the regression relationship, the R2 value is reduced by <1%. According to the AIC values for the f PTF, the PTF developed using BSMR analysis was identified as the better (and simpler) than the

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

39

Table 3 Pedotransfer function models for mobile water content ( m ) and mobile water fraction (f) using two analysis methods (multiple linear regression, MLR and backward stepwise multiple regression, BSMR) that were based upon basic soil properties (Clay %; Sand %; SOC %; and BD g cm−3 ) from ten vineyards in NSW, Australia. PTF no.

Prediction variable

Analysis method

Regression relationship

RMSEa

R2 b

AICc

1 2 3 4

m m f f

MLR BSMR MLR BSMR

−0.344 + (0.00564 × Clay) + (0.00315 × Sand) − (0.0208 × SOC) + (0.0978 × BD) −0.149 + (0.00482 × Clay) + (0.00269 × Sand) − (0.0297 × SOC) −2.258 + (0.0216 × Clay) + (0.0136 × Sand) + (0.932 × BD) − (0.0290 × SOC) −2.495 + (0.0224 × Clay) + (0.0143 × Sand) + (1.027 × BD)

0.0106 0.0127 0.0376 0.0396

0.88 0.82 0.89 0.88

−20.97 −23.25 −9.28 −11.32

a b c

RMSE is the root mean squared error. R2 is the coefficient of determination. AIC is the Akaike Information Criterion.

(a)

θ

(b)

Fig. 2. Predicted vs. measured values for (a)  m and (b) f from 10 vineyard soils in NSW, Australia. The regression fit and R2 values are given in the top left corner of each plot.

PTF calculated with MLR analysis (Table 3). After removing the SOC, remaining soil properties (clay, sand, and BD) had a statistically significant (P < 0.05) contribution in developing the PTF (using BSMR). Most of the PTFs reported by others in Table 2 have only 2 soil properties (except 1 with 3 properties). A reduced number of soil properties in the PTF means there are less to measure. This highlights the issue of trade-offs between the number of soil properties required and the efficiency of the predictive relationship. One may measure fewer soil properties, but compromise the prediction percentage. For example, the PTFs reported by others have 2 properties with prediction accuracy of 44 to 75% (Table 2). In this study a prediction accuracy of between 80 to 90% was found with 3 parameters in the PTFs for  m and f (Table 3). This means that the measurement of an extra soil property can almost double the prediction accuracy. A comparison was made between the 2 parameter ‘old’ PTFs (Table 2) with the 3 parameter new PTFs (Table 3) using the data collected in this study. Here we also show the trade-offs between

the prediction accuracy and the number of variables in the PTFs. This comparison is an important finding and we suggest the same comparison could be used in other situations. This has implications for the prediction of solute transport properties for different land uses across the globe, in particular for perennial agricultural or horticultural crops with similar management practices. We have developed PTFs for a globally important type of agricultural land use and shown the risk of using empirically derived PTFs too widely. Thus, the relationships between soil properties with solute transport or hydraulic processes and the variables which describe such processes is often much simpler than is realised. Moreover, in the PTFs reported (Table 2) there was only one particle size (either sand or silt or clay) fraction that was used to determine a predictive relationship; in contrast 2 particles size fractions (clay and sand) were selected in this study. Because the same method (e.g. the hydrometer (Gee and Or, 2002) measurement) provides 3 particle size fractions simultaneously and does not increase the number of measurements/experiments to be undertaken. In addition, we also included SOC in  m PTF. SOC is an important soil property for perennial crops such as grapevines as the management practices do not include frequent tillage and thus the potential breakdown of the soil structure (Cockroft and Olsson, 2000) and the loss of SOC (Chan et al., 2011; Steenwerth and Belina, 2008). Our PTFs were able to account for large amount of variability of  m and f with the R2 at least 0.82 (Table 3), however greater prediction power could be possible for these relationships through the incorporation of additional soil properties. Wösten et al. (2001) listed an extensive range of properties which should be considered for developing soil hydraulic PTFs. Additional soil properties included water-stable aggregates, soil morphological properties and detailed particle size data e.g. different sand fractions. Waterstable aggregates directly relate to macroporosity (Zhang et al., 2007) which is influenced by bulk density (McKenzie et al., 2002). Therefore, this provides an opportunity for future development of PTFs for  m or f including the above mentioned properties (such as aggregate stability and size distribution). Inclusion of more difficult to measure soil properties in the predictive relationship may improve the prediction accuracy, but the additional cost and time can question the feasibility of the PTF development. Moreover, there has not been any research on the relationship between  m and direct and detailed measures of soil structure such as pore connectivity or pore path length. The PTFs developed in this study are land use specific as all measurements were taken on vineyard soils. In the future additional measurements of  m and f on a greater selection of land use types are required to test the wider application of the new PTFs given in Table 2. The  m variable is useful for determining the leaching potential of a given soil (Al-Jabri et al., 2002; Jaynes et al., 1995) and this study contributes to better understanding of the leaching of vineyard soils. Finally, the  m and f data can be used for the simulation of nutrient leaching processes that are controlled by these solute transport variables.

40

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

4. Conclusion

exists on the application of these new PTFs and future testing is recommended.

This study focused on the relationship between basic soil properties and solute transport properties for vineyard soils in NSW, Australia. This was undertaken by exploring predictive relationships between soil properties and 2 related variables: the  m and f. Established PTFs were tested, but these were found to be poor predictors of both  m and f. Consequently, new PTFs were developed with excellent prediction accuracy with R2 values of at least 0.82. The PTFs developed were based upon basic soil properties which are useful for practical applications. Analysis of regression methods showed that there was a trade-off between prediction capacity and the number of predictor variables for each PTF model calculated. In this study the BSMR analysis was superior to MLR for calculating the best and most simple PTFs. The results showed much stronger influence of clay content than BD as a predictor variable for  m . The prediction of  m from the easily measured soil properties can provide understanding on the potential for nutrient movement through the soil, in particular for mobile chemicals. Beyond vineyard soils in NSW some uncertainty

Acknowledgements This work was partially funded by the Wine Growing Futures Program, a joint initiative of the Grape and Wine Research and Development Corporation and the National Wine & Grape Industry Centre. We thank Mr Peter Carey for his assistance with the fieldwork and for the co-operation of the vineyard owners and managers in providing access to their properties. We also thank Dr. Hamish Cresswell and Dr. Keith Bristow (both CSIRO) for their helpful comments on an earlier draft of the manuscript.

Appendix A. Appendix Table A1. Table A2.

Table A1 Descriptive statistics of the input variables (including bulk density (BD), organic carbon (SOC), unsaturated hydraulic conductivity (K), sorptivity (S) and the volumetric soil water content () used to develop the new pedotransfer functions (PTFs) for mobile water content ( m ) and mobile water fraction (f) for the surface soil (0–12 cm depth) from each sampling location and for the overall dataset from ten vineyard soils in NSW, Australia. Location

Statistic

Sand (%)

BD (g cm−3 )

a

Clay (%)

SOC (%)

K (mm h−1 )

S (cm s−1 )

 m (cm3 cm−3 )

 (cm3 cm−3 )

f

Hunter valley 2

Median n SD

NA 1 NA

NA 1 NA

1.21 8 0.043

1.76 4 0.152

55.15 4 18.507

10.74 4 2.077

0.12 4 0.021

0.41 4 0.021

0.30 4 0.038

Murrumbateman

Median n SD

NA 1 NA

NA 1 NA

1.36 12 0.043

1.76 6 0.896

16.21 6 37.874

4.95 6 1.269

0.12 6 0.045

0.27 6 0.038

0.44 6 0.118

Cowra

Median n SD

NA 1 NA

NA 1 NA

1.29 12 0.059

1.91 6 0.206

124.37 6 28.689

15.36 6 1.215

0.08 6 0.039

0.33 6 0.031

0.27 6 0.099

Gundagai

Median n SD

NA 1 NA

NA 1 NA

1.37 16 0.069

0.88 8 0.223

38.60 8 10.375

7.17 8 3.567

0.10 8 0.022

0.25 8 0.018

0.41 8 0.110

Dareton

Median n SD

NA 1 NA

NA 1 NA

1.46 12 0.055

0.73 6 0.238

134.15 6 200.023

6.97 6 2.547

0.10 6 0.025

0.21 6 0.035

0.49 6 0.078

Mudgee

Median n SD

NA 1 NA

NA 1 NA

1.21 10 0.050

1.97 5 0.224

78.17 5 21.835

15.22 5 2.864

0.10 5 0.019

0.32 5 0.024

0.35 5 0.051

Mathoura

Median n SD

NA 1 NA

NA 1 NA

1.51 8 0.016

1.83 4 0.134

19.23 4 4.104

5.19 4 0.673

0.07 4 0.008

0.23 4 0.024

0.35 4 0.019

Hunter valley 1

Median n SD

NA 1 NA

NA 1 NA

1.32 12 0.063

1.41 6 0.132

16.59 6 10.288

4.48 6 2.120

0.05 6 0.021

0.34 6 0.019

0.15 6 0.063

Young

Nedian n SD

NA 1 NA

NA 1 NA

1.40 10 0.086

1.38 5 0.163

116.70 5 137.052

15.63 5 3.303

0.05 5 0.016

0.24 5 0.018

0.20 5 0.060

Orange

Median n SD

NA 1 NA

NA 1 NA

1.25 12 0.108

2.63 6 0.404

126.60 6 21.856

9.40 6 8.195

0.03 6 0.014

0.38 6 0.031

0.11 6 0.041

All locations

Median N SD CV% Min Max

20.3 10 15.2 58 11.1 59.6

57.7 10 19.5 34 17.8 82.5

1.36 55 0.12 9 1.01 1.57

a

NA = not available.

1.41 56 0.68 45 0.51 3.36

55 56 96 112 1.6 567.0

7.8 56 5.5 57 1.3 27.5

0.08 56 0.04 44 0.02 0.18

0.26 56 0.07 28 0.14 0.40

0.34 56 0.14 44 0.07 0.60

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

41

Table A2 Model summary statistics, analysis of variance and parameter statistics for the pedotransfer functions (PTFs) for (a) mobile water content ( m ) and (b) mobile water fraction (f) calculated using two analysis methods (multiple linear regression, MLR and backward stepwise multiple regression, BSMR) from ten vineyard soils in NSW, Australia. (a) PTFs for  m (A) Model summary statistics Statistic term

MLR

BSMR

R R2 Adjusted R2 Standard error of estimate Variables in model MSEA RMSEB

0.936 0.877 0.778 0.015 4 0.00011194 0.01058022

0.907 0.822 0.734 0.0164 3 0.00016166 0.01271468

(B) Analysis of variance Analysis method

Source

DF

SS

MS

F

P

MLR

Regression Residual Total Regression Residual Total

4 5 9 3 6 9

0.00799 0.00113 0.00912 0.00750 0.00162 0.00912

0.002 0.000225 0.00101 0.00250 0.000270 0.00101

8.873

0.017

9.261

0.011

Std. coeff.

F to remove

2.582 1.886 −0.39 0.337

21.845 12.663 3.589 2.192

2.204 1.609 −0.556

16.884 8.784 8.63

BSMR

(C) Parameter statisticsA Analysis method

Parameter

Coefficient

Std. error

t

P

MLR

Constant Clay Sand SOC BD Constant Clay Sand SOC

−0.344 0.00564 0.00315 −0.0208 0.0978 −0.149 0.00482 0.00269 −0.0297

0.153 0.00121 0.000885 0.011 0.0661 0.0851 0.00117 0.000907 0.0101

−2.252 4.674 3.559 −1.895 1.48 −1.756 4.109 2.964 −2.938

0.074 0.005 0.016 0.117 0.199 0.130 0.006 0.025 0.026

BSMR

(b) PTFs for f (A) Model summary statistics Statistic term

MLR

BSMR

R R2 Adjusted R2 Standard error of estimate Variables in model MSE RMSE

0.945 0.893 0.808 0.053 4 0.001417 0.037648

0.939 0.881 0.822 0.051 3 0.001569 0.039605

(B) Analysis of variance Analysis method

Source

DF

SS

MS

F

P

MLR

Regression Residual Total Regression Residual Total

4 5 9 3 6 9

0.118 0.0142 0.133 0.117 0.0157 0.133

0.0296 0.00283 0.0147 0.0389 0.00262 0.0147

10.44

0.012

14.844

0.003

Std. coeff.

F to remove

2.594 2.138 0.843 −0.142

25.46 18.792 15.826 0.552

2.692 2.253 0.929

31.753 24.995 29.46

BSMR

(C) Parameter statisticsC Analysis method

Parameter

Coefficient

Std. error

t

P

MLR

Constant Clay Sand BD SOC Constant Clay Sand BD

−2.258 0.0216 0.0136 0.932 −0.029 −2.495 0.0224 0.0143 1.027

0.542 0.00429 0.00314 0.234 0.039 0.422 0.00398 0.00287 0.189

−4.163 5.046 4.335 3.978 −0.743 −5.914 5.635 4.999 5.428

0.009 0.004 0.007 0.011 0.491 0.001 0.001 0.002 0.002

BSMR

A B C

MSE is the mean square error. RMSE is the root mean square error. Inclusion/exclusion criteria; F = 5.1 and P = 0.05.

42

J.E. Holland, A. Biswas / Agricultural Water Management 148 (2015) 34–42

References Al-Jabri, S.A., Horton, R., Jaynes, D.B., 2002. A point-source method for rapid simultaneous estimation of soil hydraulic and chemical transport properties. Soil Sci. Soc. Am. J. 66, 12–18. Alletto, L., Coquet, Y., Vachier, P., Labat, C., 2006. Hydraulic conductivity, immobile water content, and exchange coefficient in three soil profiles. Soil Sci. Soc. Am. J. 70, 1272–1280. Alletto, L., Coquet, Y., Vachier, P., Labat, C., 2011. Vertical and lateral variations of soil immobile water fraction in two tillage systems. Soil Sci. Soc. Am. J. 75, 498– 508. Breusch, T.S., Pagan, A.R., 1979. Simple test for heteroscedasticity and random coefficient variation. Econometrica (The Econometric Society) 47, 1287–1294. Bryan, B.A., Barry, S., Marvanek, S., 2009. Agricultural commodity mapping for land use change assessment and environmental management: an application in the Murray–Darling Basin, Australia. J. Land Use Sci. 4, 131–155. Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference: A Practical Information–Theoretic Approach, second ed. Springer-Verlag, New York, NY. Chan, K.Y., Conyers, M.K., Li, G.D., Helyar, K.R., Poile, G., Oates, A., Barchia, I.M., 2011. Soil carbon dynamics under different cropping and pasture management in temperate Australia: results of three long-term experiments. Soil Res. 49, 320–328. Chen, D., Suter, H., Islam, A., Edis, R., Freney, J.R., Walker, C.N., 2008. Prospects of improving efficiency of fertiliser nitrogen in Australian. Agriculture: a review of enhanced efficiency fertilisers. Aust. J. Soil Res. 46, 289–301. Clothier, B.E., Kirkham, M.B., McLean, J.E., 1992. In situ measurement of effective transport volume for solute moving through soil. Soil Sci. Soc. Am. J. 56, 733–736. Clothier, B.E., Heng, L.K., Magesan, G.N., Vogeler, I., 1995. The measured mobilewater content of an unsaturated soil as a function of hydraulic regime. Aust. J. Soil Res. 33, 397–414. Cockroft, B., Olsson, K.A., 2000. Degradation of soil structure due to coalescence of aggregates in no-till, no-traffic beds of irrigated crops. Aust. J. Soil Res. 38, 61–70. Drewry, J.J., Newham, L.T.H., Greene, R.S.B., Jakeman, J., Croke, B.F.W., 2006. A review of nitrogen and phosphorus export to waterways: context for catchment modelling. Mar. Freshwater Res. 57, 757–774. Gee, G.W., Or, D., 2002. Particle-size analysis. In: Dane, J.H., Clarke Topp, G. (Eds.), Methods of Soil Analysis. Part 4. Physical Methods. SSSA, Madison, WI, pp. 255–293. Gomez, J.A., Sobrinho, T.A., Giraldez, J.V., Fereres, E., 2009. Soil management effects on runoff, erosion and soil properties in an olive grove of southern Spain. Soil Tillage Res. 102, 5–13. Goncalves, M.C., Leij, F.J., Schaap, M.G., 2001. Pedotransfer functions for solute transport parameters of Portuguese soils. Eur. J. Soil Sci. 52, 563–574. Hair, J.F.J., Anderson, R.E., Tatham, R.L., Black, W.C., 1995. Multivariate Data Analysis, third ed. Macmillan, New York, NY. Heanes, D.L., 1984. Determination of total organic-C in soils by an improved chromic acid digestion and spectrophotometric procedure. Commun. Soil Sci. Plant Anal. 15, 1191–1213. Isbell, R.F., 2002. The Australian Soil Classification, Revised ed. CSIRO Publishing, Melbourne. IUSS Working Group WRB, 2006. World reference base for soil resources 2006. In: World Soil Resources Report No. 103. FAO, Rome. Jaynes, D.B., Logsdon, S.D., Horton, R., 1995. Field method for measuring mobile/immobile water content and solute transfer rate coefficient. Soil Sci. Soc. Am. J. 59, 352–356. Lee, J., Casey, F.X.M., 2005. Development and evaluation of a simplified mechanistic–stochastic method for field-scale solute transport prediction. Soil Sci. 170, 225–234. Legout, A., Legout, C., Nys, C., Dambrine, E., 2009. Preferential flow and slow convective chloride transport through the soil of a forested landscape (Fougères France). Geoderma 151, 179–190. ˜ López-Urrea, R., Montoro, A., Manas, F., López-Fuster, P., Fereres, E., 2012. Evapotranspiration and crop coefficients from lysimeter measurements of mature ‘Tempranillo’ wine grapes. Agric. Water Manage. 112, 13–20. Marshall, T.J., Holmes, J.W., Rose, C.W., 1996. Soil Physics, third ed. Cambridge University Press, Cambridge, UK. McBratney, A.B., Minasny, B., Tranter, G., 2011. Necessary meta-data for pedotransfer functions. Geoderma 160, 627–629. McKenzie, N.J., Coughlan, K.J., Cresswell, H.P., 2002. Soil Physical Measurement and Interpretation for Land Evaluation. CSIRO Publishing, Melbourne.

Minasny, B., McBratney, A.B., 2000. Evaluation and development of hydraulic conductivity pedotransfer function for Australian soil. Aust. J. Soil Res. 38, 905–926. Minasny, B., Perfect, E., 2004. Development of pedotransfer functions in soil hydrology—solute adsorption and transport parameters. Dev. Soil Sci. 30, 195–224. O’Connell, D.A., Ryan, P.J., 2002. Prediction of three key hydraulic properties in a soil survey of a small forested catchment. Aust. J. Soil Res. 40, 191–206. Okom, A.E., 1998. Estimating Soil Hydraulic Properties and the Solute Transport Volume Using Surrogate Variables. Institute of Land and Food Resources. University of Melbourne, Melbourne. Okom, A.E., White, R.E., Heng, L.K., 2000. Field measured mobile water fraction for soils of contrasting texture. Aust. J. Soil Res. 38, 1131–1142. Oliver, Y.M., Smettem, K.R.J., 2003. Parameterisation of physically based solute transport models in sandy soils. Aust. J. Soil Res. 41, 771–788. Paydar, Z., Ringrose-Voase, A.J., 2003. Prediction of hydraulic conductivity for some Australian soils. Aust. J. Soil Res. 41, 1077–1088. Philip, J.R., 1968. The theory of absorption in aggregated media. Aust. J. Soil Res. 6, 1–19. Picard, R., Cook, D., 1984. Cross-validation of regression models. J. Am. Stat. Assoc. 79, 575–583. Ridley, A.M., White, R.E., Helyar, K.R., Morrison, G.R., Heng, L.K., Fisher, R., 2001. Nitrate leaching loss under annual and perennial pastures with and without lime on a duplex (texture contrast) soil in humid southeastern Australia. Eur. J. Soil Sci. 52, 237–252. Shaw, J.N., West, L.T., Radcliffe, D.E., Bosch, D.D., 2000. Preferential flow and pedotransfer functions for transport properties in Sandy Kandiudults. Soil Sci. Soc. Am. J. 64, 670–678. Steenwerth, K., Belina, K.M., 2008. Cover crops enhance soil organic matter, carbon dynamics and microbiological function in a vineyard agroecosystem. Appl. Soil Ecol. 40, 359–369. Stevens, R.M., Harvey, G., Partington, D.L., 2011. Irrigation of grapevines with saline water at different growth stages: effects on leaf, wood and juice composition. Aust. J. Grape Wine Res. 17, 239–248. Timm, L.C., Dourado-Neto, D., Bacchi, O.O.S., Hu, W., Bortolotto, R.P., Silva, A.L., Bruno, I.P., Reichardt, K., 2011. Temporal variability of soil water storage evaluated for a coffee field. Soil Res. 49, 77–86. van Genuchten, M.T., Weirenga, P.J., 1976. Mass transfer studies in sorbing porous media 1. Analytical solutions. Soil Sci. Soc. Am. J. 40, 473–484. van Tol, J.J., Le Roux, P.A.L., Hensley, M., 2012. Pedotransfer functions to determine water conducting macroporosity in South African soils. Water Sci. Technol. 65, 550–557. Vervoort, R.W., Minasny, B., Cattle, S.R., 2006. The hydrology of Vertosols used for cotton production: II. Pedotransfer functions to predict hydraulic properties. Aust. J. Soil Res. 44, 479–486. Vogeler, I., Cichota, R., Sivakumaran, S., Deuer, M., McIvor, I., 2006. Soil assessment of apple orchards under conventional and organic management. Aust. J. Soil Res. 44, 745–752. White, I., Sully, M.J., 1987. Macroscopic and microscopic capillary length and time scales from field infiltration. Water Resour. Res. 23, 1514–1522. White, I., Sully, M.J., Perroux, K.M., 1992. Measurement of surface-soil hydraulic properties: disc permeameters, tension infiltrometers and other techniques, advances in measurement of soil physical properties: bringing theory into practices. In: Soil Science of America Special Publication No. 30. Soil Science Society of America, Madison, Wisconsin. White, R.E., 2006. Principles and Practice of Soil Science—The Soil as a Natural Resource, fourth ed. Blackwell Science, Oxford. White, R.E., Dyson, J.S., Haigh, R.A., Jury, W.A., Sposito, G., 1986. A transfer function model of solute transport through soil 2. Illustrative applications. Water Resour. Res. 22, 248–254. White, R.E., Okom, A.E., Edis, R., Heng, L.K., 1999. Estimating a soil’s solute transport volume from surrogate variables,. In: Symposium on Preferential Flow Dynamics, Soil Science Society of America, Salt Lake City, Utah, p. 177. Williams, L.E., 2012. Effects of applied water amounts at various fractions of evapotranspiration (ETc) on leaf gas exchange of Thompson Seedless grapevines. Aust. J. Grape Wine Res. 18, 100–108. Wösten, J.H.M., Pachepsky, Y.A., Rawls, W.J., 2001. Pedotransfer functions: bridging the gap between available basic soil data and missing soil hydraulic characteristics. J. Hydrol. 251, 123–150. Zhang, G.S., Chan, K.Y., Oates, A., Heenan, D.P., Huang, G.B., 2007. Relationship between soil structure and runoff/soil loss after 24 years of conservation tillage. Soil Tillage Res. 92, 122–128.