Dynamics of shrub encroachment in an African savanna: relative influences of fire, herbivory, rainfall and density dependence


  • K.G. Roques,

    1. Centre for Ecology, Evolution and Conservation, Schools of Environmental and Biological Sciences, University of East Anglia, Norwich NR4 7TJ, UK; and
    Search for more papers by this author
    • *

      Present address and correspondence: Box 192, Malkeras, Swaziland, Southern Africa ( ecology@sntc.org.sz)

  • T.G. O'Connor,

    1. Department of Range and Forage Resources, Faculty of Agriculture, University of Natal, Pietermaritzburg 3209, Natal, South Africa
    Search for more papers by this author
  • A.R. Watkinson

    1. Centre for Ecology, Evolution and Conservation, Schools of Environmental and Biological Sciences, University of East Anglia, Norwich NR4 7TJ, UK; and
    Search for more papers by this author


  • 1Shrub encroachment has been widely observed in savanna regions. This study analysed the causes of shrub encroachment in the lowveld savanna of north-eastern Swaziland, southern Africa, and highlighted management regimes that can be used to reduce or prevent it.
  • 2The rates and dynamics of shrub encroachment were quantified for the period 1947–90 using aerial photographs, and for 1997 using a ground survey. Five similar areas with different land-use histories were compared to investigate the relative importance of fire, herbivory, rainfall, soil type and shrub density in driving shrub dynamics.
  • 3In the study area as a whole, shrub cover increased from a mean of 2% in 1947 to 31% in 1990. Dichrostachys cinerea accounted for most of the increase in cover, contributing 81% to total shrub cover during 1997. Shrub cover was strongly correlated with shrub density and weakly negatively correlated with tree cover.
  • 4Shrub encroachment varied across land-use fence lines. The key determinants of shrub dynamics were grazing, through its negative effect on fire frequency and an interaction between drought frequency and high shrub cover. Browsing pressure had a significant but minor impact on dynamics, while soil type had no significant effect. High grazing pressures through their effect on fire frequency were critical throughout the study period in promoting shrub encroachment, while the interaction between drought and high shrub cover produced declines in the later stages. Browsing had an impact on encroachment only in the early stages.
  • 5Frequent fires, facilitated by low grazing pressures, were capable of preventing shrub encroachment. When coupled with drought, frequent fires could reduce high shrub densities.
  • 6As cover and density were strongly correlated, it can be inferred from the negative correlation between change in cover (density) and initial cover (density) that the rate of shrub encroachment was cover (density)-dependent and that there was a shrub equilibrium of 40% cover, approximating to 2400 plants ha−1. Shrub population growth was driven by events (fire, drought) as well as continuous agents (density dependence, mean browsing and grazing pressure).
  • 7Bush encroachment can be reversed by a combination of management (frequent fires) and climatic events (drought). The implications for savanna management are discussed.


The question of what determines the distribution and abundance of woody plants in savanna is of global interest. Recent trends towards increased shrub abundance (encroachment) in savannas have been documented in North America (Hobbs & Mooney 1986; Archer 1995), South America (Adamoli et al. 1990), Australia (Burrows et al. 1990), India (Singh & Joshi 1979) and Africa (Van Vegten 1983). In South Africa it has been estimated that 13 million hectares (ha) of savanna have been subject to thorn bush encroachment (Trollope et al. 1989). However, despite the wide extent of this occurrence, little is known about the dynamics and causes of encroachment (Archer 1995).

The time scale over which ecological change takes place poses problems (Ormerod, Pienkowski & Watkinson 1999; Watson 1999). While it has been shown that encroachment in some areas is successional and takes place over centuries (Archer 1989), reports from other areas indicate event-driven pulses of shrub recruitment (Prins & Van der Jeugd 1993). Moreover, there is debate about the reversibility of this process (Walker & Noy-Meir 1982). Past management policy has been based on the idea that the changes are reversible, but some reports argue for a stable encroached state that may be maintained even after the reversal of causal factors (Dublin, Sinclair & McGlade 1990; Jeltsch et al. 1997). Many of these dynamic analyses are based on simulations and projections (Walker et al. 1981; Dublin, Sinclair & McGlade 1990; Menaut et al. 1990; Hochberg, Menaut & Gignoux 1994; Jeltsch et al. 1996, 1997) and there is a requirement for corroboration by field study. However, the broad temporal and spatial scales associated with encroachment have made field study difficult, but these problems can be overcome through the use of satellite imagery (Russell-Smith et al. 1998) and aerial photography.

The encroachment of shrubs into grassland and savanna can alter soil moisture (Pressland 1973), nutrient and microclimate conditions (Belsky 1992), and can suppress grass productivity (Stuart-Hill & Tainton 1989). Although the factors considered as the main causes of encroachment (grazing, fire, browsing and rainfall) have been studied individually (Walker et al. 1981; Trollope 1984; Prins & van der Jeugd 1993; O’Connor 1995), their relative and combined effects are poorly understood (Peet et al. 1999) and the influence of density dependence has not been integrated into studies of encroachment.

It is widely accepted that the use of savannas for domestic livestock has facilitated shrub invasion through associated overgrazing combined with changed fire regimes (Aucamp 1976; Walker et al. 1981; Van Vegten 1983; Archer 1995; Milton & Dean 1995). This may be general enough to account for global trends, but few rigorous tests of the assumption have been reported and there have been few attempts to separate the mechanisms underpinning the influence of grazing. The overgrazing hypothesis is based on the premise that sustained heavy grazing of grasses reduces their above- and below-ground biomass (Holland & Detling 1990), leading to increased resource (especially moisture) availability for the establishment of shrubs (Caldwell et al. 1987) and reduced fire frequency and intensity (Norton-Griffiths 1979). Intense fires usually kill the aerial tissue of savanna shrubs, most of which subsequently coppice from the base (Trollope 1984). Woody meristems within the flame zone (< 5 m) are generally more exposed to fire damage than grass meristems and the latter can recover more efficiently in the short term (Trollope 1974). Frequent fires therefore benefit grasses and suppress the recruitment of mature woody plants (Gertenbach & Potgieter 1975; Watkinson & Powell 1997).

Although shrub invasion is often associated with ‘overgrazing’, high browsing pressure can, in contrast, prevent the establishment of woody seedlings (Prins & Van der Jeugd 1993) and retard the growth of shrubs (Pellew 1983), prolonging their exposure to fire and suppressing their recruitment into the mature stage. On the other hand, the browsing of seed pods and subsequent dissemination of the seed in dung can have a positive influence on shrub recruitment (Brown & Archer 1987; Reyes et al. 1994); evidence from spatial simulations indicates that seed production and dispersal are critical parameters for encroachment (Hochberg, Menaut & Gignoux 1994; Jeltsch et al. 1996). Other factors known to influence shrub encroachment include high rainfall, which may enhance the establishment of shrub seedlings (O’Connor 1995), and severe droughts, which can cause mortality among all growth stages. Competition among shrubs can lead to density-dependent mortality largely owing to limited soil moisture (Walter 1971).

This paper reports a study of shrub dynamics in the lowveld savanna of Swaziland, southern Africa, where there has been particularly severe encroachment by thorny shrubs, especially Dichrostachys cinerea (L.) Wright & Arn. The study was designed to describe the dynamics of shrub encroachment over 50 years within five areas of land use that differed in their history of grazing, fire and browsing. Together with data on temporal variation in rainfall and spatial variation in soils across the region, these data enabled investigation of the relative importance of (i) fire frequency, (ii) herbivory, (iii) rainfall history, (iv) soil type, and (v) initial shrub cover in driving shrub dynamics. Aerial photographs and ground surveys were used to estimate changes in shrub cover from 1947 onwards, and because cover and density were strongly correlated this allowed population dynamics to be inferred.

Study area

The study area encompassed 216 km2 in the north-eastern region of Swaziland, southern Africa (31°50′ E, 26°10′ S–31°57′ E, 26°23′ S).

Geomorphology and soils

The topography and geology across the study area were essentially uniform (Swaziland Government 1992, 1996); the terrain is relatively flat, approximately 240 m above sea level, and is situated on the Karoo geological system, comprising basalt with occasional dolerite intrusions (Cleverly 1979). The soils are generally dark coloured, highly fertile, but shallow (10–70 cm) clays (Harmse 1975) and can be grouped into four classes: vertisols, lithosols, red acidic loams and solonedzites. The soil type, and hence soil class, at any location was determined from the 1 : 50 000 soil map (Murdoch, Baillie & Andriesse 1968). Each class was included as a separate variable for analysis by assigning a score for a particular soil class according to the presence (score = 1) or absence (score = 0) of the relevant soils. In rare cases, where two soil classes were found on one plot, each class was assigned a score of 0·5.


Rain falls predominantly during the summer months (October–April) and fluctuates about an annual mean of 675 mm (Swaziland Government, unpublished data). Rainfall was estimated from records at the Simunye weather station (within the study area) after 1979 and, prior to that, from a combination of records at the Homestead and Nokwane weather stations (13 km and 6 km from the study area, respectively). The weather station records were used to calculate the average annual rainfall and the frequency of various high and low rainfall events for each of the four periods between census. In contrast to the considerable rainfall variability between years over the study period (Fig. 1), it can be assumed, over an area of this size, that there was a general uniformity of rainfall across the study area (Masson 1976; Tyson 1986).

Figure 1.

Annual (open circle) and 7-year moving average (closed circle) rainfall patterns from 1917 to 1997 within the study area, derived from records at Simunye, Homestead and Nokwane in north-eastern Swaziland. Census dates for shrub cover are indicated on the x-axis by an arrow.

There was no overall trend in rainfall but periods of alternate high and low rainfall occurred, each approximately 15 years in duration (Fig. 1; Tyson, Dyer & Mametse 1975). There was a rainfall peak during the 1970s marginally greater than that of any other 7-year period since 1946 and, similarly, there was a rainfall trough during the 1990s marginally greater than that of any other 7-year period since 1946 (Fig. 1).

Monthly means of daily maximum and minimum air temperatures range between 31 °C and 20·7 °C during January to 24·9 °C and 5 °C during June (Simunye sugar estates, unpublished data).


The vegetation can be classified as lowveld savanna (Compton 1966; Acocks 1988), comprising open savanna where large trees (Acacia nigrescens and Sclerocarya birrea) are scattered within a continuous grass layer dominated by Themeda triandra and Panicum maximum (Gertenbach & Potgieter 1978). According to land managers, the shrub layer was virtually non-existent prior to the 1960s, but it now forms a pervasive feature of the vegetation, with D. cinerea being by far the most conspicuous species. Species nomenclature follows that of Gibbs Russell et al. (1987, 1990).

Land use

Stone Age humans inhabited the vicinity of the study area from at least 10 000 years ago (Prior & Price Williams 1985), while Iron Age Bantu people moved into the area, bringing cattle and goats with them, approximately 800 years ago (R. Patricks & F. Okinoto, personal communication). During the late 1800s European hunters and a rindapest outbreak contributed to the decimation of wildlife in the region and a number of species went extinct locally (Reilly, Reilly & Raw 1994). The study area was concessioned out as a cattle ranch in 1902, after which the numbers of wildlife recovered despite heavy hunting in the region during the early to mid-1900s. This ranch was divided into areas of different land use around 1967, which resulted in a wealth of land-use comparisons within one relatively uniform vegetation community: 124 km2 was set aside as the Hlane Wildlife Sanctuary (Hlane site); 39 km2 remained a cattle ranch called Ndukuyamangedla that became commercialized and heavily stocked after the mid-1980s (Ranch site); 15 km2 was left relatively unused (simply because of its fenced-out location) as an area into which wild and domestic animals dispersed in drought periods (Malahleni site); and 38 km2 became used as communal land for subsistence cattle and goat farming (Communal site). During 1980 a neighbouring cattle ranch was cleared for Simunye sugar cane and residential settlement. A 0·5-km2 plot of this ranch between Hlane Wildlife Sanctuary and the settlement was fenced off in 1980 as an area devoid of herbivory (Simunye site).

Thus the original ranch was divided into five major areas of land use that differed in their histories of grazing, browsing and fire. Shrub encroachment may have been affected by limited tree felling in the communal area and altered soil moisture through irrigation of the cane lands, but these factors affected only a very small part of the study area and were not quantified.

Grazing and browsing

Grazing and browsing pressure were initially quantified in terms of live stock units (LSU) per unit area (km2) of animals that selectively feed on herbaceous plants or woody plants, respectively (Fig. 2). The grazing animals (nomenclature follows Skinner & Smithers 1990) in the study region included blue wildebeest Connochaetes taurinus, zebra Equus burchellii, white rhino Ceratotherium simum, impala Aepyceros melampus, warthog Phacochoerus aethiopicus and inguni cattle Bos indicus. The browsing animals included kudu Tragelaphus stepsiceros, indigenous goats Capra ibex and impala. Data for grazing and browsing were obtained from wildlife census or stock dip tank registers for all areas and periods, except Malahleni throughout the study period, Hlane prior to 1967, the Ranch and Communal sites prior to 1986, and Simunye prior to 1980. The known biomass records were arranged into a rank order against which each unrecorded interval and place was ranked by managers who had both historical and current knowledge. This resulted in the identification of 32 grazing ranks and 14 browsing ranks that could be related to the changes in cover at each of the various sites over the census intervals.

Figure 2.

Mean grazing pressure (open circle) and burning frequency (closed circle) patterns for the different areas of land use during the study period: (a) Communal land; (b) Hlane Wildlife Sanctuary; (c) Ranch; (d) Simunye exclosure; and (e) Malahleni buffer sites.

A prolonged drought during 1991–93 (Fig. 1) caused widespread herbivore mortality, when the number of grazing animals declined by more than 50% in the study area as a whole (see declines in herbivory for the Hlane, Ranch and Communal sites in Fig. 2). This herbivore population crash was followed by a marked increase in burning when normal rainfall resumed (Fig. 2). Similar herbivore mortality was observed specifically in Hlane during 1976 owing to ‘overgrazing’, when the number of impala, among other animals, declined from 6000 to 2540 (Reilly 1981). Herbivore population recovery from this former crash was, however, rapid; grazing pressure increased by 60% during 1976–79, so grass fuel was prevented from accumulating and burning did not resume until after the 1992 crash. There has been a general trend toward increased grazing and increased browsing over the past 50 years in all areas except the Simunye site, where herbivory declined owing to the exclusion of herbivores after 1980 (Fig. 2).


Reliable records of fire are available for Hlane since 1975 and for the Ranch, Communal and Simunye sites since 1992, enabling determination of the number of years that any particular plot had burned. For unrecorded places and periods, estimates of the frequency of fire were derived from consultation with land managers and aerial photographs taken approximately once every 5 years during 1971–90. These estimates could not be related to individual plots, but a mean fire frequency could be assigned to each area of land use.

There was a general trend towards reduced burning between 1947 and 1997 (Fig. 2). In the study area, fire frequency was negatively related to grazing pressure measured in livestock units (y = −0·0134x + 0·404, r2 = 0·48, n = 412, P < 0·001). An average plot was burnt approximately once every 3 years in the study site prior to 1973. Fire frequency declined to zero in the Communal, Hlane and Ranch sites at various stages thereafter, coincident with increases in grazing pressure (Fig. 2). The Malahleni site was frequently burnt throughout the study period, with an average plot experiencing fire approximately every 2 years (Fig. 2).


Change in woody plant cover and species composition

A stratified sampling technique was used to locate 103 randomly positioned plots within the five major land-use areas. These plots were monitored for change in woody plant cover using aerial photographs from 1947, 1971, 1979 and 1990. All photographs were of the scale 1 : 30 000 and were viewed under a stereoscope with a magnification of × 9. The absolute change in cover during each period was converted to an annual rate of absolute change.

An acetate sheet containing a 9 × 9-mm grid with 100 regularly spaced dots of 0·18 mm diameter, was lain over each of the 103 plots on each set of photographs as described by Norton-Griffiths (1979). The number of dots covering shrubs and trees, respectively, gave the percentage cover of each. If a dot covered both growth forms it was assigned that which was covered most by the dot. The grid size corresponded with a ground area of 270 × 270 m, which was therefore the size of each plot. In nearly all cases shrubs and trees were distinctly different in size and therefore easy to distinguish. A tree was considered to be a distinct woody plant with a mean crown diameter of 3·6 m or more (two-thirds the size of a dot on the grid). Shrubs were defined as a thicket of indistinguishable individuals or, more rarely, distinct woody plants with a mean crown diameter of less than 3·6 m. This definition was suited to the scattered distribution of large trees with shrubs tending to form dense thickets.

To quantify the change in cover during 1990–97, the 1990 aerial photograph estimates were compared with field-based measures taken during 1997. The latter were achieved using the line-intercept method (Walker 1976; Kent & Coker 1992). Five 25-m line transects were placed at regular intervals along the central north–south axis of each 270 × 270-m plot. The line length intercepted by shrubs and trees determined the cover of each. To facilitate comparisons between the ground-based and photograph-based estimates of cover, plants that would not be visible on aerial photographs were not included in the comparison (i.e. those situated beneath tree canopies and those concealed beneath the grass layer). The cover of these was negligible in relation to total shrub cover.

To determine the change in species composition through time, a previous survey of Hlane, undertaken in 1976 (Gertenbach & Potgieter 1978), was repeated in 1997 using the same methodology. The cover of each woody species present on 16 400-m2 plots was classed visually according to a Braun–Blanquet nine-point scale (for details see Gertenbach & Potgieter 1978). The methods of plot relocation and cover measurement were imprecise so this exercise served as a supplement to interpretation of the aerial photographs.

Plant density

To relate changes in cover (above) to some understanding of population dynamics, the relationship between shrub cover and density was determined. Thirty random plots covering a range of shrub densities were censused during 1997. The total number of shrubs that had emerged from the grass layer was counted on each plot. This was compared with the percentage cover of shrubs on each plot measured using the line intercept method already described. All shrub species were pooled for the analysis.

The 103 plots in the five land-use areas were surveyed on the ground during June 1997 to determine the distribution and abundance of uncoppiced D. cinerea individuals that were less than 1 m in height (hereafter termed recruits). Five 25 × 1-m transects were walked along the central axis of each 270 × 270-m plot, within which total counts of D. cinerea recruits were made. The abundance of recruits gave a measure of recent establishment success (Tiver & Andrew 1997).


The relative spatial uniformity of topography, geology and rainfall in the study region and the large spatial variation in land use permitted inference about the causes of change in shrub cover. The change in shrub cover on individual plots could be related to (i) the frequency of fire; (ii) herbivory as measured by grazing and browsing pressure; (iii) average patterns of rainfall; (iv) soil type; and (v) the initial cover of shrubs and trees. Some of these data were available for individual plots (i, iv, v), others for areas of land use (ii and sometimes i) and others for the region as a whole (iii).

Linear regression was initially carried out between these environmental variables (fire, grazing, browsing, rainfall, soil type, tree and shrub cover) and change in shrub cover over the entire study period. Various interactions between the variables were tested, with one being significant: the multiplication of initial shrub cover by the probability of having a year with low rainfall (less than 600 mm).

Multiple regression was then used to relate the same variables to change in shrub cover during each of the three more recent time periods (1971–79, 1979–90 and 1990–97) as well as the complete study period (1947–71 + 1971–79 + 1979–90 + 1990–97). Preliminary analysis indicated that fire frequency and grazing pressure were collinear. A principal components analysis (PCA) was therefore performed on the standard normal deviates of these two variables in order to extract two factors that were, by definition, uncorrelated. This technique is appropriate for eliminating collinearity and preferable to other techniques such as ridge regression (Draper & Smith 1981). The PCA identified a component that accounted for the correlated variation in fire and grazing (axis I, which explained 84·5% of the variance; henceforth referred to as the fire–grazing correlation) and a component that accounted for the uncorrelated variation in fire and grazing (axis II, which explained 15·5% of the variance). Grazing and fire had an equal loading of 0·919 for factor I and 0·394 for factor II; axis I scores are positively correlated with fire frequency and negatively correlated with grazing. The relative importance of variables in the multiple regression analysis was tested by comparing the magnitude of the standardized β coefficients. In a similar way the density counts of D. cinerea seedlings taken during 1997 were related to fire frequency, grazing pressure, browsing pressure, soil type and initial shrub cover for the period 1990–97.

As data on the variables in this study had been collected from the same plots during different time periods, there was potential for temporal autocorrelation in the data. This was examined for but was found to be negligible (r2 = 0·025, n = 309) owing to the long periods, and it was therefore not considered further.


Change in species composition and shrub cover

Shrub cover increased from 2% to 31% over 43 years. Dichrostachys cinerea accounted for most of the increment in shrub cover, increasing from a mean cover of 5% to 19% during 1976–97 over the study area as a whole. During 1997 D. cinerea accounted for 81% of total shrub cover. Two other species that increased were Acacia nilotica and Grewia bicolour, although their combined cover in 1997 was only 4%.

Contrasting patterns of change in shrub cover were evident on different areas of land use. During the study period as a whole, cover on the Communal and Hlane sites increased more than 13-fold, while it increased on the Ranch and Simunye sites by approximately ninefold but barely doubled on the Malahleni site (Fig. 3). The standard errors for mean shrub cover were notably greater in the communal area than elsewhere (Fig. 3), reflecting large spatial variation in cover in the communal area. The coefficient of variation in shrub cover declined with time in all areas that experienced encroachment, from a mean of 160% in 1947 to 76% in 1997, indicating that the spatial distribution of shrubs became more uniform as encroachment progressed.

Figure 3.

Mean percentage shrub cover for the different areas of land use during the study period. Bars represent standard errors for (a) Communal land (n = 9 plots); (b) Hlane Wildlife Sanctuary (n = 52 plots); (c) Ranch (n = 26 plots); (d) Simunye exclosure (n = 6 plots); and (e) Malahleni buffer (n = 10 plots) sites. The points marked with an asterisk derive from ground surveys, while all the others are from aerial surveys.

There was a strong positive correlation between shrub cover and the logarithm of shrub density (Fig. 4). This allowed density to be inferred from cover measurements and also indicated that the changes in shrub cover were primarily the result of changes in plant numbers as opposed to plant size.

Figure 4.

The relationship between shrub density (log scale) and shrub cover for the study area during 1997 (log y = 0·026x + 2·34: r2 = 0·73, n = 30, P < 0·0001).

Causes of shrub dynamics

There was a negative relationship between initial shrub cover and change in shrub cover, with areas of low initial cover (< 40%) being susceptible to recruitment and areas of high initial cover (> 40%) being prone to mortality (Fig. 5a). The data indicated a dynamic population equilibrium somewhere around 40% cover, which translates to a population density of 2400 plants ha−1, using the given relationship between density and cover in Fig. 4. The lower values for change in cover and the smaller standard errors at 0–10% cover were a consequence of a number of low cover plots showing little annual change in cover. Analysis of the residuals indicated that these were plots that were frequently burnt and lightly grazed.

Figure 5.

The relationships between the annual change in shrub cover and (a) initial shrub cover, (b) grazing pressure, (c) fire frequency and (d) drought frequency. Bars represent the standard errors of the mean. The sample size for each graph is 412 measures of change in shrub cover. Points marked with an asterisk are those where the number of data points was too low (1 or 2) to calculate a meaningful standard error.

Simple regression analysis revealed that the change in shrub cover was not related to soil type (which reaffirmed the comparability of sites) but was related to a number of rainfall and land-use variables (Table 1). There was a positive relationship between shrub encroachment and grazing pressure (Fig. 5b) and a negative relationship between shrub encroachment and fire frequency (Fig. 5c). Fire frequencies greater than one in every 3–4 years were typically associated with declines in shrub cover, while values less than this were generally associated with encroachment. Shrub cover had a negative although relatively minor impact on fire frequency (r2 = 0·14, n = 412, P < 0·001). A high frequency (P = 0·71) of low rainfall years (less than 600 mm) resulted in decline in shrub cover, while lower frequencies of low rainfall years resulted in encroachment (Fig. 5d). It must be remembered, however, that the simple linear correlations are purely descriptive as a number of the variables tend to be confounded.

Table 1.  Results of linear regression analysis of change in shrub cover with a range of variables; the statistics given are the slope, coefficient of determination and probability of the slope being significantly different from zero. Those variables that were also significant in the multiple regression analysis are shown in bold
 Fire frequency−4·77  0·15< 0·001
 Grazing pressure 0·116  0·23< 0·0001
 Browsing pressure0·319  0·13< 0·001
 Mean annual rainfall 0·0114  0·25< 0·001
 Flood frequency 5·13  0·20< 0·001
 Drought frequency4·80  0·25< 0·001
 Lithosol presence−0·243< 0·01  0·30
 Red soil presence 0·235< 0·01  0·35
 Vertisol presence 0·00053< 0·01  0·64
 Solodnedzite soils−0·163< 0·01  0·64
Vegetation cover   
 Initial shrub cover0·0487  0·20< 0·001
 Initial tree cover 0·0206  0·01  0·06
 Drought frequency × initial shrub cover0·108  0·40< 0·001
 Grazing-fire correlation (PCA axis I)0·975  0·22< 0·001

From the multiple regression analysis for the study period as a whole, it can be seen that the fire–grazing correlation (PCA axis I) and the interaction between drought frequency and initial shrub cover (Table 2) had similar magnitudes of impact on shrub dynamics. Browsing pressure was significant but less influential. Plots of the partial regressions (Fig. 6) indicated the influence of the drought–initial cover interaction, fire–grazing correlation and browsing pressure on changes in cover, having taken into account variation between plots in the other variables. High values of the drought–cover interaction, fire–grazing correlation and browsing were all associated with zero or negative changes in shrub cover. High values of the drought–cover interaction were particularly associated with decreases in cover (Fig. 6a), while intense grazing and infrequent burning were associated with increases in shrub cover (Fig. 6b). Browsing was not closely related to change in cover, but its absence was typically associated with encroachment, while very high values were associated with declines in cover (Fig. 6c). The fire–grazing correlation was important throughout, but its impact was greatest during 1971–79. The effect of initial shrub cover on the change in shrub cover was strongest during 1990–97. Browsing only inhibited shrub encroachment significantly during 1971–79, thereafter it was unimportant. The positive value of the intercept from the multiple regression (Table 2) indicated that high levels of shrub encroachment were associated with high levels of grazing coupled with low fire frequencies (low PCA axis I scores), a low frequency of drought and low browsing pressures. While a reduction in grazing and an associated increase in fire could slow and sometimes reverse encroachment, the major factor impacting on the declines in shrub cover observed during the latter periods of the study was the interaction between high shrub cover and drought frequency.

Table 2.  The regression coefficients and standardized β coefficients for the multiple regression variables, following multiple regression, indicating their relative importance in affecting shrub cover changes during the 1970s, 1980s, 1990s and the entire study period. As rainfall was essentially uniform across all plots, the analyses covering a single period of census (1971–79 1979–90 and 1990–97) could not investigate the influence of rainfall. The drought frequency–initial shrub cover interaction was therefore excluded from the regressions of a single time period and replaced with initial shrub cover. * P < 0·05, ** P < 0·01, *** P < 0·001, NS = non-significant
 Entire study period1971–791979–901990–97
Intercept  1·47***  2·69***  1·08***  1·75***
Fire–grazing correlation −0·524*** −0·962*** −0·486*** −0·466***
Initial shrub cover  −0·108NS −0·273* −0·695***
Browsing pressure −0·097** −0·352** −0·047NS −0·016NS
Drought frequency × initial shrub cover −0·606***   
r2  0·66  0·50  0·15  0·72
Figure 6.

The partial regression plots of the change in cover and (a) drought frequency × initial cover, (b) the fire–grazing correlation (PCA axis I) and (c) browsing pressure. The plots indicate the influence of each of the three variables on the changes in cover, taking into account variation between plots in the other two variables. Bars represent the standard error of the mean. The sample size for each graph is 412 measures of change in shrub cover. Points marked with an asterisk are those where the number of data points was too low (1 or 2) to calculate a meaningful standard error.

The predicted values of change in cover from the fitted regression model for the whole study period corresponded well with observed values for changes of +2% to −4% cover annum−1 (Fig. 7), indicating that the model can be relied upon to predict the response of shrub cover within this range to changes in management. Only 15% of plots experienced change in cover outside this range. For changes in cover more negative than −4% and greater than +2% annum−1, the model becomes increasingly unreliable and would result in underestimation of change.

Figure 7.

The relationship between the mean (± SD) change in cover predicted by the regression model (see Table 2) and that observed during the entire period of study.

On the 25-m transects (as opposed to the 270 × 270-m plots) shrub cover was negatively correlated with tree cover (y = −0·0706x + 1·76, r2 = 0·15, n = 356, P < 0·0001), indicating that shrub recruitment was hindered in areas of high tree cover. However, the proportion of variance accounted for by the relationship was relatively small. The density of seedlings during 1997 was not related to any variable.


During the past 40 years the lowveld of Swaziland has seen a number of changes in land use. Land clearance for sugar cane caused large numbers of game to seek refuge in a few protected areas. Overabundance of game was compounded by a no-cull policy in Hlane Wildlife Sanctuary. This, together with commercial cattle ranching and increased human settlement, coincided with a dramatic increase in the number of shrubs in remaining natural vegetation. This study documents the history of this encroachment, provides insight into the reasons behind it, and suggests some interesting dynamics.

Population dynamics

Between 1947 and 1997 there was dramatic growth in the population of shrubs and most notably of D. cinerea. The time scales over which the process took place (less than 40 years; Fig. 3) are short in historical terms and in comparison with encroachment in other regions, for example 400 years in Texas, USA (Scanlan & Archer 1991). Under high grazing pressure and low fire frequencies, initial rates of shrub encroachment were typically in the order of 1·5% annum−1 (Table 2). However, shrub cover did not usually exceed 50% and on only two plots did it exceed 80%.

The correlation between shrub cover and density (Fig. 4) implies that it is possible to interpret the changes in cover in relation to density and hence in a population dynamics framework. The implication of Fig. 5a is that the growth rate of the shrub population is density-dependent and that density dependence was especially important in limiting population growth rates during 1990–97 (Table 2). The density dependence implies an inherent population equilibrium at a cover of approximately 40% (2400 plants ha−1). In the absence of density dependence and with high grazing pressures and low fire frequencies, initial rates of shrub encroachment were typically of the order of 1·5% annum−1 (Table 2). Although there is other evidence that density dependence has an effect on the mortality and fecundity of woody plants, there is little published evidence to date to indicate that population growth rates are density-dependent (Wills et al. 1997), especially in fire-prone ecosystems (Bond & van Wilgen 1996). For savanna grasses, on the other hand, there is evidence that density dependence influences population growth rates (e.g. Sorghum intrans in Australia; Watkinson, Lonsdale & Andrew 1989). Here we have clear evidence that density dependence regulates population size in a savanna shrub.

In an analysis of the dynamics of the shrub Cytisus scoparius, Rees & Paynter (1997) concluded from a population model that the fraction of sites occupied by the shrub was dependent on the probability that a site becomes suitable for colonization following plant senescence (this can be reduced by encouraging interspecific competition), the maximum longevity of plants, and the probability of disturbance. Disturbance represents a large range of factors (e.g. fire, drought, grazing, browsing and trampling) that may cause adult mortality and also create suitable recruitment sites. The analysis here indicates that disturbance, through light grazing and the consequent build up of grass biomass and frequent fires, prevents shrub encroachment and vice versa. The fact that no relationships between land-use agents and the abundance of recruits in 1997 could be revealed, while these agents clearly influenced population dynamics (Tables 1 and 2), suggests that recruitment of shrubs into larger, grass-emergent, height classes may be a more critical parameter than actual establishment of seedlings. Moreover, interspecific competition from the trees may be viewed as a mechanism that prevents sites from being potentially invasible by shrubs, although the effect is typically weak.

Determinants of shrub dynamics

The fire–grazing correlation had a major impact on shrub encroachment (Table 2), indicating that the critical mechanism by which grazing influences encroachment is through its effect on fire. This is significant because, although grazing has previously been the focus in bush encroachment studies, its impact is often believed to operate through its effects on competition between grasses and the shrubs for soil moisture (Walker et al. 1981). The importance of the fire–grazing correlation in explaining increases in cover (Table 2) suggests rather that it is fire that plays a major role in suppressing shrub encroachment in this system. The close negative correlation between grazing and fire frequency can be explained by the fact that sustained heavy grazing of grasses removes combustible herbage, which reduces the probability of ignition and restricts the spread of fire (Norton-Griffiths 1979). The fact that encroachment occurred on the Simunye site during 1990–97 when grazing and burning were prevented, indicates that light grazing on its own, and consequently interspecific competition from the grasses, is not sufficient to limit encroachment and that it must be coupled with frequent fire. However, in arid savannas, resource competition may be a more important mechanism affecting grass–shrub dynamics (Moore et al. 1988). The augmented importance of the fire–grazing correlation during 1971–79 (Table 2) may reflect greater impact when young shrubs are prevalent and/or when rainfall is particularly high.

Frequent fires, facilitated by low grazing pressure, appear to maintain a large number of initially unencroached plots in a state of low shrub cover (Fig. 5a), indicating that the suppression of fire is required to trigger shrub encroachment. There was little evidence from the census of recruits to suggest that burning reduces the number of woody recruits. In support of this, a long-term burning experiment showed, for the same vegetation type (A. nigrescens/D. cinerea savanna), that the effect of burning on woody plants was revealed in the reduced abundance of individuals between 1·6 and 5 m in height (Gertenbach & Potgieter 1975) rather than smaller individuals. The importance of this fire–grazing correlation supports the hypothesis that recent global trends of shrub encroachment can primarily be explained by land-use trends towards increased grazing and reduced burning, a phenomenon widely observed (Noy-Meir, Gutman & Kaplan 1989; Abel 1993; Archer 1994; Milton & Dean 1995).

The interaction between initial shrub cover and drought frequency was an equally important determinant of shrub dynamics (Table 2). While high rainfall and low shrub densities are associated with high rates of shrub encroachment (see also O’Connor 1995), the regression coefficient for the interaction between shrub cover and drought (Table 2) indicates that the interaction is particularly associated with shrub mortality. Presumably competition for water at high shrub densities results in the death of plants (Figs 5a and 6a). Drought can also restrict plant growth and seed germination, further limiting increases in shrub cover. One might expect that high rainfall could positively influence fire frequency because it accelerates grass growth and the subsequent accumulation of combustible herbage (Norton-Griffiths 1979) and hence increases shrub mortality. However, no relationship between rainfall and fire frequency was evident in this study, although it may be that it was concealed by the broad time periods used. Moreover, an increase in shrub cover had a negative although relatively minor impact on fire frequency, which might be expected to result in higher shrub survival. It can therefore be inferred that the major mechanism underpinning the effect of the interaction between drought and initial shrub cover on mortality was competition among shrubs for soil moisture.

The overriding importance of shrub cover in determining decreases in plant cover during 1990–97 (Table 2), when shrub density was particularly high and droughts particularly frequent, reaffirms this conclusion. A more detailed analysis of the effect of drought might be achieved through the contrast of areas that differ significantly in rainfall. A comparison of this with other studies indicates that the regions in which encroachment proceeds rapidly are generally more mesic than the regions in which encroachment proceeds slowly (Van Vegten 1983; Scanlan & Archer 1991; Prins & Van der Jeugd 1993; Van Auken 1993). This supports the notion that high rainfall enhances shrub encroachment, presumably through its effect on shrub vigour and survival.

Browsing pressure was the third significant variable affecting encroachment but it was substantially less important overall than the first two (Table 2). Encroachment was particularly associated with the absence of browsing, while declines in cover were associated with very high levels of browsing (Fig. 6c). The only period in which browsing was significant was during 1971–79 (Table 2) when the establishment of seedlings was generally most prolific. Impala, which can curtail the establishment of woody seedlings (Prins & Van der Jeugd 1993), were abundant in the study area, especially during the early periods, but experienced a notable population crash in Hlane during 1976. Kudu and goats, which browse large plants as well as seedlings, were substantially less abundant than impala and were most numerous during the late periods (1990–97). The importance of browsing during 1971–79 might therefore be largely attributed to the influence of impala on seedling recruitment. The analysis of the distribution of recruits failed to support this hypothesis, although the effects of impala browsing may be to suppress seedling growth (and hence development beyond the 1-m height class) rather than survival. The scarcity of browsers that focus on large woody plants may account for the relative insignificance of browsing during periods when large shrubs were prevalent. If browsing was important in limiting shrub encroachment, one might expect encroachment in Hlane to be more advanced than in the communal area, following the crash in browsing animal numbers during 1976 on Hlane. Comparison of the Hlane and Communal sites (Fig. 2), however, indicates that the decrease in browsing animals had little impact on shrub encroachment. Finally it should be noted that soil type across the study area had no impact on shrub encroachment.

Management considerations

There are considerable concerns in a range of systems over how rainfall (Fensham & Holman 1999; Fernandez-Gimenez & Allen-Diaz 1999; Fynn & O’Connor 2000), grazing (Hulme et al. 1999; Bokdam & Gleichman 2000; Sternberg et al. 2000) and fire (Russell-Smith et al. 1998; Perry, Sparrow & Owens 1999) should be taken account of in terms of management. The results from this study indicate that low grazing pressure can reduce or prevent shrub encroachment. However, the effect of low grazing pressure is primarily through the increased opportunities for managers to burn more frequently. The close dependence of fire frequency on grazing emphasizes the need for managers to address these agents in tandem. Clearly this interaction needs to be the focus of control to achieve shrub population management objectives, as it is the most influential determinant of shrub encroachment. Also, the circumstantial evidence for competition between trees and shrubs implies that selective removal of trees will result in a compensatory increase in shrubs. Because substantial tree felling for fire-wood and building has been observed in certain areas outside the study area, this warrants further investigation.

Frequent fires, facilitated by light grazing, are able to prevent shrub encroachment and also reverse it, but effective reductions in shrub cover seem most likely to occur when drought acts in concert with light grazing and frequent fires (Table 2). The reversal of encroachment in Hlane after severe drought and under conditions of light grazing and frequent burning during 1990–97 (Fig. 3) reaffirms this and indicates the remarkable capacity of this vegetation to recover from shrub encroachment. But while managers can influence the fire–grazing mechanism, they have little control over drought and self-thinning. This suggests that the reversal of encroachment is possible but unpredictable, while the suppression of encroachment can be more reliably achieved, i.e. manipulating recruitment is easier than manipulating mortality. Considering an average plot, it appears that burning more often than once every 3 years would be relatively inefficient in terms of the use of grass fuel to combat encroachment (Fig. 5c). On the other hand, small changes in fire frequency below this level have large positive effects on encroachment. In an erratic climate consistent objectives are difficult to achieve, but it would appear that an average area should be burnt at least once every 4 years if the management objective is to control shrub encroachment. This corresponds to a grazing pressure of approximately 9–14 LSU km−2 (Fig. 2).

Savanna regions are particularly prone to rising human pressure on vegetation resources (Abel 1993; Milton & Dean 1995; Scholes & Archer 1997) and this study documents just one example where increases in population together with livestock grazing have resulted in dramatic vegetation change. Importantly, the study demonstrates that long-term variations in grazing, browsing, burning, rainfall and vegetation structure can be quantified to provide insight into broad-scale vegetation dynamics. It should also be noted that the management implications presented here are generic and hold across a limited range of soil types in north-eastern Swaziland. Furthermore, it highlights that although forces external to an encroaching population of shrubs can have important influences on population growth rates conferring the capacity to manipulate a system, there may be population equilibria within the system that are not readily apparent.


We thank Rob Freckleton for his advice on the statistical analysis and comments on the manuscript, Mr T.E. Reilly for facilitating the work in Hlane National Park and help with the collection of records, the members of the mapping office in the Swaziland Ministry of Public Works who assisted with aerial photographs and the loan of a stereoscope, and Mr T. Dlamini, M. Reilly and Dr P. Danso for help with collecting records. We are grateful to Mr J. Culverwell for identifying land-use contrasts and the Swaziland National Trust Commission for providing reference material. Financial assistance was provided by the University of East Anglia and the British Ecological Society.

Received 6 April 1999; revision received 7 June 2000