Printer Friendly

Monitoring changes in soil salinity and sodicity to depth, at a decadal scale, in a semiarid irrigated region of Australia.

Received 26 March 2018, accepted 31 July 2018, published online 2 October 2018


Soil salinity and sodicity are two of the most important and widespread soil constraint and degradation issues in Australia and throughout the globe (Shahid 2013; Mora et al. 2017). Soil salinity has typically received more recognition and attention in Australian agricultural regions, but soil sodicity far overshadows salinity in regards to impact on agriculture and land area covered (Northcote and Skene 1972; Orton et al. 2018). Arid and semiarid ecosystems cover ~40% of Earth's surface, and are particularly known for possessing saline and sodic soils, largely due to the minimal leaching of salts from the soil profile (Ford et al. 1993; Corwin et al. 2003; Rengasamy 2006; Jalali et al. 2008; Hulugalle et al. 2013). Despite receiving low rainfall, these landscapes often have an abundance of groundwater and surface water available for irrigated agriculture, and many semiarid areas in eastern Australia are used for irrigated cotton (Gossypium hirsutum) production.

Excessively saline soils can detrimentally impact crop production, but soil salinity can also positively impact the physical properties of a soil, as a high electrolyte concentration encourages particles to bind together into aggregates (McKenzie and Orange 2003). In contrast, soil sodicity encourages the dispersion of soil, as excessive sodium ions on exchange sites weaken bonds between clay particles (Rengasamy and Olsson 1991; Sumner 1993). The relative proportion of salinity and sodicity in a soil is crucial, as this determines whether a soil will disperse or stay aggregated upon wetting (McNeal 1968). Consequently, the salt concentration of applied irrigation water can have an important role in altering the physical structure of soils, with saline irrigation water encouraging flocculation, and less saline water promoting dispersion (Shainberg et al. 1981).

Although soil salinity and sodicity are often inherent issues in Australia's arid and semiarid landscapes, the effects of these can be further accentuated by agricultural management practices (Cattle and Field 2013). The salt content of a soil horizon can increase through the movement of water from other parts of the soil, including the percolation of salty water from the surface (e.g. salty irrigation water or dissolved fertiliser salts), or the rising of salty water from an underlying watertable. This connectivity is why it is crucial to concurrently monitor salinity and sodicity at multiple depths in the soil, although this is rarely considered. The Australian cotton industry is dominated by irrigated production, accounting for as much as 79% of the production area, and 90% of total lint produced between 2009 and 2014 (Roth 2014). During a typical growing season of an irrigated cotton crop, the equivalent of ~10001400 mm of water is applied, and with this ~1.5 tonnes per hectare of salt may be deposited into the soil within the root zone (McKenzie and Orange 2003). In general, sodium is the dominant cation in this deposited salt, causing potentially detrimental problems of both soil salinity and sodicity (McKenzie and Orange 2003). The quality of irrigation in Australia can be highly variable, with groundwater typically possessing higher sodium levels than that of water extracted from rivers (Speirs et al. 2011). During the 2000s, much of eastern Australia experienced the 'Millennium Drought', which meant that surface water for irrigation was scarce and only groundwater was available. However, the emergence of La Nina in 2010 resulted in several consecutive years of high rainfall and floods in some locations, making surface water for irrigation more available again.

Although there are a few examples of studies that have monitored spatio-temporal changes in soil salinity (e.g. Herrero and Perez-Coveta 2005), most studies that have mapped soil sodicity have only done this at one time point at the field (Corwin et al. 2003) and regional scale (Ford et al. 1993). This is not due to a lack of interest in sodicity, but rather because the electrical conductivity (EC) of a soil is easily measured by traditional laboratory techniques, whereas analysing exchangeable sodium percentage (ESP) by traditional laboratory methods is both extremely time and resource intensive. There has been much recent focus on using inexpensive and rapid soil spectroscopic techniques to predict soil properties as a way of overcoming this issue (Viscarra Rossel et al. 2006). These techniques are a promising avenue for predicting soil ESP and other difficult-to-measure soil properties, and this is likely to encourage monitoring studies to focus on environmentally- and agriculturally-important soil properties that are often disregarded due to their cost-prohibitive nature.

In this work, we analyse the impact of recent shifts in rainfall, variability of irrigation water quantity and quality, and the intensive management practices of irrigated cotton production has on soil condition. In the relatively new southern cotton-growing regions of New South Wales (NSW) there is particular interest in the state of the soil, as there has been little monitoring of the potential impacts of cotton production in these areas (Holland and Eastwood 2014). Australia-wide maps of soil salinity and sodicity (Northcote and Skene 1972) indicate that the soils in these landscapes are typically sodic. Long-term studies of spatio-temporal changes in soil salinity and sodicity are scarce, and although these issues are important throughout the soil profile, most studies focus solely on the topsoil. This study monitors the change in soil EC at five depth increments to 1.2 m and ESP at two depth increments (0-0.1 and 0.8-1.2m) from 2002 to 2015 in a semiarid, irrigated cotton-growing region of south-west NSW, Australia. Due to cost limitations, visible near infrared (VisNIR) spectra are combined with traditional laboratory methods to predict soil ESP on a portion of samples, and the benefits and limitations of this are discussed. The primary aim of the study is to analyse the impact that different land uses, management practices and recent shifts in rainfall patterns have had on the trajectory of change in soil salinity and sodicity.

Materials and methods

Study area

The study area is 265 000 ha in size and surrounds the township of Hillston (-33.4853[degrees], 145.5328[degrees]) in the lower Lachlan River catchment of south-west NSW, Australia. The landscape consists of mostly flat alluvial floodplains, with some rocky outcrops. The study area primarily comprises the Vertosol soil type (Isbell and NCST 2016), and these are the most important for agricultural production in the region. The Vertosols can be split into Red, Brown and Grey Vertosols, ranging from the least fertile and oldest, to the most fertile and youngest respectively. There are also some less agriculturally productive, sandier soil types of alluvial and aeolian origin, which are combined together as 'non-Vertosols' for the purpose of this study.

The temperature at Hillston during summers is hot, and the winters are cool. The long-term mean annual rainfall is 372 mm, and rainfall is uniformly spread throughout the year (BOM 2017). Like much of eastern Australia, Hillston experienced the 'Millennium Drought' from 2002 to 2009, which put substantial pressure on rain-fed and irrigated farms. As the drought ceased in 2010, Hillston received almost double the average rainfall in the three consecutive years from 2010 to 2012, meaning that water was much more available again.

The land use in Hillston is diverse, and consists of areas of native vegetation, rangeland grazing, dryland cropping, irrigated annual and perennial horticulture, and irrigated cotton, with the latter the most economically important crop for the district (Fig. 1). Despite the lengthy drought, the area under irrigated horticulture and cotton expanded during the 2002 to 2015 period as a result of increased utilisation of groundwater aquifers. More information on the climate, land use and other details of the study area can be found in Filippi et al. (2018).

Soil datasets

This study used samples from 115 soil cores extracted to 1.5 m in a soil survey conducted at Hillston in 2002, along with samples from 160 soil cores extracted to 1.5 m from a repeat soil survey conducted in 2015. Most of these soil cores (n= 103) were sampled at the same spatial location in the different surveys, using georeferenced coordinates. The sampling design of both surveys is considered to be purposive, as the locations of sampling sites were not selected probabilistically. The soil cores were then subsampled into depth increments of 0-0.2, 0.3-0.4, 0.55-0.65, 0.8-0.9, 1.1-1.2 and 1.35-1.5 m for the 2002 survey; and 0-0.1, 0.1-0.3, 0.3-0.5, 0.5-0.8, 0.8-1.2 and 1.2-1.5 m for the 2015 survey. Because of the differences in these subsampling depths, equal-area quadratic smoothing splines (Bishop et al. 1999) were used on soil EC and ESP data for the 2002 survey, to standardise these to the 2015 subsampling depths.

Laboratory analyses

All of the soil subsamples were air-dried, gently crushed and passed through a 2-mm sieve. From this point on, we refer to subsamples (single depth of a soil core) as a 'sample'. To estimate soil salinity, EC was determined for 1 : 5, soil: water extracts, using a CDM 83[TM] conductivity meter for the 2002 samples (n = 548), and using a Mettler Toledo SevenCompact[TM] conductivity meter for the 2015 samples (n = 906). For all 2002 samples (n = 548) and a subset of 2015 samples (n = 138), exchangeable basic cation ([Ca.sup.2+], [Mg.sup.2+], [K.sup.+] and [Na.sup.+]) contents were determined after samples were extracted with alcoholic 1 M ammonium chloride (pH 8.5), and then analysed by atomic absorption spectrometry (Rayment and Lyons 2011). Prior to this, samples were pre-treated to remove soluble salts using a combination of 60% aqueous ethanol and 20% aqueous glycerol. The effective cation exchange capacities (ECEC) were estimated by summing the exchangeable basic cations, and ESP was calculated by dividing the exchangeable sodium content by ECEC, then multiplying by 100.

VisNIR spectroscopy analyses

Although EC is rapid and inexpensive to analyse by traditional laboratory methods, the estimation of ESP and ECEC is quite labour intensive. Consequently, not all soil samples were analysed for ESP and ECEC by traditional laboratory methods in the 2015 dataset, despite all 2002 soil samples being fully analysed by this approach (Fig. 2). Instead, both ESP and ECEC were directly predicted using VisNIR spectroscopic techniques in those samples from the 2015 survey that were not measured by traditional laboratory techniques (Fig. 2). The training dataset consisted of 385 samples from the 2002 survey, as well as 138 samples from the 2015 survey, and this was used to predict onto the remaining 768 un-analysed samples from 2015. Spectroscopic measurements were made with an Agrispec portable spectrophotometer with a contact probe attachment on dried and ground samples (Analytical Spectral Devices, Boulder, CO, USA). To reduce signal-to-noise ratios of the spectra, three scans of each sample were performed, from which an averaged reflectance spectrum was derived. Calibration of the instrument was made with a Spectralon white tile and was re-calibrated after every 15 scans, or five samples.

To ensure that this subset of 138 samples represented the remaining 768 samples from the 2015 survey appropriately, conditioned Latin hypercube (CLHC) sampling was used (Minasny and McBratney 2006). Initially, a principal component analysis (PCA) was performed on the spectra from the VisNIR wavelengths (350-2500 nm) of the 906 samples from the 2015 survey. The first two principal components (PCs) from this PCA explained >95% of the cumulative variation. We aimed to select 20 out of the 160 sites from 2015 based on the whole profile, so the first two PCs of the first four sampling depths (as not all cores had six sampling depths) were used as input criteria for the CLHC, which ensured that most of the soil profile was considered during site selection. Land use and soil type were also used as input in the CLHC to ensure that each soil-land use complex of the study area was represented. An additional 10 samples were selected at the 0-0.1 and 0.8-1.2 m sampling depths by a similar approach. The training dataset (n = 523) was then used to directly predict soil ESP on the un-analysed samples (n = 768) using Cubist models, which is a machine-learning technique (Quinlan 1986). The predictor variables included wavelengths within 500-2450 nm, averaged into segments of 10 nm, and the mid-depth of the sample, to ensure that the depth of the sample was taken into account when predicting.

To test the prediction quality of the Cubist models, 75% of the dataset was used as calibration, and the remaining 25% was used as validation, as is common practice when undertaking this procedure (e.g. O'Rourke et al. 2016). These two datasets were selected by using the VisNIR spectra and measured ESP values as inputs in a conditioned Latin hypercube, and this ensured that both the validation and calibration datasets were appropriately represented. During model testing, it was ensured that other depth samples from the same soil core for that year were not included in the calibration dataset when validating. It is assumed there is a strong correlation between these different depths of the same core and this may bias results.

Modelling and mapping


Various spatial covariates were collated to use as predictor variables in the development of the soil EC and ESP models for mapping. In total, 21 predictor variables were collated, and these can be categorised into four types: terrain attributes (n = 10), gamma radiometric data (n = 9), soil type (n = 1) and land use information (n = 1). Because there were 19 individual numerical covariates (terrain attributes and gamma radiometrics), a PCA was performed and the main PCs were used as potential predictors. The covariates used in this study and the preparation and processing of these covariates is described in greater detail in Filippi et al. (2018).

Bivariate linear mixed models (BLMMs) and statistical significance of change

The digital soil maps of EC and ESP were created with linear mixed models (LMMs). In LMMs, the response variable is modelled as a combination of fixed effects (linear relationship between the response variable (EC or ESP) and the covariates) and random effects (spatial correlation of the residuals). In this study, a BLMM is used, because two response variables are modelled in conjunction: (1) soil property data in 2002 and (2) soil property data in 2015. The parameters of the LMM were

fitted by restricted maximum likelihood, using a linear model of coregionalisation (Marchant and Lark 2007). The theory is discussed in detail in Lark and Papritz (2003), and further details of the modelling approach used in this study are presented in Filippi et al. (2018). The BLMM was fitted using customised code, and all analyses were performed in R (R Development Core Team 2017).

One BLMM for each soil property at each sampling depth was created. This allows for the covariance and correlation at colocated sites through time to be utilised (Lark and Papritz 2003). The spatial auto-correlation of the residuals for each time point, and the cross-correlation between the time points is modelled, which allows for more precise estimates of the soil property for each time point, as well as the subsequent change in the soil property over time. Despite the advantages of adopting this approach for soil-monitoring studies over univariate approaches, it is implemented in few studies (Papritz and Fliihler 1994).

The variance of the change (contrast variance) is an important element of BLMMs, and is integral in calculating a statistically significant change in a soil property over time. It is expressed as:

[mathematical expression not reproducible] (1)

where V([DELTA][??]) is the variance of the change in the soil property between two points in time, V([[??].sub.t1]) is the variance of the soil property for the observations at the first point in time, V([[??].sub.t2]) at the second point in time and V ([[??].sub.t1,t2]) is the covariance between the measurements of soil properties at the two time periods. In univariate approaches, the method that most soil-monitoring studies use, the covariance term is considered to be zero. However, in BLMMs the covariance term is included, which results in a smaller contrast variance, and therefore an increased sensitivity in predicting a statistically significant change. The statistical significance of change is analysed with the z-statistic:

[mathematical expression not reproducible] (2)

where z is the z-statistic score, [[??].sub.t1] is the value of the prediction of the soil property at time point one, [[??].sub.t2] at the second time point and V([DELTA][??]) is the contrast variance, which is obtained using Eqn 1. For a two-tailed test, a z-statistic score of greater, or less than [+ or -] 1.28, [+ or -] 1.65, [+ or -] 1.96 and [+ or -] 2.58 denotes a prediction interval of 80%, 90%, 95% and 99% respectively. Most soil-monitoring studies do not place a prediction interval around predicted change in soil attributes, but this is crucial in determining whether the change observed is statistically significant (Filippi et al. 2016). Failing to do this can lead to incorrect interpretation of the magnitude and accuracy of actual change in soil properties. Throughout this paper, any reference to statistically significant change in soil properties over time will be explicitly stated, whereas a simple use of the word 'change' refers to all detected change in soil properties over time, not solely statistically significant ones.

Model selection, map production and assessment of model quality

In this study, maps for the five depth increments to 1.2 m for EC are displayed; however, only the 0-0.1 and 0.8-1.2 m depths are shown for soil ESP, as these depths had fewer ESP values predicted by VisNIR and more samples measured for ESP by traditional laboratory methods. Models/maps of the 1.2-1.5 m layer for either soil properties are not presented, due to the small number of samples that represented the natural, dryland and horticulture land uses. Overall, one model for each of the layers for EC (n = 5) and each layer for ESP (n = 2) at both time points was created, giving a total of seven models and 14 individual soil property status maps. The soil EC data were transformed with a natural logarithm at all sampling depths because the data and the residuals lacked a normal distribution in all instances. For the ESP data, it was only necessary to transform the topsoil (0-0.1 m), and this was done using a square-root transformation. The predictor variables to be included in the final model were determined by backwards elimination, where the predictor variables with the highest P-value were removed from the model, and this continued until all predictor variables in the model possessed P<0.05. If at least one level for categorical data had P<0.05, the predictor variable was retained in the model. After the model trends were determined, the PCs (numerical covariates) that were common to both the 2002 and 2015 trend were then merged if the upper and lower confidence intervals around each associated partial regression coefficient overlapped, and the model was then re-computed. This results in a more parsimonious model and indicates that the relationship between the response and the covariate was the same between the two time periods. Filippi et al. (2018) describes the model selection process in greater detail. Once the final models were fitted, digital soil maps were produced by predicting onto a 100-m resolution grid of the study area, using an Empirical Best Linear Unbiased Predictor.

For the soil data that were transformed by applying a square-root, back-transformation was achieved by simply squaring the transformed predictions. For the data that had a natural logarithmic transformation applied, the transformed predictions were back-transformed using the following:

[y.sub.i] = exp ([x.sub.i] + ([[sigma].sup.2.sub.i]/2) (3)

where [y.sub.i] is the back-transformed prediction at point i, [x.sub.i] is the transformed prediction at point i and [[sigma].sup.2.sub.i], is the prediction variance at point i.

The quality of the models was tested by using leave-one-site-out cross-validation (LOSOCV), which involves fitting one model and then validating by predicting at a location where data from both time points for that location have been removed, which is then repeated at all locations. The standardised squared prediction errors (SSPEs) were analysed to determine if the mean and median values were close to the expected values of 1.0 and 0.455 respectively (Lark 2000; Orton et al. 2014). Lin's concordance correlation coefficient (LCCC) was used as the primary tool for assessing model quality due to the ability to compare models at different depths where the values of the response variable have a different magnitude (Lin 1989).


Soil EC summary statistics

Soil in the majority of the Hillston cotton-growing district contained low levels of salinity, and soil EC generally increased with increasing soil depth (Table 1). It is also clear that land use has had an impact on both the EC levels of soil and the magnitude of change over time. The baseline (2002) EC levels of soil under irrigated cotton production were more than double those under the natural land use in the topsoil (0-0.1 m). An overall desalination trend was observed throughout the profile in soils under irrigated cotton production in the 13-year study period, but no appreciable change in EC was observed in soils of natural sites at most depths (Fig. 3). Soils under horticultural production experienced a mean increase in salinity at all sampling depths, as well as those under dryland cropping from 0.3 m and deeper (Table 1). The deepest sampling depth of 1.2-1.5 m possessed particularly high EC values, with no discernible differences in EC between land uses at this depth (Table 1). The influence that soil type had on the changes in EC over the study period was less clear than that of the influence of land use (Fig. 3). Consistent decreases in soil EC were observed for Red Vertosol soils, while slight, but consistent increases in the subsoil were observed for the Grey/Brown Vertosols (Fig. 3). There appeared to be minimal change in EC in the upper half of the soil profile for non-Vertosols, but larger increases in EC were observed in the lower half (Fig. 3).

Soil ESP prediction by VisNIR

The quality statistics of the Cubist model, with VisNIR spectra and sample mid-depth as predictor variables, showed that ESP could be predicted to high accuracy. The validation dataset (25%) was predicted with the calibration dataset (75%) to an accuracy of LCCC = 0.79, [R.sup.2] = 0.65 and an RMSE (root mean square error) = 4.6% (Fig. 4).

Soil ESP summary statistics

In general, soils of the Hillston district were non-sodic to marginally-sodic in the upper 0.5 m of the soil profile, with mean ESP values ranging within 2.7-8.9%, and highly sodic in the subsoil (0.5-1.5 m) with mean ESP values of 12.3-15.2% (Table 2). The VisNIR-predicted ESP values generally reflected the laboratory-measured values well for 2015 samples (Table 2). Widespread trends of increasing soil ESP at most depths in the study area between 2002 and 2015 were observed at co-located sites (Fig. 5). The impact of land use on soil sodicity levels was not as obvious as it was for soil salinity, and ESP increased under all land uses at most depths (Table 2). The main soil types in Hillston exhibited different average levels of soil ESP throughout the profile (Fig. 6). The Red Vertosols consistently contained the highest ESP values at all depths in the soil profile, followed by the Grey/Brown Vertosols, with the non-Vertosols having the lowest ESP (Fig. 6).


The predictor variables that were used in the final BLMMs varied for each sampling depth and soil property, but there were particular variables that were consistently important. For the EC models, PC2 was an important predictor in all depths, and land use and soil type were included for the first three layers, but were less important in the lower depths. In the ESP models, land use, soil type and PC2 were important predictors for both depths (0-0.1 and 0.8-1.2 m). It was common for similar predictor variables to be included in both the 2002 and 2015 trend for each model. Further detail about results from the PCA of numerical predictor variables used in modelling, and the proportion that each original predictor variable contributes to different PCs is described in Filippi et al. (2018). The final models generally possessed mean and median SSPEs that were close to the expected values of 1.0 and 0.455 respectively (Table 3). The ESP models had reasonable LCCC values, and EC also showed a relatively good fit throughout the soil profile, with the topsoil EC showing the best fit of LCCC = 0.59 (Table 3).

Modelled maps

Soil EC modelled maps

In the topsoil (0-0.1 m), the alluvial soils along the Lachlan River and areas on the floodplains possessed the highest EC values (Fig. la). The change map for the 0-0.1 m depth revealed that changes in soil EC over time occurred primarily on irrigated farms, with no notable changes predicted under the natural or dryland land uses. Some cotton farms experienced a decrease in EC, and several areas under irrigated horticultural production underwent an increase in EC at the 0-0.1 m depth (Fig. 1a). The 0.1-0.3 m depth revealed very similar EC values to the 0-0.1 m maps, as well as similar levels and patterns of change over the 13-year period, but a greater number of irrigated cotton farms were predicted to have experienced a desalination trend (Fig. la). The 0.3-0.5 m depth displayed comparable spatial patterns of soil EC to those depths above (Fig. lb). Similarly, no change in EC under the natural and dryland land uses was observed at this depth, and the same desalination trend was observed at some cotton farms, but the majority of horticultural farms continued to experience an increase in soil EC at the 0.3-0.5 m depth (Fig. lb). The spatial patterns of the 0.5-0.8 m depth maps were quite contrasting to those of the upper three depths. The EC values were also substantially higher than the preceding layers. The change map showed predicted decreases in EC in the south-west of the study area, with a predicted increase in EC in the north and north-east (Fig. lb). The 0.8-1.2 m depth maps of EC also showed little influence of land use on the spatial patterns of soil EC. The pattern of temporal EC change was similar to that of the 0.5-0.8 m depth, but the predicted change was more widespread (Fig. 7c).

Soil ESP modelled maps

The difference in magnitude of ESP values at the two sampling depths was stark (Fig. 8). In the 0-0.1 m depth, higher ESP values were found on the alluvial floodplains west of the Lachlan River, although only small parts of the study area had ESP > 6% at both time points. The change map revealed that parts of the study area were predicted to have experienced an increase in ESP between the two surveys, with most of this occurring in the cotton-growing and natural areas west of the Lachlan River. There were also some very isolated areas that were predicted to have undergone a decrease in ESP in the 0-0.1 m depth. The 0.8-1.2 m depth maps displayed a similar spatial pattern of ESP to those of the topsoil, with the higher ESP values located on the alluvial floodplains; however, almost the whole study area possessed ESP >6%, apart from some isolated areas at higher elevations (Fig. 8). The change map showed that large parts of the study area were predicted to have experienced an increase in ESP over time, with this trend not confined to any particular location (Fig. 8). Although the impact of land use was still somewhat discernible in these ESP maps, it was apparent that land use did not drive the spatial patterns of ESP as strongly as it did in the EC maps (Fig. 8).

Statistical significance of predicted change

Soil EC

The five sampling depths to 1.2 m analysed in this study underwent varying directions, amounts and degrees of statistically significant change in EC (Fig. 9). Most of the statistically significant increases at the different sampling depths were confined to areas under irrigated horticultural production. Much of the statistically significant decreases in EC were located on irrigated cotton farms, as well as areas at higher elevation (Fig. 9).

Soil ESP

Although the standalone change maps showed widespread increases in soil ESP in both the 0-0.1 and 0.8-1.2 m depths (Fig. 8), only isolated parts of the study area underwent a statistically significant change in soil sodicity over time at these sampling depths (Fig. 10). The 0-0.1 m depth experienced relatively more change than the 0.8-1.2 m depth, with several locations predicted to have undergone a statistically significant increase in ESP (P-values of 0.05-0.10). There seemed to be no clear spatial pattern of statistically significant change at either depth, but most of this change occurred on either irrigated cotton farms, or soils under irrigated horticultural production.


Changes in soil salinity and sodicity

Although many semiarid regions throughout the world possess moderately to highly saline soils, the soils of the Hillston district were generally non-saline. Although soil salinity increased with depth, only isolated parts of the study area had deeper subsoil layers with EC values that would inhibit crop growth. Cotton is considered relatively tolerant to soil salinity, and can endure EC values up to 1.03 dS/m in a medium clay soil before it affects plant growth (Hazelton and Murphy 2007). However, important crops that are rotated with cotton in the Hillston district, such as wheat, are less tolerant of soil salinity, and EC>0.80dS/m in medium clay soils can impact yield (Hazelton and Murphy 2007).

The 2002 results showed that initial topsoil (0-0.1 m) EC levels of irrigated cotton sites were more than double those of natural sites, and the outline of some cotton farms can be seen in the 2002 EC maps (e.g. Fig. la). This suggests that irrigated cotton production caused an increase in topsoil salinity from natural levels pre-2002. This trend, however, did not continue into 2015, with a mean decrease in topsoil EC for most sites under cotton production. This desalination trend was also observed throughout much of the soil profile for cotton sites; however, no appreciable changes in soil salinity were observed for natural sites. In contrast, soils under irrigated horticulture generally experienced an increase in soil EC throughout the soil profile over time. Although the standalone change maps for EC showed widespread shifts for the various sampling depths (Fig. 7), not all of this change was statistically significant (Fig. 9). The z-score maps of temporal EC change indicated that both statistically significant increases and decreases occurred, but this does not necessarily mean that this change was large enough to be agriculturally important. This was particularly true for topsoil (0-0.1 m), where no parts of the study area possessed EC values that would impact plant production (Fig. 9). In the three sampling depths in the upper 0.5 m of the soil profile, most of the statistically significant decreases occurred in areas under irrigated cotton production, and all of the statistically significant increases occurred under irrigated horticulture. These changes in soil salinity are likely linked to different agricultural management practices in these production systems.

The lack of change in soils under the natural land use was somewhat expected, as it is generally accepted that these parts of the landscape are in equilibrium, and that insignificant shifts in EC would occur over a decadal time period (Hatton et al. 2003). In contrast, soil salinity is expected to be temporally less stable under irrigated land uses, as irrigation can either increase or decrease salinity levels, even over very short time periods. Studies have found that the introduction of irrigation can markedly decrease soil salinity levels (Cetin and Kirda 2003), as well as increase salinisation (Shirokova et al. 2000) over time periods of as little as two years. The quantity and quality of water applied plays an important role in determining the type and extent of these changes; low electrolyte ('clean') irrigation water can cause salts to be flushed out of the soil profile, but the use of high electrolyte ('low quality') irrigation water can increase soil salinity from baseline levels.

Although groundwater quality varies across the Hillston region, it is generally considered to be of high quality, and often contains lower levels of salt than surface water extracted from the Lachlan River. A study by Speirs et al. (2011) noted an EC level of 0.45 dS/m for groundwater in the Hillston district. Data from WaterNSW suggests that surface-water quality is slightly lower, with an average EC of 0.54 dS/m during 2002-15 (Water NSW 2017). Surface-water quality was also highly variable, with a minimum EC of 0.10 dS/m observed in 2010 (period of high rainfall), and a maximum of 1.04 dS/m in 2004 (drought) (Water NSW 2017). It is likely that the use of saltier, poorer quality surface-water used by irrigated cotton farms leading up to the 2002 survey could be the reason for the higher baseline soil salinity levels observed in the upper 0.5 m of the soil profile, compared with natural sites.

Over the 13-year study period, the source and availability of irrigation water fluctuated considerably, due to the lengthy drought and subsequent periods of high rainfall, and this varied for each water-licence holder that produced cotton. During the 2002-09 Millennium Drought in Hillston, surface water was scarce and increasingly of poor quality, and only irrigated cotton farms that had access to good quality groundwater (-50% of the total irrigated area of the Hillston district) were able to continue to irrigate. The remaining cotton farms without access to groundwater were primarily used for dryland cropping, or left in fallow. When rainfall patterns changed abruptly in 2010, there was suddenly an abundance of good quality surface water available to all irrigators in the Hillston district for several years. Most of the changes in soil EC detected under irrigated cotton farms in our study were decreases at various depths over time. This suggests that the use of good quality groundwater during the drought, followed by the use of good quality surface water after the drought broke, resulted in a general desalination trend of those soils that were kept under continuous irrigated cotton production. Comparable temporal trends in soil salinity in Vertosols were also detected in an irrigated cotton-growing system in southern Queensland for 1996-2014 (Melland et al. 2016). In this region, the use of marginal quality groundwater for irrigation had increased soil salinity levels initially, but over the subsequent two decades of higher quality surface water irrigation, much of the accumulated salt had leached out of the upper root zone and beyond one metre (Melland et al. 2016).

The contrasting shifts in soil EC of irrigated cotton and irrigated horticultural production is likely linked to the different management and irrigation practices implemented in these production systems. The horticulture land use is primarily comprised of irrigated perennial orchards that require continual irrigation for survival; at Hillston, orchard crops such as almonds receive ~1600-2000 mm of irrigation water per year, whereas cotton receives ~1000-1400mm of irrigation water during its growing season. Although horticultural and cotton enterprises at Hillston had the same access to good quality irrigation water sources during the study period, higher EC values and statistically significant temporal increases in EC were observed in soils under the horticulture land use, suggesting that the amount of irrigation water used was not a cause of salinisation. Instead, it is plausible that this salinisation trend is a result of the use of certain fertilisers containing various salts (Darwish et al. 2005), particularly potassium (K)-based fertilisers. Such fertilisers are typically not used in cotton production, but are commonly applied via drip irrigation ('fertigation') to the almond and citrus orchards in the Hillston district. Almond industry data suggests that K application rates via fertigation range within 200-580 kg K/ha, and growers are advised to apply 20-30% more K than the expected K removal in harvested fruit (150-200 kg K/ha) to account for leaching and 'nutrients being locked-up in the soil' (Rosenzweig 2012). It is this continual application of soluble salts to the topsoil that is the likely reason for the observed high levels of salinity and modelled increases over time at most sampling depths under this land use.

According to the Australian Soil Classification, a sodic soil is defined by an ESP [greater than or equal to] 6% (Isbell and NCST 2016), and a hypersodic Vertosol by ESP [greater than or equal to] 15%. Our study found only small sections of sodic topsoil across the study area, but large sections of hypersodic subsoil, especially in 2015. Although the soils in the Hillston district are inherently sodic, it is likely that irrigation practices have impacted on the degree of sodicity. In this study, widespread trends of increasing soil ESP in both the topsoil and subsoil were found across the study area between 2002 and 2015, but only isolated areas underwent a statistically significant increase in ESP at the 0-0.1 and 0.8-1.2 m depths (Fig. 10). Most of the statistically significant increases in ESP occurred in areas under irrigated cotton production or irrigated horticulture.

It is therefore speculated that during the drought period (2002-09) soil ESP gradually increased under irrigated production systems; even though generally good quality irrigation water was used, there was still a source of sodium being supplied to the soil surface which would have been absent in other dryland systems. Although much of the supplied sodium would have been flushed from the soil profile during the high rainfall years at the end of the study period, presumably a fraction of this sodium will have exchanged onto clay particles, displacing other cations and increasing the soil sodicity (Rengasamy and Olsson 1991). Any periods of irrigation with poor quality water will have heightened this effect (Rengasamy et al. 1994). A study on Vertosols under irrigated cotton production in northern NSW demonstrated that irrigation with saline water increased ESP over time (Hulugalle et al. 2002). The three years of high rainfall and good quality water availability from 2010-12 will have removed any excess soluble salts from the irrigated soils, but will have had a smaller effect on levels of sodicity (Rengasamy and Olsson 1991; Rengasamy et al. 1994). The results from our study reveal that very few soils within the region would be deemed salinesodic and are primarily only sodic. Thus the likelihood of some degree of soil dispersion after irrigation is high, which in turn may inhibit the leaching of subsequently applied sodium.

Prediction of ESP by VisNIR

The ESP is generally not well predicted by VisNIR spectra, as it does not have a primary response in the VisNIR region (Zornoza et al. 2008). Studies have reported poor prediction accuracies in the range of [R.sup.2] = 0.09-0.44 for exchangeable sodium (Stenberg et al. 2010). In contrast, ESP was predicted with relatively high accuracy ([R.sup.2] = 0.65 or LCCC = 0.79) in this study, for several plausible reasons. Rather than identifying specific areas in the spectra that correspond to exchangeable sodium, our models may be recognising areas that correspond to other soil properties that are correlated with ESP. Clay content and CEC of soil are known to be well predicted by VisNIR spectroscopic techniques and generally assumed to have a good correlation with ESP (Viscarra Rossel et al. 2006); however, we found poor (Pearson) correlations of ESP with CEC (r = 0.05) and clay content (r = 0.08) in our study, which is confounding. VisNIR identifies colour very well (Viscarra Rossel et al. 2009), and it is possible that the visible part of the spectra is contributing to these accurate predictions of ESP. At Hillston, there is a strong correlation between soil colour and ESP, with the Red Vertosols typically possessing yellow subsoils that are highly sodic, while the lesssodic soils are typically darker in colour (grey/brown) (Fig. 6).

In addition to this, the large and comprehensive calibration dataset available for the study area undoubtedly contributed to the quality of the predictions. This is likely accentuated by the resampling of the same sites in the second survey (Lark 2009), which consequently increases the probability that a similar soil would be present in the calibration dataset. Despite the relatively accurate predictions of ESP in this study, the limitations of using spectroscopic techniques in soil monitoring must be acknowledged. Although it is generally assumed that ESP is poorly predicted due to the lack of a spectral signature, and that any good predictions are a result of correlation with other well predicted soil properties, we observed that both the predicted data and laboratory-measured data showed increases in ESP over time. Time was not included as a predictor variable, which suggests that there must be some aspect of the soil that has changed over time that is correlated with ESP. Overall, the results presented here suggest that using VisNIR spectra and Cubist models to predict soil ESP shows promise for the rapid estimation of soil ESP in the future.

In our soil-mapping approach, we treated the laboratory-measured and the spectroscopic-predicted values as equal, whereas the latter contains a larger amount of error. It is possible to include the error of the VisNIR predictions of ESP for each corresponding sampling depth in the calculation of the contrast variance (Eqn 1). However, doing so would provide a conservative estimate of the change, as not all of the samples were predicted by VisNIR, and this may consequently mask some of the statistically significant change. Instead, the different uncertainty in the measurement sources could be treated more formally in the spatial modelling, as suggested by Orton et al. (2014), and should be considered in future work. This would involve giving greater weighting to the laboratory-measured data as it is a more accurate representation of actual ESP values in the soil.


The seven fitted BLMMs had modest predictive power, with the LOSOCV results showing an LCCC range of 0.41-0.59. All fitted models displayed mean and median SSPEs that were close to the expected values of 1.0 and 0.455 respectively, indicating that the prediction variances represented actual errors. Unlike most soil-monitoring studies, we took advantage of the colocation and correlation of soil information from different time points. This resulted in digital soil maps with a logical connection through time, which is particularly beneficial for soil-monitoring purposes. The advantages of using this BLMM approach for digital soil monitoring compared with conventional approaches is discussed in detail in Filippi et al. (2018).

Conclusions and future directions

This study showed that the semiarid, cotton-growing region of Hillston in the lower Lachlan catchment generally had low levels of soil salinity, with some high and potentially limiting EC levels in the subsoil. Results showed that the study area experienced both increases and decreases in soil salinity over the survey period, and all sampling depths experienced small amounts of statistically significant change. Land use played an important role in the trajectory and degree of change in salinity at the various depths, and no substantial changes in EC were observed in soils under natural land use. Overall, a trend of a decrease in salinity was observed in some areas under irrigated cotton production, with this attributed to the leaching of salts out of the profile through the application of high quality irrigation water. In contrast, several irrigated horticultural farms experienced an increase in soil EC over time at various sampling depths, with this linked to the continual use of fertilisers that contain salts. Irrigated horticulture is expanding in the study area, and soil salinity under this land use should continue to be monitored. Soil sodicity at Hillston was generally low to moderate in the upper 0.5 m of the soil profile, but very high in the lower 0.5-1.5 m depths, leading to potentially undesirable impacts on crop production. A trend of increasing soil sodicity through time was observed in some parts of the study area. Although the impact of land use on changes in soil ESP was less clear than on changes in EC, most of the statistically significant change in ESP occurred in areas under irrigated cotton and horticultural production. The observed trend of increasing ESP and decreasing EC is concerning for areas under irrigated agriculture, as the relative values of sodicity and salinity are crucial in determining soil structural integrity upon irrigation, and the changes observed could lead to heightened waterlogging issues and losses in crop productivity. It is therefore essential that soil salinity and sodicity both continue to be monitored in conjunction, to analyse the direction of any future changes. More detailed information on the quantity and quality of irrigation water applied on irrigated farms should be considered in any similar studies in the future, as this would be extremely useful in understanding the influence of different irrigation management on temporal shifts of both soil salinity and sodicity. Very few studies monitor soil salinity and sodicity, particularly in the subsoil, and further research on this should be performed, as the subsoil is often where these soil constraints are most limiting to agricultural production. This study showed that soil ESP could be predicted to relatively high accuracy using VisNIR spectra and Cubist models, which shows promise for using this as a viable approach of rapidly detecting soil ESP in the future. The spatio-temporal BLMMs used to model soil EC and ESP in this study showed relatively high predictive power, and proved advantageous for soil-monitoring purposes.

Conflicts of interest

The authors declare no conflicts of interest. Supplementary material

Tables SI and S2 are available from the Journal's website. To simplify the interpretation of the soil maps, salinity and sodicity ratings appropriated from Hazelton and Murphy (2007) were used. These ratings and the corresponding legends used in the soil maps can be viewed in the supplementary material.


The authors would like to thank the Cotton Research and Development Corporation (CRDC) for funding this research.


Bishop TFA, McBratney AB. Laslett GM (1999) Modelling soil attribute depth functions with equal-area quadratic smoothing splines. Geoderma 91. 27-45. doi: 10.1016/S0016-7061 (99)00003-8

Bureau of Meteorology (BOM) (2017) Monthly rainfall--Hillston airport. Available from: < =075032> [25 April 2017]

Cattle SR, Field DJ (2013) A review of the soil science research legacy of the triumvirate of cotton CRC. Crop & Pasture Science 64, 1076-1094.

Cetin M, Kirda C (2003) Spatial and temporal changes of soil salinity in a cotton field irrigated with low-quality water. Journal of Hydrology 272, 238-249. doi: 10.1016/S0022-1694(02)00268-8

Corwin DL, Kaffka SR, Hopmans JW, Mori Y, van Groenigen JW, van Kessel C, Lesch SM, Oster JD (2003) Assessment and field-scale mapping of soil quality properties of a saline-sodic soil. Geoderma 114, 231-259. doi: 10.1016/S0016-7061(03)00043-0

Darwish T, Atallah T, El Moujabber M, Khatib N (2005) Salinity evolution and crop response to secondary soil salinity in two agro-climatic zones in Lebanon. Agricultural Water Management 78, 152-164. doi: 10.1016/j.agwat.2005.04.020

Filippi P, Minasny B, Cattle SR, Bishop TFA (2016) Monitoring and modeling soil change: the influence of human activity and climatic shifts on aspects of soil spatiotemporally. Advances in Agronomy 139, 153-214. doi: 10.1016/bs.agron.2016.06.001

Filippi P, Cattle SR, Bishop TFA, Odeh IOA, Pringle MJ (2018) Digital soil monitoring of top- and sub-soil pH with bivariate linear mixed models. Geoderma 322, 149-162. doi:10.1016/j.geoderma.2018. 02.033

Ford G, Martin J, Rengasamy P, Boucher S, Ellington A (1993) Soil sodicity in Victoria. Soil Research 31, 869-909. doi:10.1071/SR9930869

Hatton TJ, Ruprecht J, George RJ (2003) Preclearing hydrology of the Western Australia wheatbelt: target for the future. Plant and Soil 257, 341-356. doi: 10.1023/A: 1027310511299

Hazelton P, Murphy B (2007) 'Interpreting soil test results--What do all the numbers mean?' (CSIRO Publishing: Collingwood, Australia)

Herrero J, Perez-Coveta O (2005) Soil salinity changes over 24 years in a Mediterranean irrigated district. Geoderma 125. 287-308. doi:10.1016/j.geoderma.2004.09.004

Holland J, Eastwood C (2014) An evaluation of the current understanding of cotton-growing soils and soil management practice issues in Southern NSW. NSW Department of Primary Industries. Available at <> [Verified 13 March2018]

Hulugalle NR, Entwistle PC, Weaver TB, Scott F, Finlay LA (2002) Cotton-based rotation systems on a sodic Vertosol under irrigation: effects on soil quality and profitability. Australian Journal of Experimental Agriculture 42, 341-349.

Hulugalle NR, Weaver TB, Finlay LA (2013) Soil water storage, drainage, and leaching in four irrigated cotton-based cropping systems sown in a Vertosol with subsoil sodicity. Soil Research 50, 652-663. doi: 10.1071/SR12199

Isbell RF, National Committee on Soil and Terrain (NCST) (2016) 'The Australian Soil Classification", 2nd edn. (CSIRO Publishing: Melbourne)

Jalali M, MerikhpourH, Kaledhonkar MJ, Van DerZee SEATM (2008) 95. 143-153.

Lark RM (2000) A comparison of some robust estimators of the variogram for use in soil survey. European Journal of Soil Science 51, 137-157. doi: 10.1046/j.1365-2389.2000.00280.X

Lark RM (2009) Estimating the regional mean status and change of soil properties: two distinct objectives for soil survey. European Journal of Soil Science 60, 748-756. doi:10.1111/j.1365-2389.2009.01156.x

Lark RM, Papritz A (2003) Fitting a linear model of coregionalization for soil properties using simulated annealing. Geoderma 115. 245-260. doi: 10.1016/S0016-7061(03)00065-X

Lin LIK (1989) A concordance correlation coefficient to evaluate reproducibility. Biometrics 45, 255-268. doi: 10.2307/2532051

Marchant BP, Lark RM (2007) Estimation of linear models of coregionalization by residual maximum likelihood. European Journal of Soil Science 58. 1506-1513. doi:10.1111/j.1365-2389.2007,00957.x

McKenzie D, Orange PLM (2003) Salinity and sodicity--what's the difference. The Australian Cottongrower 24, 28.

McNeal BL (1968) Prediction of the effect of mixed-salt solutions on soil hydraulic conductivity. Soil Science Society of America Journal 32. 190-193. doi: 10.2136/sssaj 1968.03615995003200020013x

Melland A, Biggs A, Silbura M, Shaw R (2016) Changes in soil salinity, sodicity and nitrate under irrigated cotton. The Australian Cottongrower 37, 26-31.

Minasny B, McBratney AB (2006) A conditioned Latin hypercube method for sampling in the presence of ancillary information. Computers & Geosciences 32, 1378-1388. doi:10.1016/j.cageo.2005.12.009

Mora JL, Herrero J, Weindorf DC (2017) Multivariate analysis of soil salination-desalination in a semi-arid irrigated district of Spain. Geoderma 291, 1-10. doi:10.1016/j.geoderma.2016.12.018

Northcote KH, Skene JKM (1972) Australian soils with saline and sodic properties. CSIRO Soil Publication No. 27, CSIRO Division of Soils, Melbourne, Victoria.

O'Rourke SM, Stockmann U, Holden NM, McBratney AB, Minasny B (2016) An assessment of model averaging to improve predictive power of portable vis-NIR and XRF for the determination of agronomic soil properties. Geoderma 279, 31-44. doi:10.1016/j.geoderma.2016.05.005

Office of Environment and Heritage (OEH) (2017) NSW Landuse 2007. NSW Government, Australia. Available at https://datasets.seed.nsw. [verified 20 Septemer 2018],

Orton TG, Pringle MJ, Page KL, Dalai RC, Bishop TFA (2014) Spatial prediction of soil organic carbon stock using a linear model of coregionalisation. Geoderma 230-231. 119-130. doi:10.1016/j.geoderma. 2014.04.016

Orton TG, Mallawaarchchi T, Pringle MJ, Menzies NW, Dalai RC, Kopittke PM, Searle R, Hochman Z, Dang YP (2018) Quantifying the economic impact of soil constraints on Australian agriculture: a case study of wheat. Land Degradation & Development, doi: 10.1002/1dr.3130

Papritz A, Fliihler H (1994) Temporal change of spatially autocorrelated soil properties: optimal estimation by cokriging. Geoderma 62, 29-43. doi: 10.1016/0016-7061 (94)900264

Quinlan JR (1986) Induction of decision trees. Machine Learning 1, 81-106. doi: 10.1007/BF00116251

R Development Core Team (2017) 'R: A language and environment for statistical computing, R version 3.2.1.' (R Foundation for Statistical Computing: Vienna)

Rayment GE, Lyons DJ (2011). 'Soil chemical methods: Australasia, 3." (CSIRO Publishing: Collingwood, Vic.)

Rengasamy P (2006) World salinization with emphasis on Australia. Journal of Experimental Botany 57, 1017-1023. doi: 10.1093/jxb/erj108

Rengasamy P, Olsson K (1991) Sodicity and soil structure. Soil Research 29, 935-952. doi: 10.1071/SR9910935

Rengasamy P, Walters L, Cresswell H, Abbott T, Lynch B, Shaw R, Churchman J, Doyle RB, Olsson K, Cochrane H (1994) Introduction to soil sodicity. Technical Report. CRC for Soil and Land Management, Adelaide, Australia.

Rosenzweig B (2012) Crop nutrition: it's not just NPK. All about Almonds Fact Sheet. Almond Board of Australia, Berri, South Australia.

Roth G (2014) Australian grown cotton sustainability report. Cotton Research and Development Corporation and Cotton Australia. Available at [verified 13 March 2018],

Shahid SA (2013) Developments in soil salinity assessment, modeling, mapping, and monitoring from regional to submicroscopic scales. In 'Developments in soil salinity assessment and reclamation: innovative thinking and use of marginal soil and water resources in irrigated agriculture.' (Eds SA Shahid, MA Abdelfattah, FK Taha) pp. 3-43. (Springer Netherlands: Dordrecht)

Shainberg I, Rhoades JD, Prather RJ (1981) Effect of low electrolyte concentration on clay dispersion and hydraulic conductivity of a sodic soil. Soil Science Society of America Journal 45, 273-277. doi: 10.2136/sssaj1981.03615995004500020009x

Shirokova Y, Forkutsa I, Sharafutdinova N (2000) Use of electrical conductivity instead of soluble salts for soil salinity monitoring in Central Asia. Irrigation and Drainage Systems 14, 199-206. doi: 10.1023/A:1026560204665

Speirs SD, Cattle SR. Melville GJ (2011) Impact of sodium adsorption ratio of irrigation water on the structural form of two Vertosols used for cotton production. Soil Research 49, 481-493. doi: 10.1071/SR11083

Stenberg B, Viscarra Rossel RA, Mouazen AM, Wetterlind J (2010) Chapter Five--Visible and near infrared spectroscopy in soil science. Advances in Agronomy 107, 163-215. doi: 10.1016/S0065-2113(10) 07005-7

Sumner ME (1993) Sodic soils: new perspectives. Australian Journal of Soil Research 31, 683-750. doi: 10.1071/SR9930683

Viscarra Rossel RA, Walvoort DJJ, McBratney AB, Janik LJ, Skjemstad JO (2006) Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131, 59-75. doi:10.1016/j.geoderma.2005.03.007

Viscarra Rossel RA, Cattle SR, Ortega A, Fouad Y (2009) In situ measurements of soil colour, mineral composition and clay content by vis-NIR spectroscopy. Geoderma 150, 253-266. doi: 10.1016/j.geoderma.2009.01.025

Water NSW (2017) 'Lachlan River at Hillston Weir (Site 412039)--Daily electrical conductivity data (2002 to 2015).' (NSW Government, Australia)

Zomoza R, Guerrero C, Mataix-Solera J, Scow KM, Arcenegui V, Mataix-Beneyto J (2008) Near infrared spectroscopy for determination of various physical, chemical and biochemical properties in Mediterranean soils. Soil Biology & Biochemistry 40, 1923-1930. doi:10.1016/j.soilbio.2008.04.003

Patrick Filippi (iD) (A,C) Stephen R. Cattle (A), Thomas F. A. Bishop (A), Matthew J. Pringle (B), and Edward J. Jones (A)

(A) The University of Sydney, School of Life and Environmental Sciences, Sydney Institute of Agriculture, Sydney, New South Wales, Australia.

(B) Department of Science, Information Technology and Innovation, Queensland Government, Australia.

(C) Corresponding author. Email:

Caption: Fig. 1. Areas under irrigated cotton, irrigated horticulture and non-irrigated land uses for the study area in the year 2015 (Office of Environment and Heritage 2017).

Caption: Fig. 2. Locations of soil cores extracted in the 2002 and 2015 soil surveys, and sites analysed by traditional laboratory methods and predicted from VisNIR spectra.

Caption: Fig. 3. Mean change in soil electrical conductivity (EC, dS/m) over time (2015-2002) of co-located sites by land use (left) and soil type (right) to 1.2 m with standard error bars (Ve = Vertosol).

Caption: Fig. 4. Observed and predicted exchangeable sodium percentage (%) of calibration (left) and validation (right) datasets predicted with VisNIR spectra and Cubist models (1:1 line overlaid).

Caption: Fig. 5. Difference in soil exchangeable sodium percentage (ESP, %) of laboratory-measured samples only at co-located sites between 2002 and 2015 for the six sampling depths.

Caption: Fig. 6. Mean soil exchangeable sodium percentage (ESP, %) of laboratory-measured samples from 2002 by soil type to 1.5 m with standard error bars (Ve = Vertosol).

Caption: Fig. 7. Electrical conductivity (EC, dS/m, 1: 5 soil: water) for 2002 and 2015 (a) for 0-0.1 and 0.1-0.3 m depths and the change in EC over time (2015-2002) (b) for 0.3-0.5 and 0.5-0.8 m depths and the change in EC over time (2015-2002) (c) for 0.8-1.2 m depth and the change in EC over time (2015-2002).

Caption: Fig. 8. Exchangeable sodium percentage (ESP, %) for 2002 and 2015 for the 0-0.1 and 0.8-1.2 m depths and the change in ESP over time (2015-2002).

Caption: Fig. 9. The z-scores showing the statistical significance of the change in electrical conductivity (EC) for the six sampling depths.

Caption: Fig. 10. The z-scores showing the statistical significance of the change in ESP for the 0-0.1 and 0.8-1.2 m sampling depths.
Table 1. Electrical conductivity (EC, 1: 5 soil: water)
statistics for all sites at different sampling depths,
time points and land uses

                         EC (dS/m)

          Depth (m)     Cotton        Dryland

                      2002   2015   2002   2015

n=           0-0.1     60     65     14     21
           0.1-0.3     60     62     14     20
           0.3-0.5     60     62     14     20
           0.5-0.8     60     62     12     19
           0.8-1.2     60     56     11     19
           1.2-1.5     54     56     4      17

Mean         0-0.1    0.22   0.16   0.13   0.09
           0.1-0.3    0.22   0.15   0.14   0.15
           0.3-0.5    0.24   0.20   0.18   0.28
           0.5-0.8    0.36   0.31   0.36   0.51
           0.8-1.2    0.61   0.48   0.45   0.72
           1.2-1.5    1.10   0.77   0.74   0.90

Median       0-0.1    0.17   0.15   0.12   0.08
           0.1-0.3    0.20   0.14   0.14   0.14
           0.3-0.5    0.23   0.19   0.16   0.28
           0.5-0.8    0.29   0.30   0.30   0.40
           0.8-1.2    0.41   0.40   0.35   0.52
           1.2-1.5    0.66   0.49   0.62   0.74

Minimum      0-0.1    0.06   0.07   0.04   0.02
           0.1-0.3    0.06   0.07   0.04   0.03
           0.3-0.5    0.06   0.07   0.03   0.02
           0.5-0.8    0.06   0.10   0.04   0.03
           0.8-1.2    0.06   0.06   0.04   0.12
           1.2-1.5    0.06   0.04   0.28   0.21

Maximum      0-0.1    0.93   0.32   0.33   0.2
           0.1-0.3    0.66   0.34   0.26   0.46
           0.3-0.5    0.54   0.44   0.54   0.74
           0.5-0.8    1.54   0.69   0.98   1.25
           0.8-1.2    2.74   1.34   1.30   1.94
           1.2-1.5    4.21   3.89   1.44   2.32

                         EC (dS/m)

          Depth (m)     Natural     Horticulture

                      2002   2015   2002    2015

n=           0-0.1     38     61     3       13
           0.1-0.3     37     61     3       12
           0.3-0.5     37     61     3       12
           0.5-0.8     33     60     3       12
           0.8-1.2     23     60     3       12
           1.2-1.5     3      52     1       12

Mean         0-0.1    0.10   0.10   0.09    0.32
           0.1-0.3    0.15   0.13   0.10    0.22
           0.3-0.5    0.25   0.21   0.13    0.23
           0.5-0.8    0.43   0.46   0.19    0.39
           0.8-1.2    0.58   0.75   0.27    0.63
           1.2-1.5    0.83   1.07   0.31    1.08

Median       0-0.1    0.14   0.12   0.11    0.21
           0.1-0.3    0.18   0.13   0.11    0.18
           0.3-0.5    0.24   0.18   0.12    0.20
           0.5-0.8    0.41   0.35   0.19    0.30
           0.8-1.2    0.68   0.48   0.28    0.37
           1.2-1.5    0.95   0.67   0.29    0.41

Minimum      0-0.1    0.02   0.02   0.06    0.10
           0.1-0.3    0.03   0.01   0.07    0.06
           0.3-0.5    0.02   0.01   0.08    0.05
           0.5-0.8    0.02   0.02   0.08    0.09
           0.8-1.2    0.02   0.02   0.14    0.05
           1.2-1.5    0.50   0.05   0.31    0.15

Maximum      0-0.1    0.19   0.23   0.12    1.64
           0.1-0.3    0.19   0.53   0.14    0.65
           0.3-0.5    1.14   0.86   0.19    0.59
           0.5-0.8    1.84   2.31   0.30    0.81
           0.8-1.2    2.32   3.90   0.38    2.64
           1.2-1.5    1.01   4.33   0.31    3.48

Table 2. Soil exchangeable sodium percentage (ESP)
statistics for 2002 and 2015 laboratory-measured (Lab)
samples, and 2015 VisNIR-predicted (Pred) samples at
different soil depths, time points and land uses

                         ESP (%)

           Depth            Cotton
            (m)      2002    2015    2015   2002
                     Lab     Lab     Pred   Lab

n           0-0.1     60      12      53     14
          0.1-0.3     60      7       55     14
          0.3-0.5     60      7       55     14
          0.5-0.8     60      7       55     12
          0.8-1.2     60      11      45     11
          1.2-1.5     53      6       50     4

Mean        0-0.1    2.6     2.2     3.7    1.8
          0.1-0.3    4.5     5.3     5.5    2.7
          0.3-0.5    8.1     9.6     12.7   5.2
          0.5-0.8    12.6    15.3    15.3   10.4
          0.8-1.2    14.2    19.6    17.2   12.6
          1.2-1.5    13.9    15.7    16.7   17.6

Median      0-0.1    2.2     2.1     3.4    1.3
          0.1-0.3    3.9     3.8     5.5    2.2
          0.3-0.5    6.9     7.4     11.7   5.8
          0.5-0.8    11.9    13.6    15.3   11.9
          0.8-1.2    13.4    21.0    17.6   15.1
          1.2-1.5    14.1    20.7    15.8   16.5

Minimum     0-0.1    0.4     1.1     1.2    0.0
          0.1-0.3    0.3     1.8     1.1    0.3
          0.3-0.5    0.9     1.3     5.1    0.2
          0.5-0.8    2.0     3.2     5.3    0.5
          0.8-1.2    2.3     3.6     0.0    0.4
          1.2-1.5    2.1     2.5     3.6    14.1

Maximum     0-0.1    10.6    3.3     9.2    5.8
          0.1-0.3    13.3    9.4     10.9   6.0
          0.3-0.5    28.1    20.0    24.4   12.2
          0.5-0.8    30.3    33.6    25.8   20.4
          0.8-1.2    25.0    35.5    28.5   20.1
          1.2-1.5    25.5    23.3    31.7   23.1

                           ESP (%)

           Depth     Dryland                 Natural
            (m)       2015     2015   2002    2015
                       Lab     Pred   Lab      Lab

n           0-0.1       4       17     38      13
          0.1-0.3       3       17     38       9
          0.3-0.5       3       17     37       9
          0.5-0.8       3       16     33       9
          0.8-1.2       4       15     23      13
          1.2-1.5       3       14     3        9

Mean        0-0.1      3.2     2.3    2.1      3.4
          0.1-0.3      7.5     4.2    4.0      7.3
          0.3-0.5     13.5     8.5    7.0     11.3
          0.5-0.8     15.5     12.2   11.0    15.5
          0.8-1.2     17.5     14.5   11.8    15.8
          1.2-1.5     14.3     19.9   15.1    14.6

Median      0-0.1      3.5     1.6    1.7      1.9
          0.1-0.3      8.1     4.5    2.8      4.8
          0.3-0.5     13.8     7.7    5.6     10.0
          0.5-0.8     14.3     12.9   8.7     18.6
          0.8-1.2     18.7     14.0   12.4    17.5
          1.2-1.5     18.6     19.4   14.4    16.2

Minimum     0-0.1      1.2     0.0    0.1      0.3
          0.1-0.3      4.8     0.0    0.3      1.0
          0.3-0.5     12.2     0.0    0.2      2.7
          0.5-0.8     13.0     2.1    0.4      3.1
          0.8-1.2     11.9     4.5    0.4      0.8
          1.2-1.5      4.4     7.5    9.5      4.3

Maximum     0-0.1      4.7     5.1    17.4    12.1
          0.1-0.3      9.5     10.4   21.4    18.9
          0.3-0.5     14.5     21.4   27.6    20.5
          0.5-0.8     19.3     25.8   29.7    28.5
          0.8-1.2     20.8     24.7   33.3    24.8
          1.2-1.5     20.0     33.1   23.2    24.4

                         ESP (%)

           Depth                   Horticulture
            (m)      2015   2002       2015       2015
                     Pred   Lab        Lab        Pred

n           0-0.1     48     3          1          12
          0.1-0.3     52     3          1          11
          0.3-0.5     52     3          1          11
          0.5-0.8     51     3          1          11
          0.8-1.2     47     3          1          11
          1.2-1.5     43     1          1          11

Mean        0-0.1    2.5    0.8        0.3        2.3
          0.1-0.3    3.8    0.5        2.6        2.9
          0.3-0.5    8.5    2.0        7.3        8.5
          0.5-0.8    10.7   6.0        6.5        11.4
          0.8-1.2    13.0   5.0        7.2        12.5
          1.2-1.5    13.7   0.7        7.7        15.5

Median      0-0.1    2.3    0.4        0.3        1.6
          0.1-0.3    3.5    0.4        2.6        2.5
          0.3-0.5    8.0    2.4        7.3        5.6
          0.5-0.8    11.2   6.9        6.5        10.8
          0.8-1.2    12.8   5.2        7.2        10.1
          1.2-1.5    13.6   0.7        7.7        14.7

Minimum     0-0.1    0.0    0.1        0.3        0.0
          0.1-0.3    0.0    0.3        2.6        0.4
          0.3-0.5    0.0    1.1        7.3        0.0
          0.5-0.8    0.0    3.9        6.5        0.0
          0.8-1.2    0.0    3.2        7.2        2.9
          1.2-1.5    0.0    0.7        7.7        4.6

Maximum     0-0.1    7.3    1.9        0.3        7.7
          0.1-0.3    8.7    0.8        2.6        5.4
          0.3-0.5    24.4   2.5        7.3        19.0
          0.5-0.8    21.0   7.2        6.5        24.7
          0.8-1.2    26.5   6.6        7.2        31.6
          1.2-1.5    22.5   0.7        7.7        24.4

Table 3. Model quality statistics for EC and ESP bivariate linear
mixed models based on leave-one-site-out cross-validation

Attribute   Soil depth   LCCC   RMSE     Mean     Median
               (m)                       (SSPE)   (SSPE)

EC             0-0.1     0.59   0.44 *   1.01     0.52
             0.1-0.3     0.48   0.50 *   1.03     0.46
             0.3-0.5     0.44   0.60 *   1.03     0.44
             0.5-0.8     0.47   0.65 *   1.01     0.46
             0.8-1.2     0.41   0.80 *   1.01     0.46

ESP            0-0.1     0.50   0.54 *   1.06     0.44
             0.8-1.2     0.54   5.48     1.01     0.42

* Root mean square error (RMSE) values are redundant as values were
transformed; LCCC, Lin's concordance correlation coefficient; SSPE,
standardised squared prediction errors
COPYRIGHT 2018 CSIRO Publishing
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2018 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Filippi, Patrick; Cattle, Stephen R.; Bishop, Thomas F.A.; Pringle, Matthew J.; Jones, Edward J.
Publication:Soil Research
Article Type:Report
Geographic Code:8AUST
Date:Oct 1, 2018
Previous Article:Influence of manure, compost additions and temperature on the water repellency of tropical soils.
Next Article:Soil erodibility affected by vegetation restoration on steep gully slopes on the Loess Plateau of China.

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters