Spatial variability and the grazing gradient method
A whole suite of methods for rangeland condition assessment has been developed using data on plant species composition (e.g. Holm, Burnside & Mitchell 1987; Bastin 1989). However, the only data set with the potential to describe rangeland conditions over large areas, which is widely and regularly available at a reasonable cost, comes from high resolution sensors on remote sensing satellites such as Landsat MSS or TM (Thematic Mapper). These data can be transformed into vegetation indices that are closely related to vegetation cover (e.g. Graetz, Pech & Davis 1988).
Although there is a well-established link between lack of vegetation cover and various forms of land degradation (e.g. Frank 1984; Warren & Hutchinson 1984; Eldridge & Rothon 1992), a reduction in cover is not, in itself, sufficient grounds to indicate a decline in land condition. Indeed, most changes in cover occur because of the short-term rainfall variability characteristic of arid and semi-arid environments. It is therefore necessary to have methods that separate grazing effects from natural change and that distinguish long-term grazing impact from that which is short-term. This leads to an approach to land degradation assessment based on loss of resilience, here defined as a reduction in the ability of a landscape to recover after it has changed (Holling 1973). This reduction arises from processes such as soil erosion, reduction in the infiltration or moisture-holding capacity of soils, loss of seed banks, and increases in unpalatable woody shrubs to the extent that herbage growth is limited (Pickup, Bastin & Chewings 1994). These processes are difficult to reverse in most landscape types and recovery from them may require from decades to centuries, even if grazing is removed.
Plant cover not only changes through time with rainfall, it also varies in space, both as a function of grazing and as a result of natural variability. The principal sources of natural variability are differences in geology, soils and geomorphic history, expressed as land units or land systems; surficial patterns of run-off, run-on, erosion and deposition expressed as erosion cell mosaics; and history of burning (Pickup 1989; Stafford Smith & Pickup 1990). In arid and semi-arid Australia, most of the spatial variability due to grazing occurs because animals are confined by fences and must rely on wells or dams for drinking water. Animal movement to and from these watering points produces radial patterns in uniform country, with grazing impact decreasing with distance from water. In non-uniform country, the radial patterns are distorted into star shapes with axes of concentration extending into those landscape types that are more preferred by animals due to more palatable forage (Pickup & Chewings 1988).
Grazing gradient methods use the spatial pattern produced by grazing animals as a spatial filter to separate the impact of grazing on vegetation cover or cover change over time from that of other factors. The methods are implemented in a Geographic Information System (GIS) environment and, in this paper, use the PD54 vegetation index derived from Landsat MSS data as a measure of cover (Pickup, Chewings & Nelson 1993). The index may also be derived from Landsat TM data (Bastin, Chewings & Pearce 1996). Essentially the index uses the data space that occurs when reflected radiance data in the green and visible red spectral bands are plotted against each other. The upper limit of this data space usually indicates bare soil, while the lower area is characteristic of areas with 100% vegetation cover. Intervening points are closely correlated with the amount of vegetation cover present and can be scaled to reduce or remove differences in brightness or greenness of ground cover. The GIS used in this study covers a 60 000 km2 area of central Australia and contains vegetation index data at 1 ha resolution spanning every major vegetation growth pulse between 1982 and 1995 as well as some earlier data. These data were acquired 6–8 weeks after major rainfalls when cover was close to the maximum, during intervening droughts when cover was at its lowest, and at intervening times. The GIS also contains information on the location of water points, fence lines and other barriers to the movement of grazing animals (e.g. mountain ranges), and a stratification of the area into different landscape types based on supervised classification of MSS data (see Bastin et al. 1993 for details). The filtering is carried out by calculating the distance from water in the GIS, dividing the distance into discrete classes, and determining mean vegetation cover or cover change in each class after stratifying by landscape type. The result can then be plotted as a graph of cover against distance from water.
Grazing gradient analyses produce different information over time. For example, changes in the shape of the cover–distance from water curve, as the landscape moves into drought and forage is depleted over time by grazing, have been used to infer animal distributions (Pickup & Chewings 1988; Cridland & Stafford Smith 1993) and differences in forage palatability (Pickup 1994). Alternatively, the shape of the curve after a major vegetation growth period can be used as an indicator of the extent of degradation, while the pattern of change from before, to the height of the growth period, shows the type of degradation (Pickup, Bastin & Chewings 1994). Until now, however, there has been no method to determine whether the amount of degradation is increasing or the landscape is recovering from past grazing management practices.
Using patterns of change through time
Plant growth in arid and semi-arid areas occurs as a series of pulses followed by periods of decline as vegetation is consumed by grazing and lost by natural decay. While the size of these pulses depends on the amount of rainfall and the pre-existing cover (e.g. Hacker et al. 1991; Hobbs, Sparrow & Landsberg 1994; Pickup 1995), it also varies with land condition. Indeed, Bastin et al. (1993) have defined rangeland degradation as a grazing-induced long-term reduction in the ability of a landscape to respond to rainfall. Other things being equal, changes in pulse magnitude through time should therefore indicate changes in landscape resilience and, if they vary systematically along grazing gradients, measure grazing-induced degradation or recovery from it. This principle provides the key to degradation assessment and the identification of trend. The next step is to turn it into a practical method given the limitations imposed by poor rainfall data and the limited number of growth pulses in the archived remote sensing data record.
There is some argument about the time scale over which landscapes are affected by, and recover from, grazing in environments characterized by episodic rainfall and short-term climatic variability (e.g. Condon 1986; Hayes 1987; Perkins & Thomas 1993; Ellis, Coughenour & Swift 1993). It is therefore important to know whether a particular sequence of growth pulses has the ability to reflect a real trend in land conditions or merely short-term variation.
A typical pattern of plant cover change through time for central Australia is shown in Fig. 1 and illustrates the variability problem. On a broad scale, there are three long periods of below-average growth resulting from major droughts starting in the late 1890s, the late 1920s and the early 1960s, and two exceptional growth pulses, one during 1920–21 and the other during 1973–75. These pulses resulted from rainfalls three times greater than the annual average (260 mm) sustained over 1–2-year periods. However, most vegetation cover change occurs with greater frequency and in response to smaller rainfall events in which 50–150% of the annual average falls in a period of a few days to several months. There were five of these events in central Australia between 1982 and 1995 and several periods of relatively minor drought. This is reasonably typical of, or perhaps slightly better than, the normal run of conditions in the 120-year record.
Figure 1. Modelled vegetation biomass in central Australia from 1880 to 1995. The data were generated from rainfall observations for Alice Springs using the model developed by Pickup (1995, 1996) and show herbage biomass on a 500-km2 paddock of nearby Owen Springs Station occupied by mixed ephemeral and perennial grasslands, shrub-dominated alluvial fans, and eucalypt and acacia woodlands.
Download figure to PowerPoint
The major growth pulses in the early 1920s and 1970s had a substantial impact on the vegetation of central Australia. Anecdotal evidence, air photographs and some observations (Foran et al. 1982; Friedel 1985; Griffin & Friedel 1985; Cunningham 1996) indicate that major recruitment of tree and woody shrub species occurred. Herbage probably also became established in areas where it had previously been lost due to grazing or during drought (Friedel 1984; Purvis 1986). Little is known about the 1920–21 pulse but comparisons of satellite imagery from 1972 and the early 1980s suggests that this was the case for the 1973–75 pulse. Even so, most long-term grazing gradients did not disappear.
While some authors have argued that the appearance of a landscape after major rainfall pulses provides the most accurate assessment of its state (e.g. Condon 1986; Hayes 1987; Palmer 1991), we suggest that assessment on this basis is skewed towards unusual climatic situations and could mask underlying trends in land condition related to anthropogenic effects. It also seems unrealistic to use growth events with a 50-year recurrence interval when most vegetation growth depends on rainfall events of a much higher frequency. We therefore argue that trends in range conditions should be defined as a consistent pattern of change in the ability of a landscape to respond to rainfall prevailing over one or two decades. Thus, the five growth pulses in the remote sensing record between 1982 and 1995 should be sufficient to identify such a pattern.
Trend detection procedure
Three problems must be overcome to detect systematic change over time in a series of growth pulses. First, the effect of initial vegetation cover on the size of the growth pulse must be removed. Secondly, the five pulses in the remote sensing record all resulted from different amounts and sequences of rainfall so there is a need to standardize the vegetation cover change data if they are to be compared. Thirdly, there are few high-quality rainfall recording stations in central Australia so the precise amount of rainfall producing a given pulse is often unknown. Also, even where rainfall is known at a specific location, it cannot be extrapolated with sufficient accuracy to describe the actual variation across the 60 000 km2 region. This prevents the use of rainfall in the standardization procedure. We dealt with these difficulties as follows.
On average, the increase in plant cover during a growth pulse varies inversely with initial cover in a more or less linear fashion (Fig. 2) (see Pickup, Bastin & Chewings 1994 for an explanation of this pattern). The slope of the relationship seems to increase with the amount of rainfall and there may also be small differences in the position of the intercept. One way of standardizing the vegetation response for different rainfalls might be to derive a relationship between rainfall magnitude and the slope of the vegetation response curve. However, where rainfall is unknown or uncertain, this is not feasible.
Figure 2. Increases in vegetation cover in relation to initial plant cover, as measured by PD54 cover values, for major rainfalls in March 1983 (▪), September 1986 (+), June 1988 (;) and March 1989 (E) for an area with sandy soils in central Australia occupied by a mix of ephemeral and perennial grasses and forbs. Data were derived from a time series of Landsat MSS data and are expressed in units of the PD54 vegetation index with values of 80 = no cover and 254 = 100% cover.
Download figure to PowerPoint
An alternative approach is to examine how the vegetation response to rainfall changes over time across a grazing gradient. For example, as a landscape becomes degraded, after allowing for variations in initial cover, we would expect the greatest reduction in response to rainfall in those areas close to water since that is where grazing impact is highest. Alternatively, where a landscape has some capacity to recover from grazing, the grazing gradient should tend to disappear after rainfall, which requires that the vegetation response is greater close to water than at a distance. These patterns can be detected without information on rainfall by calculating the ratio of the vegetation response in the high grazing impact area close to water to that in a neighbouring benchmark region where the impact of grazing is limited or significantly smaller. Where the ratio decreases over time, the grazing gradient is intensifying, and the landscape is progressively degrading. Where it increases, the landscape is showing progressively more ability to respond to rainfall in the more intensively grazed areas and the grazing gradient is disappearing. If the ratio remains constant, the landscape is neither degrading nor improving.
Some flexibility is possible in defining these areas. For example, sheep do not graze as far from water as cattle and some cattle breeds will walk further than others. For conditions prevailing in central Australia where most grazing activity occurs within 4–6 km from water, a good approximation is to use the area within 4 km of water as the high grazing impact area and select the area beyond 6 or 8 km as the benchmark. However, there are a few extreme situations where both areas need to be closer in or further out from watering points.
The method is suitable for use with the full range of grazing gradient types recognized by Pickup, Bastin & Chewings (1994) because, as airborne video data show, most of the response to rainfall over the growth periods used consists of an increase in herbage cover rather than the greening up of trees and shrubs (G. Pickup, G. N. Bastin & V. H. Chewings, unpublished data). Also, the PD54 index is useful in this regard because it was developed to detect changes in vegetation cover and to reduce the influence of simultaneous shifts in greenness on the cover signal. Concentrating on the herbage response means that the effect of degradation will be detected even with inverse grazing gradients where total cover decreases with distance from water due to an increase of woody shrubs in the vicinity of water. These shrubs suppress herbage growth, creating a reduced total vegetation response to rainfall close to water.
Implementing the method is relatively simple and uses the standard grazing gradient software described by Bastin, Chewings & Pearce (1996). Using the GIS, the vegetation response is calculated by subtracting a vegetation index map showing cover before a growth pulse from a similar map containing data on cover after the pulse. Masks are applied stratifying the response data by landscape type to remove the effects of differences in vegetation cover and soil colour and by distance from water. A vegetation response data set is then derived for the whole of each stratified area by calculating the mean vegetation response for each one of a set of initial cover classes and a weighted vegetation response ratio for each growth pulse calculated as:
in which R is the ratio; Rg and Rb are the mean vegetation response values for each initial cover class, i, for the high grazing impact and benchmark areas, respectively; Ag is the area occupied by each initial cover class; and n is the number of classes. Thus, the stratified areas remain the same over time and the whole landscape within those areas is analysed. This removes much of the localized variability associated with analysis of a limited number of individual and supposedly representative sites common in rangeland vegetation analysis (e.g. Bastin et al. 1993).
The equation allows for the effect of variations in tree and initial herbage cover on the growth pulse by calculating a ratio for each of a series of initial cover classes, weighting that ratio by area and then summing values over the range of cover classes. It circumvents the need for rainfall data by comparing data for particular water points with that of surrounding areas where rainfall should be similar. There is no need to convert vegetation index values to percentage cover or biomass because the ratio is dimensionless. The use of a ratio also reduces or removes any errors or inconsistencies caused by differences in vegetation index values calculated from remotely sensed data acquired at different times. These might arise because of problems with atmospheric correction, uncorrected sensor drift, differences in scene brightness not removed by sun angle corrections, and variations in the amount of shadow present that cannot be removed by sun angle correction, etc. We expect these problems to be minor given the rigorous standardization procedures applied to the remotely sensed data (Pickup, Chewings & Nelson 1993). However, ratioing makes for a fairly robust procedure even where such standards have not been applied.
A further advantage of ratioing is that, because it compares like with like, there is no problem with differences between vegetation response to summer and winter rainfall. These are not of great significance in the rangelands of central Australia, where lower temperatures during winter months result in germination and growth of different species rather than setting limits to increase in cover. However, they could be of importance in southern Australia where winters are cooler.