Description and spatial inference of soil drainage using matrix soil colours in the Lower Hunter Valley, New South Wales, Australia

Introduction

Soil colour is arguably one of the most obvious and easily observed soil morphological characteristics. Soil scientists use soil colour to differentiate genetic soil horizons as well as for the classification of soil types, e.g. Isbell (1996). From a trained or untrained eye, some inference on soils may be made from observation of soil colour in relation to organic carbon content (Schulze et al., 1993; Aitkenhead et al., 2013; Pretorius, Van Huyssteen & Brown, 2017), mineral composition (Schwertmann & Taylor, 1977), soil water content and moisture regime (Bouma, 1983; Blavet, Mathe & Leprun, 2000). Our interest in this study is making inference of a soils’ capacity to drain or soil drainage, based on observed characteristics of soil colour.

For agricultural and environmental applications, soil drainage is an important property that affects plant growth, water flow and solute transport in soils (Kravchenko et al., 2002). It has long been established that soil colour patterns can be related to a soils’ capacity to drain water (Evans & Franzmeier, 1988; Pickering & Veneman, 1984; Vepraskas & Wilding, 1983). Naturally there are exceptions to this, but often, soil colour can be interpreted as a reflection of oxidative and reductive soil processes. Reductive processes are caused by periodic or continuous water saturation. This could be due to position in the landscape (Chaplot, Walter & Curmi, 2000) and/or the presence of a permanently or fluctuating water table near the soil surface. Described in Bouma (1983), reductive soil conditions occur when a soil is saturated. Microbial activity depletes the soil of any free oxygen (O₂) causing the soil to become anaerobic. Under anaerobic conditions, and in the presence of organic carbon, ferric iron (Fe³⁺) is microbially converted to ferrous (Fe²⁺). This process is referred to as iron reduction and causes the Fe pigmented coatings on soil particles (Fe³⁺ oxides) to dissolve off the particles and into the soil solution. This results in a washed out and ultimately, grey matrix soil colour, indicating the natural colour of the soil mineral grains. In addition, other redoximorphic features such as mottling and precipitation of manganese are symptomatic of soils which experience periodic or prolonged periods of soil saturation.

Explanations for the causes of soil to remain saturated for prolonged periods include the proximity of a water table or a watercourse line. Related to these physical features is the topographical position of a particular site or landscape. For example, soil saturation occurs in the landscape (from Chaplot et al., 2004), when the accumulated water flux, the product of the catchment area A_s and the area drainage flux q, passing across an element of contour length b, exceeds the product of local soil transmissivity T and the local surface gradient S (O’Loughlin, 1986). Thus terrain attributes such as slope gradient, elevation from, and distance to watercourse lines, and terrain wetness index are generally useful for understanding, but more importantly describing the spatial variation of saturated soils in a particular landscape. The other important variable, which determines soil drainage, is related to its permeability (transmissivity of water). Soil texture and pore size distribution are principal factors which determine the ability of soils to transmit water (Bouma, 1983).

Soil drainage classes have been used widely in soil survey to characterise the wetness (drainage capacity) of soil and describe the fluctuations and proximity of the water table at site locations (Kidd et al., 2014). The distinctions between different drainage classes are based on tacit knowledge of the soil surveyor, or better, through physical measurements. These measurements may include observations of soil water tables via a well or core, and/or measurement of the soil water status. Measuring the soil water status for estimations of drainage class requires prolonged monitoring, and though possible, the procedure is complicated, costly, and time consuming (Bouma, 1983). With these logistical issues, it is not surprising that soil colour and assessment of soil redoximorphic features are often used as an indicator for making assessments of soil drainage.

The implementation of quantitative indexes of soil drainage, inferred from soil colour and/or redoximorphic features is not a new concept. Some include that of Evans & Franzmeier (1988) which requires numeric indexing of Munsell notation, in addition to information regarding mottle characteristics and abundance. Blavet et al. (2002) also numericalised Munsell colour notations in addition to using of a soil redness index (Torrent et al., 1983) to derive a continuous index for describing the duration of water-logging. Chaplot, Walter & Curmi (2000) developed a continuous index (0–100) of soil hydromorphy based on the cumulative thickness of soil horizons with redoximorphic features, combined with information regarding the Munsell Hue and Value numbers. These studies exemplify the value of using low-cost soil morphological information for making inference of soil drainage characteristics.

Our interest in this study is to develop a different type of continuous index of soil drainage. It is necessarily simple, because the soil database we are using is limited in terms of direct measurements of soil drainage and is inconsistent, even unreliable in terms of descriptions of the abundance or even presence of redoximorphic features such as mottles. In the simplest terms, the drainage index we develop in this study combines some tacit knowledge with actual observations made in the field of the soil matrix colour (each genetic soil horizon), to derive a continuous whole-profile index of soil drainage.

The motivation for deriving a soil drainage index is that we are particularly interested in understanding its spatial distribution across the landscape, as this is probably more useful from a land management and assessment perspective. Studies such as Schaetzl et al. (2009) demonstrate this. Furthermore, after a number of years surveying the area described in this study, we have developed a mental concept of how soil drainage varies across the landscape. It is a useful exercise to validate such mental models with empirical information. Given the relationship between topography and soil saturation, there is considerable benefit in applying digital soil mapping methods (McBratney, Mendonca-Santos & Minasny, 2003) for inferring the spatial distribution of soil drainage. A number of studies have constructed soil spatial inference models of soil drainage class using topographical variables (Kravchenko et al., 2002; Campling, Gobin & Feyen, 2002). Bell, Cunningham & Havens (1992) used multivariate discriminant analysis using topography and geological information to spatially predict drainage classes. Chaplot et al. (2004) were interested in mapping the soil hydromorphic index using topographic indices derived from the land surface and saprolite upper boundary. Less invasive techniques of mapping soil drainage classes through the use of remote sensing platforms have also been demonstrated (Peng et al., 2003; Cialella et al., 1997).

The aims of this study are threefold: (1) To develop an index of soil drainage combining tacit knowledge and empirical information of soil matrix colour; (2) to determine whether an empirical relationship exists between the estimated drainage index and landscape features; and (3) to develop a soil spatial prediction function for estimating the spatial distribution of soil drainage across our study area in the Lower Hunter Valley, NSW (New South Wales, Australia).

Materials and Methods

Study area

The area of this study is the Hunter Wine Country Private Irrigation District (HWCPID), situated in the Lower Hunter Valley, NSW. The HWCPID covers approximately 220 km² and encompasses the localities of Pokolbin and Rothbury, NSW (32.83°S 151.35°E), which are approximately 140 km north of Sydney, NSW (Fig. 1). Topographically, this area consists mostly of undulating hills that ascend to low mountains to the south–west. The underlying geology of the HWCPID is predominantly Early Permian, with some Middle and Late Permian formations. Described in the Newcastle Coalfield Regional 1:100,000 Geology Map (Hawley, Glen & Baker, 1995), the most extensive formation is the Rutherford Formation (Early Permian) which consists of siltstones, marl, and some minor sandstone. Much of the southern and eastern part of the HWCPID is underlain by the Rutherford Formation. Other extensive formations include: the Mulbring Siltstone (Late Permian siltstones), the Branxton Formation (Middle Permian conglomerates, sandstones, and siltstones) and the Farley Formation (Early Permian silty sandstones). These formations occupy the north-western extents of the HWCPID. In terms of land use, dryland agricultural grazing systems are predominant, followed by an expansive viticultural industry. While most of the land has been dedicated for these uses, tracts of remnant natural vegetation (dry forest) are apparent, particularly towards the south-western area—which is bordered by Broken Back Range, Werakata National Park situated to the east, and some areas situated in the northern extents.

Our knowledge of the soils across the HWCPID was first informed from legacy soil survey which is described in detail within the Soil Landscapes of the Singleton 1:250,000 Sheet Map and Report (Kovac & Lawrie, 1990). This knowledge has since evolved through annual soil surveying campaigns by students and members of our research group, which began in 2001 and continue to the present time. These annual surveys, while concentrated to the south of the study area, form a densely populated database of soil information and descriptions. This information and soil knowledge has been supplemented with two area-wide soil surveys of the HWCPID which have been described in Malone et al. (2011) and Odgers, McBratney & Minasny (2011). Based on these various soil surveying campaigns we have found—based on the sub-order level of the Australian Soil Classification system (Isbell, 1996)—that the most dominant soils across the HWCPID are both Brown and Red Dermosols and Chromosols. Generally, Dermosols and Chromosols are the most prolific; there are few Kurosols by comparison. Hydrosols and Rudosols are few, but generally concentrated near watercourse lines. Calcarosols are also few, yet exist in areas where the Rutherford Formation exists, particularly where the occurrence of the calcareous marl parent material is present. Corresponding WRB (Food and Agriculture Organisation (FAO), 1998) soil classes to the ASC soil orders are: Calcisols (Calcarosols), Luvisols (Chromosols and some Dermosols), Acrisols (Kurosols and some Dermosols), Fluvisols (Hydrosols), and Regosols (Rudosols).

Results

Pearson’s coefficient of correlation between the derived soil profile drainage index and each of the covariate data sources from highest to lowest were: TWI (−0.34), MSP (−0.29), MRVBF (−0.29), VDCN (0.26), SH (0.22), SL (−0.18), S (0.11), E (0.09), and AH (−0.03). These correlation coefficients indicate some general features of soil drainage in the HWCPID, for example there is a positive correlation of the drainage index with vertical proximity to watercourse lines. Indices such as TWI and MRVBF, which inform us about the hydrological characteristics of the area, are negatively correlated with the drainage index. Thus based on landscape position, where the soil is more prevalent to concentration of water, the drainage index is also lower. The correlations of the slope indices S, SH, and MSP with the drainage index indicate a relationship whereby gentle slopes (relatively low S and SH, and high MSP) soils more likely to have a lower drainage index. Similarly, longer slopes (SL) result in a negative correlation with the drainage index.

Fitting of the Cubist model to the calibration data resulted in the partitioning of two simple rules for the spatial distribution of the drainage index. Each rule defining a different regression model:

• Rule 1

  if VDCN ≤ 7.8 m

  then

  DI = 6.58 + 0.06(VDCN) − 0.20(TWI) − 0.02(E) − 0.80(AH)

• Rule 2

  if VDCN ≤ 7.8 m

  then

  DI = 6.58 − 0.11(TWI) − 0.01(E) − 0.09(MSP) + 0.001(VDCN)

These simple linear regressions are pre-empted by a recursive split of all data based on a threshold value of 7.8 for the VDCN. Essentially this means that vertical proximity to a watercourse line is a defining characteristic of soil drainage. Common parameters to each linear model were VDCN, TWI, and E, while AH was only included in the first rule and MSP was only included in the second rule.

Examining where each Cubist rule was applied shows clearly the relationship of the rules with proximity to watercourse lines (Fig. 2A). For approximately one-third of the area, rule 1 was applied.

The associated drainage index map which resulted from the regression kriging model is shown in Fig. 2B. With the blue lines indicating the watercourse lines, it is clear from the map that proximity to them has a considerable effect on the soil drainage. From a basic statistical analysis, where rule 1 was applied, the mean drainage index was 2.70 with 95% of the area between 1.85 and 3.60. Where rule 2 was applied, the mean drainage index was 3.40 with 95% of the area between 2.51 and 4.00.

Validation of the regression kriging model based on 446 withheld data indicated a RMSE of 0.90, meaning that, the predictions of the drainage index on average deviate approximately 0.9 away from the observed value. The concordance between the observed and predicted values was a reasonable 0.49. The plot in Fig. 3 show the observations and corresponding predictions with respect to the 1:1 relationship (draw as a red dashed line). Predictions appear to be strongest around drainage index values between 2.5 and 4.0. From visual inspection of the plot it is clear there does not seem to be any bias, such as prevalence for over or under predictions (mean error was calculated as −0.08). A linear model fitted to the observed and fitted data resulted in a coefficient of determination of 31%, indicating a reasonable correlative relationship (green solid line).

Discussion

The continuous index of soil drainage proposed in this study requires little information other than tacit knowledge of soil drainage spatial variability, and observed soil matrix colour descriptions. The fundamental limitation of this is that we have assumed there is a direct correlation between soil colour and soil drainage. There is good physical evidence though to support this relationship (e.g. Bouma, 1983; Evans & Franzmeier, 1988). It is likely, since we have been able to establish a correlative relationship between our index of soil drainage and some topographical variables, that some validation in terms of soil water measurements or measurements of water table proximity is warranted in further studies.

There is significant value from the land management perspective in quantifying soil drainage as a spatially continuous variable across the landscape. In this study, we have avoided mapping the well-known and established drainage classes. While there may be value in doing this, the grading between classes is qualitative and in the absence of direct measurement, allocation to a particular class is a subjective designation. Nevertheless, it was by necessity (due to the data used) that we had to develop our own model, which by default treats soil drainage as a continuous variable. Subsequently, mapping the continuously varying drainage index across the HWCPID revealed spatial patterns more attuned to what one would observe in the landscape; that is, continuously varying rather than discreetly apportioned. In the HWCPID, it is believed that discreet variations of soil drainage are the exception rather than the rule.

On the basis of using digital soil mapping methods, we have been able to validate quantitatively the spatial model of soil drainage. Comparatively with other digital soil mapping studies, the results found in this study are acceptable (Grunwald, 2009). Other studies that have examined the relationship between soil drainage and environmental information have reported stronger correlations than that reported in this study (Campling, Gobin & Feyen, 2002; Chaplot et al., 2004). It is suspected that scale may be one cause for this discrepancy. For example, we have attempted to describe the variations of soil drainage over a much larger area of land which we know to be rather complex (topographically and lithologically). While these results may be improved upon, the map, from a soil surveyor’s perspective, adequately coincides with the knowledge we have developed over the years of survey in the HWCPID.

In terms of the spatial prediction model, we used the Cubist models as an attempt to mirror what a soil surveyor would observe in the landscape. That is, given particular combinations of features or characteristics of the landscape, a particular soil or characteristic of the soil will behave similarly. The quantitative interpretation of this and what was found in this study, was that vertical distance to a channel network was a divisive and important physical attribute determining the estimation of soil drainage; such that, given a certain threshold, different predictive models were applied. More generally, we have found that using such rule-based spatial prediction functions makes them more interpretable (from the soil survey perspective) and particularly useful for digital soil mapping.

Correcting the deviation between what was observed (from the data) and what was predicted using a spatial model is a worthwhile pursuit. What was clear in this study is that the observed variations of soil colour described something much more complex than what the spatial model was able to describe. There could be many reasons and explanations for this. One of them is that soil colour alone and attribution thereof can have significant influence on the interpretation of soil processes. While fuzzy set theory is embedded within our model, which by definition embraces the subjectivity around soil colour attribution, our model is by no means immune to poorly attributed soil colour descriptions. Ultimately, this can have flow-on effects when soil colour is then used in some sort of quantitative model, e.g. soil drainage (Chaplot et al., 2004). O’Donnell et al. (2010) have proposed a standardised procedure of soil colour attribution based on image processing. Or perhaps usage of a colorimeter would be enough to make standardised assessments of soil colour. Currently, standardised assessments of soil colour are not made during soil survey around the world, so it is likely it will be some time before we can test the applicability of our method using such assessments.

The spatial model of soil drainage in this study principally used topographical variables as predictive covariate information. We also incorporated a regression kriging model with the intention of further modelling spatial trend that was not detectable from the topographic information. Regression kriging made some improvement of the prediction in comparison to just using a deterministic model. Nevertheless, due to a limitation in the availability of additional sources of predictive information, we were unable to explore more complex relationships of soil drainage with other environmental variables. For example, parent material or underlying geology has been shown to be a useful variable (Bell, Cunningham & Havens, 1992). Intuitively, different lithologies will impart differing soil physical characteristics, such that the drainage characteristics of a soil developed from limestone will be different from those developed on siltstone or sandstone etc. The best available geological survey of the HWCPID (1:100,000; Hawley, Glen & Baker, 1995) informs us that while siltstones are the most predominant lithology, there are also sandstones and silty sandstone parent materials. A limitation of our mental model of soil drainage is that it has been refined where the lithology is predominantly siltstone—as most of the soil sampling has been conducted on this lithology. There is potential bias regarding estimation of soil drainage to contend with where other lithologies are found. However, the question is whether the current geological survey could be used to refine our model of soil drainage? It is unlikely that it would, because while instructive, it is neither comprehensive or of the appropriate scale. Furthermore, soil processes such as colluviation and alluviation have often created soil profiles of complex and mixed lithology that is near impossible to disentangle from geological survey maps. We envisage that in the future, gamma-radiometric survey will provide us with information regarding the lithology and lithological processes at the necessary detail to be included within our soil drainage model. Gamma radiometry refers to the measurement of naturally occurring gamma radiation which is emitted from the ground surface (Cook et al., 1996). Such information has been shown to describe the distribution of soil-forming materials and weathering processes over large areas (Wilford, 2012).

In the absence of detailed lithological information, a pragmatic solution may be to examine whether the digital mapping of soil texture grades (or soil variables derived from them such as bulk density etc.) are useful for interpreting variations of soil drainage. In the HWCPID, where soil textures are recorded predominantly as that derived from hand bolusing, we need to explore methods of how to incorporate these data within a digital soil mapping framework.

Conclusion

By necessity of the data available, we have developed an index of soil drainage which incorporates tacit knowledge of the soil surveyor and observed soil matrix colour. Soil drainage is evaluated as a whole-profile, weighted combination of the soil colour at each generic soil horizon. Fuzzy set theory is built into the drainage index model as a means to dampen the subjectivity of soil colour attribution. We believe the approach can be generalised to other areas once the unique soil colour and soil drainage relationships have been defined by an expert.

In our study, we found that the topographical variables most strongly correlated with soil drainage are TWI, MSP, MRVBF, and vertical distance above channel network. Cubist models were used to model the relationship of the drainage index with a suite of topographic variables with the dual purpose of understanding the spatial variation of soil drainage and to validate our mental model of soil drainage developed over the years from successive field surveys. Validation of the spatial model of soil drainage was adequate in consideration of the scale of mapping and nature of the data. The associated map corresponds meaningfully to what we have generally observed in the field. The incorporation of new information specifically from gamma-radiometry or soil texture may be useful solutions in improving our understanding of soil drainage in the HWCPID.

Supplemental Information