Applied Geography 53 (2014) 389e401
Contents lists available at ScienceDirect
Applied Geography journal homepage: www.elsevier.com/locate/apgeog
Climate variability as a dominant driver of post-disturbance savanna dynamics Cerian Gibbes a, *, Jane Southworth b, Peter Waylen b, Brian Child b a
Department of Geography and Environmental Studies, University of Colorado, Colorado Springs, 1420 Austin Bluffs Parkway, Colorado Springs, CO 80918, USA b Department of Geography, University of Florida, TUR 3141, Gainesville, FL 32611, USA
a b s t r a c t Keywords: Precipitation variability Thresholds NDVI Savanna Dryland systems Time-series analyses
How do climate variability and climate shifts influence vegetation patterns in dryland ecosystems? A landscape scale assessment of the effect of changes in precipitation on vegetation in three watersheds, spanning four southern African nations is undertaken and tests of statistical significance of vegetation changes developed. Concepts of resilience provide the framework for examining the influence of changes in both the mean and variance of annual precipitation on the ecological systems of southern Africa. They illustrate thresholds of change and their manifestation in the state of savanna ecosystems, which could previously only be postulated. Time series analyses indicate the fundamental role of precipitation mean and variability in modulating the states of savanna ecosystems. Here we show that the savanna dryland ecosystems have overall responded with increasing NDVI measures despite decreased precipitation since the mid to late 1970's. Areas which experienced diminished vegetative cover over time, are related to specific vegetation types, and declines in the variance of precipitation (even in the presence of overall increases in annual mean). The work highlights the importance of time-series analyses and explicit vegetation-precipitation linkages across this highly vulnerable region. © 2014 Elsevier Ltd. All rights reserved.
Introduction Ecosystems often occupy multiple states in which climate influences functioning and movement between states (Holling, 1973). In savanna landscapes this property is dependent primarily upon precipitationevegetation relationships (Meyer, Wiegand, Ward, & Moustakas, 2007; Zen & Neelin, 2000). Spatio-temporal trends in vegetation change can now be measured at landscape-scales and over extended temporal scales using time series of remotely sensed data that improve opportunities for the development of linked conceptual models and methodological approaches. Additionally, the testing of the statistical significance of landscape change over these time periods offers a unique approach to assessing and potentially identifying landscape changes which are (or lead to) permanent shifts in ecosystem states. General impressions of change can be evaluated objectively through a series of statistical tests developed from simple random walk processes and
* Corresponding author. E-mail address:
[email protected] (C. Gibbes). http://dx.doi.org/10.1016/j.apgeog.2014.06.024 0143-6228/© 2014 Elsevier Ltd. All rights reserved.
assumptions about the nature of the statistical distribution of values of remotely sensed indices of vegetation, which determine the probabilities of positive (increasing) and negative (decreasing) changes in such measures. This research develops a temporally and spatially multiscale understanding of the relationships between the heterogeneity and quantity of vegetation cover, and a widely observed shift in global climate. The work highlights the importance of time-series analyses and explicit vegetation-precipitation linkages across the world's highly vulnerable dryland regions (Hirota, Holmgren, Van Nes, & Scheffer, 2011; Staver, Archibald, & Levin, 2011). An expanded understanding of the response of this major ecotype to changes in the amount and variability of precipitation will improve our ability to adapt to the significant climate change, which is currently predicted for dryland regions globally (IPCC 2007). Savannas are important sustainers of human livelihoods, and act as repositories of biodiversity. We assess empirically how climate variability and shifts, as measured by changes in the mean and variability of precipitation affect the ability of different dryland ecosystems (savanna, miombo, Kalahari woodland) to recover from shocks, and to persist. Building on this resilience analogy, we develop a conceptual model, which is applied regionally to the
390
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
savannas of southern Africa but possesses potential global implications, and evaluate its ability to explain observed changes. Using the concepts of non-equilibrium and resilience we propose a model (Fig. 1) whereby the response of landscapes to a shock (e.g. a persistent meteorologic drought) is mediated by changes in the mean and/or variance of precipitation. Combinations of changes in precipitation characteristics influence the long-term vegetation patterns. As might be expected, increases in mean annual precipitation (MAP) can yield landscapes with continuously increasing vegetation amounts (greens Fig. 1). However, when MAP increases are small relative to increased inter-annual variability, the latter dominates the precipitationevegetation relationship, resulting in a landscape in which vegetation declines continuously (purples Fig. 1). Combined changes in precipitation characteristics lead to continuously increasing (greens Fig. 1) or decreasing (purples Fig. 1) vegetation amounts, assuming that critical thresholds are crossed; otherwise the system will remain in the same state and small net changes in vegetation are observed (yellow). The relative sensitivity of the system to changes in precipitation properties is also dependent upon the absolute level of mean precipitation prior to the “shock”. The centrality of vegetation amount and condition in moderating ecosystem status yields a quantifiable system characteristic, which can be monitored to identify potential shifts across thresholds. Vegetation characteristics, including quantity and heterogeneity of vegetation cover, serve as useful measures of landscape-scale changes in response to shocks to the ecosystem (Washington-Allen, Ramsey, & West, 2004; Washington-Allen, Van Niel, Ramsey, & West, 2004; Washington-Allen, West, Ramsey, & Efroymson, 2006). Although vegetation attributes are most frequently considered with regard to a reference state (Washington-Allen, Ramsey, West, & Norton, 2008), examinations of post-shock landscape dynamics offer valuable insights to system shifts and regional-scale recovery. Specifically, increases in variance within system properties, such as shifts in vegetation types, have been found to link directly to ecological transitions (Carpenter & Brock, 2006). Only recently have remote sensing records permitted the systematic quantification of changes in vegetation over sufficient temporal and spatial scales, which can be used to begin to operationalize ‘state-and-transition’ or ‘multiple stable state’ models as presented by Holling (1973). These advances in geospatial technologies combined with theoretical underpinning of ecosystem functioning hold potential to improve understandings of the
impact of a changing climate conditions on the landscape (VicenteSerrano et al. 2012). Landsat data are of limited application in biologically complex systems (Lunetta, Knight, Ediriwickrema, Lyon, & Worthy, 2006), and although used frequently to quantify land cover change, difficulties in obtaining series of images representative of similar phenological period ensure that results must be interpreted with caution. However, monthly time series of AVHRRand MODIS-derived vegetation indices, available from the early 1980s, are useful for the large-scale quantification of vegetationclimate dynamics, particularly in arid and semi-arid regions (see for example, Yang, Wylie, Tieszen, & Reed, 1998). Additionally, the scale of AVHRR and MODIS data is suitable for regional analyses of landscape change related to global environmental changes (for example, Poveda et al. 2001). To operationalize the conceptual model (Fig. 1) we employ a methodology which quantifies spatial and temporal changes in vegetation (Lanfredi, Simoniello, & Macchiato, 2004; Westman & O'Leary, 1986) and interpret responses to the globally recognized climate shift of the mid to late 1970's (Chavez, Ryan, Lluch-Cota, & Niquen, 2003; Nicholson, 2000), which is manifested regionally as a change in precipitation regime. Using a novel time-series approach more familiar in climate analyses than those of landscape change assessments, productivity and variability in response to the climate shift are conducted. The spatial persistence of vegetation productivity is interpreted in terms of changes in levels of precipitation and its interannual variability. The method allows both temporal and spatial specificity at a scale unusual in resilience or land change science studies. Additionally, we construct and apply statistical tests at the landscape level, but implemented at the pixel scale, to assign real statistical significance to the changes seen. To illustrate and evaluate the potential of the techniques we investigate a diverse region in Southern Africa, straddling a noted precipitation threshold (Sankaran et al. 2005) in the environmentally critical savanna biome. Savanna vegetation ecosystem processes and their phenological expression are heavily influenced by climate variability. Many studies consider changes in atmosphere-ocean interactions in the Pacific basin during the mid-1970s to have influenced global pre~ o-Southern Oscillation cipitation, the characteristics of the El Nin (ENSO) phenomenon, and its associated climate patterns (Ebbesmeyer, Coomes, Cannon, & Bretschneider, 1991; Graham, 1994; Namias, 1978; Trenberth, 1990; Trenberth & Hurrell, 1994). The shift has been detected in many climate and biological time
Fig. 1. Conceptual diagram of the effect of changes in mean and variance of annual precipitation on the Normalized Difference Vegetation Index (NDVI), which is used as a proxy for the response of a landscape to environmental shocks. The exact nature of the response depends upon the changes in mean and variance, and also upon the original level of mean annual precipitation (MAP).
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
series (Mantua, Hare, Zhang, Wallace, & Francis,1997; Shi, Ribbe, Cai, & Cowan, 2007), and gives rise to persistent drought conditions and stronger responses to ENSO events across much of southern Africa, (Fauchereau, Trzaska, Rouault, & Richard, 2003; Gaughan & Waylen, 2012; Mason, 2001). Impacts of drought on vegetation patterns are reported to be particularly acute in the semi-arid savannas where MAP is the dominant control on vegetation distribution (Fuller & Prince, 1996; Richard & Poccard, 1998; Townshend & Justice, 1986). Protracted desiccation, such as the 10e15% decline in MAP and increased frequency of years experiencing totals within the lowest historic tercile, witnessed after the mid 1970's (Fig. 3) could alter irreversibly the state of the savanna landscape. Across the
391
study landscape, changes of this magnitude, combined with changes in variance, have considerable effects on phenology, land use choices, and ultimately, socio-ecological functioning. As such, this research asks ‘How do climate variability and climate shifts influence vegetation patterns in dryland ecosystems?’ Study area and context The study covers 683,000 km2 of three watersheds (KavangoKwandu-Zambezi catchments), in four southern Africa countries (Zambia, Angola, Namibia and Botswana). The catchments support an unmatched diversity of large mammal species and represent
Fig. 2. Study area of the Kavango-Kwando-Zambezi catchments depicting (a) location in southern Africa covering areas of Botswana, Namibia, Zambia and Angola, and (b) portions of White's 1970's biophysical/vegetation zone map for the area (White, 1983).
392 C. Gibbes et al. / Applied Geography 53 (2014) 389e401 Fig. 3. Triangles indicate the locations of six precipitation stations and their time series of annual precipitation at the six sites divided into two periods 1950e72 and 1979e2008 separated by a six-year buffer. Historic upper and lower terciles and counts of the numbers of years in each period falling within the two extreme classes are shown.
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
major investments in the protected areas from the KavangoZambezi Transfrontier Conservation Area's component parks to community conservation areas. Historic and gridded precipitation records (1950e2009) encompassing recent climate change, and 30 years of satellite imagery, are analyzed to detect a potential driver of the persistence of vegetation productivity over this large area. The area (Fig. 2) exhibits a MAP gradient from 400 mm at the southerly extreme of the Okavango delta to 1400 mm on the CongoeZambia border and crosses a key threshold, that of 650 mm MAP, below which precipitation is thought to dominate vegetation patterns, and above which other factors such as fire and grazing play a role (Sankaran et al. 2005). The study area has a relatively low human population; with the majority of the study area having a population density of less than 10 people per km2 (Linard, Gilbert, Snow, Noor, & Tatem, 2013). Linard et al. (2013) use data with a 100 m resolution to depict the spatial distribution of the human population. At this resolution limited variation is evident across the study area. The presence of a low human population density does not suggest that human activity does not have an impact on the landscape, but rather that at the scale of study used here the low human population density reduces the noise from land use changes associated with major roads and settlements, and facilitates identification of the effects of climate on vegetation productivity. African savannas are highly heterogeneous mixed woody-herbaceous systems (Hanan & Lehmann, 2010; Scholes & Walker, 1993) whose ecosystems are closely related to non-equilibrium environmental variability (Ellis & Swift, 1988; Rodriguez-Iturbe, D'Odorico, Porporato, Ridolfi, & l, 1999; Staver et al. 2011) and occupy multiple states in which changes in climate influence savanna functioning (Fig. 1). Annual precipitation is non-stationary even in the weak statistical sense (changes in the first two moments over time), and this may induce spatially-differing impacts across the region (Fig. 3). Consequently savanna ecosystems have been discussed in the frameworks of both non-equilibrium and resilience theories. For instance, small scale studies show that burning or grazing of a well-grassed savanna can cause it to pass over an ecological threshold to a new state where woody species persist by maintaining soilewater relationships in their favor, a situation that lowers the productive capacity of the land for livestock (Dublin, Sinclair, & McGlade, 1990; Ellis & Swift, 1988). Similarly, Staver et al. (2011) found that low rainfall in Africa resulted in savanna and high rainfall in forest and that tree cover tended towards one of these two poles, with little intermediate coverage. They argued that this indicates savanna is a distinct and possibly alternate stable
393
Fig. 5. Comparison of Monte Carlo simulations of the distribution of the relative directional persistence metric, R, following 28 transitions and the anticipated results from a random walk (p ¼ 0.5) process.
state to forest cover. Our region, encompassing vegetation cover from open grassland to dense miombo woodland, presents an ideal study landscape. Materials and methods Data sets The Kavango-Kwandu-Zambezi catchments cover 683,000 km2 of four southern Africa countries (Zambia, Angola, Namibia and Botswana). The Willmotte Matsuura dataset, a global gridded monthly time series of modeled rainfall (Matsuura & Willmott, 2007) is used to describe long-term precipitation patterns. This dataset improves upon a previous global mean monthly dataset in terms of its interpolation algorithm and the use of an increased number of neighboring station points (Fekete, Vorosmarty, Roads, & Willmott, 2004). It uses a spatial interpolation of monthly total precipitation station values created a 0.5 0.5 latitude/longitude grid with grid nodes centered on 0.25 which results in a total of 232 points for the study area. Hypergeometric test The hypergeometric test compares the observed number of years displaying predefined (wet/dry) precipitation characteristics in a specified time span to a random sampling of n elements (years) among the N members of the initial population (total historic record), k of which display the desired characteristic (wettest tercile/ driest tercile) (Martineu, Caneill, & Sadourny, 1999). If precipitation is stationary from 1950 to 2009 we expect to find similar proportions of wet and dry years within each of the two time period groups (‘pre’, 1950e1972 and ‘post’, 1979eonward) relative to the sample size of each group. Across the 20 year time period from 1950 to 1970 we would expect to see approximately one third (~7) of the observations characterized as ‘wet years’ and one third (~7) characterized as ‘dry years’. Similarly across the 28 year time period from 1980 to 2008 we would expect to see approximately one third (~9) of the observations characterized as ‘wet years’ and one third Table 1 Critical values of the relative directional persistence statistic, R for the given levels of statistical significance with a 28-year time series. N (years)
Fig. 4. Hypothetical normally distributed NDVI values (lower panel) and the changing probabilities of success and failure in the following step depending upon the value of the current value of NDVI, as used in defining the relative persistence metric, R.
28
Significance 0.005
0.010
0.025
0.050
0.100
10
10
8
8
8
394
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
(~9) characterized as ‘dry years’. The hypergeometric probability distribution provides an indication of the chances that the observed numbers of events came about through random chance of the sampling of a stationary precipitation process.
395
an increase in values of NDVI over the critical observation, and 1 for decreases. Theoretically, as NDVI is a continuous variable bounded by 1 and þ1, identical values of NDVI are impossible. Relative directional persistence, R
Time-series analysis A 28-year time series of AVHRR-MODIS Normalized Difference Vegetation Index (NDVI) data is analyzed to determine the resilience of the ecological systems via a series of persistence analyses of NDVI. The AVHRR NDVI data set (1982e1999) consists of maximum composite value images with 0.1 resolution. The MODIS dataset (2000e2009), designed to complement the AVHRR data and offer continuity were originally processed by NASA Goddard and consist of cloud free composite NDVI data with a .05 resolution. . Preliminary examinations of the data indicated that April consistently represents the peak vegetation conditions. April imagery was also used as it corresponded with field observations which were used to guide the interpretation of the results. The selection of the peak vegetation month, which is representative of the cumulative effect of the wet season on the landscape, addressed the challenge of missing data values. A global shift in climate during the mid-1970s detected in many climatological and biological time series (Mantua et al., 1997; Shi et al. 2007) manifests itself regionally as a persistent drought and stronger responses to ENSO events, (Fauchereau et al. 2003; Gaughan & Waylen, 2012; Mason, 2001). To identify regional differences from this overall trend we develop three metrics of NDVI persistence by cumulating: (1) directional change in NDVI in each year compared to 1982, Directional Persistence, (2) directional change in NDVI compared to the previous year, Relative Directional Persistence and (3) absolute changes in NDVI from year to year, Massive Persistence. The first two are based on “directional” changes (increase/decrease) in NDVI. The third is simply the total amount of change summed across the time series at a pixel level (values are added for positive NDVI values which increased from time 1 to time 2, and are subtracted if time 2 is smaller than time 1). This metric indicates not only the direction of that change, but also its magnitude, a “massive” measure of persistence. The first two metrics, directional persistence and relative directional persistence, are also tested to assess the statistical significance of change in the metric values.
Relative directional persistence, R, is similar to the previous metric except directional comparisons are made to the observation in the preceding time period rather than the fixed benchmark:
Rj ¼
n X
ti;j
(2)
i¼2
Vi;j < Viþ1;j : ti;j ¼ þ1 Vi;j > Viþ1;j : ti;j ¼ 1 Massive Persistence, M accounts for the magnitude and direction, of change in NDVI annual, and is therefore fairly sensitive to abrupt “step” changes in a series, as opposed to only monotonic changes.
Mj ¼
n X Viþ1;j Vi;j
(3)
i¼2
where n is the total record length, and Vi is the value of NDVI in April of the ith year for the jth pixel. Theoretical development of a NDVI persistence analysis Throughout, the null hypothesis is defined by the assumption that, in the absence of significant perturbations to the land use land cover system, values of NDVI are normally distributed and serially independent. In theory, the ratio-nature of the NDVI variable, and the fact that the variable is itself bounded (1, þ1) contradicts the first part of this assumption. However empirical observations of NDVI values drawn from pixels across the region that retained a consistent land cover classification over time indicate that the assumption is not unreasonable. Similarly, as comparisons in this study are made between NDVI values in April in different years, serial independence is more likely.
Directional persistence, D
The directional persistence
Directional persistence, D, is a measure of the cumulative direction of change over the time series relative to the fixed benchmark observation of NDVI (April 1982), and is derived as:
The Directional Persistence metric matches closely the classic random walk process in which consecutive outcomes are viewed as a simple Bernoulli process with fixed probabilities of success (step ¼ þ1) and failure (step ¼ 1) relative to the first value of NDVI, and are independent of position in the random walk. The probability distribution of the number of successes or failures after a specified number of trials is given by Pascal's triangle. The difficulty arises in estimating the probability of a success. Under the conditions of the null hypothesis, the most likely probability associated with the first annual value is 0.5 (the expectation of a normal distribution). In the case of the Relative Directional Persistence metric, a success (failure) occurs when a value of NDVI is greater (less) than the preceding one. Thus the probabilities of a success or failure change
Dj ¼
n X
ti;j
(1)
i¼2
V1 < Vi;j : ti;j ¼ þ1 V1; > Vi;j : tij ¼ 1 where Vi,j is the April NDVI, in year i, for pixel j, e.g. V1 is NDVI value in April 1982. A value of þ1 is assigned to ti,j when the pixel records
Fig. 6. Spatially explicit representation of mean annual precipitation and standard deviation for pre- and post-1975 and the difference between the two time periods. Increases in mean annual precipitation are indicated in blue-green, while decreases are indicated in yellow-red. Increases in standard deviation, suggesting increased variability are shown in yellow-red, while decreases are indicated in blueegreen. (a) Mean annual precipitation pre-1975 (b) Standard deviation of annual precipitation pre-1975 (c) Mean annual precipitation post-1975 (d) Standard deviation of annual precipitation post-1975 (e) Difference between pre- and post-1975 mean annual precipitation (f) Difference between pre and post 1975 standard deviation.
396
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
depending upon the previous value. The greater the absolute value of the preceding NDVI the less likely the series is to continue in that direction. Fig. 4 shows hypothetical normally distributed values of NDVI in the lower panel. The upper panel displays the probability of a success (positive step) or failure (negative step), conditioned upon the previous value of NDVI (horizontal axis). The lower (higher) the preceding NDVI value the less likely that the next value will be lower (higher), thereby restricting the range of likely scores derived previously according to Pascal's triangle. Monte Carlo simulations (n ¼ 10,000) of the null condition permits the computation of a sampling distribution. We compared the frequency distributions of anticipated values of R resulting from 28 transitions (comparable to the available 28-year time series), to those expected under conditions of the standard random walk with 0.5 probability of success (Fig. 5). The constriction of the range of anticipated sums of directional changes is quite apparent. Table 1 contains critical values of R for commonly employed significance levels. As D and R only measures directional change, under the conditions of the null hypothesis, all applications, regardless of the original mean and variance of NDVI and land use/land cover type, can be reduced to a standard normal distribution and handled in this fashion. Results Changes in precipitation mean and variance before and after the climate shift The mean and variance of annual precipitation are positively correlated (Rind, Goldberg, & Reudy, 1989), and many analyses simply default to using MAP rather than investigating variability. However, our data suggests that attention should be paid to both statistics. Gridded precipitation data facilitate a broad spatial assessment of differences in MAP and its interannual variability (standard deviation) across the study region for 1950e1999 (accepting the limitations posed by extrapolation and discontinuity of records, see Gaughan & Waylen, 2012) (Fig. 6). Given the shorter temporal extent of these gridded data, 1975 was identified as the midpoint of the buffer period during which the climate shift occurred and about which precipitation patterns were assessed pre- and post-1975. Differencing MAP between the two periods suggests that precipitation increased slightly only in the extreme northeast, and declined elsewhere, particularly in the east central (Zambezi) area and in small areas in the west (central Angola) (Fig. 6). Surprisingly, many areas did not experience the same sign of changes in both mean and variability. The latter increased in portions of the study area which experienced limited changes in MAP e along the wetter (1000 mmþ) northern boundary of the study area and in the west central Batotseland area (600e800 mm). Temporal shifts in landscape productivity and spatial heterogeneity Given ENSO's noted periodicity of between 3 and 7 years and its postulated impacts on regional precipitation, 7-year running means are derived from the 1982e2009 NDVI time series to minimize potential temporary impacts of the dearth or excess of precipitation in individual years associated with its warm and cold phases (Gaughan & Waylen, 2012; Mason, 2001; Nicholson, 2000). Fig. 7 provides a graphic aspatial representation of biomass quantity and heterogeneity for the whole study region. Fig. 7(a) depicts the hypothetical relationship between mean-variance and vegetation status. Each quadrant offers a comparative measure of heterogeneity and vegetation presence. Variance is representative of the degree of landscape heterogeneity, while the mean gives an
indication of vegetation quantity. Quadrant 1 is typified by low mean and low variance, and represents degraded landscapes; with consistently less vegetation across the landscape. Quadrant 2 evinces low mean but high variance, and indicates patchy landscapes, possibly with bare ground and susceptibility to disturbance. Quadrant 3 contains landscapes with high mean and low variance, indicating that much of the landscape has vegetation cover, however this may indicate that part of landscape is susceptible to disturbance. Quadrant 4 exhibits high values of both mean and variance. Landscapes in this quadrant can be considered the most ideal, as they are well vegetated and less susceptible to disturbance. Low biomass (lower mean NDVI) and low spatial heterogeneity (lower NDVI variance) typify the study landscape in the early 1980's, a potentially lagged response to the decreased precipitation initiated in the latter half of the previous decade. The subsequent period of adaptation to the new climate regime, characterized by increases in biomass and heterogeneity, but not in precipitation (Fig. 3), extends to the present (Fig. 7). This sequence suggests that the landscape as a whole has recovered from the climate shift as is typical of a functioning savanna landscape (Hill & Hanan, 2011). Although the NDVI data series, which only begins in 1982, does not permit comparison to the pre-climate shift landscape, comparisons using data from other satellite platforms indicate that high biomass and moderate heterogeneity characterized the pre-shift condition (Cui, Gibbes, Southworth, & Waylen, 2013). These analyses suggest that this landscape is dynamic, and its levels of primary production have rebounded functionally from a significant climate shift. The increase in spatial heterogeneity, reflecting the varied responses of different dryland ecosystems (savanna, miombo, Kalahari woodland), may be an indicator of an ecosystem shift in response to the decreased precipitation (Carpenter & Brock, 2006). Spatial persistence of landscape primary productivity All three metrics (Figs. 8e10) support the notion of an overall increase in NDVI (green) at the landscape-scale, while limited subregions emerge within which NDVI has declined (purple). When examined year-by-year or in terms of absolute NDVI, the pattern is less extreme, but nonetheless reinforces the impression that, overall, this landscape has recovered (vegetation increase) since 1982 and started to stabilize (Fig. 7). Areas of persistent decline are common across all three metrics, and are clustered predominantly in the northeast and northwest of the study region in miombo woodlands (Fig. 2). Statistically significant values of the metrics (Figs. 9 and 10) highlight regions where patterns of change are most marked within the landscape. Areas of persistent decrease in NDVI-related metrics correspond to regions of highest overall NDVI and MAP. Consequently, we assess 1) whether the climate shift has affected vegetation types within the landscape differentially, and 2) the degree to which changes in the mean and variance of precipitation have impacted vegetation. Patterns in NDVI and precipitation mean and variance Declining vegetation productivity (NDVI) in the wet miombo woodlands is associated, counter-intuitively, with an increase or a small decline (relative to other changes) in precipitation. However, these regions evince marked increases in inter-annual variability (Fig. 6b, d and f). Increasing NDVI amongst the deciduous forest/ Kalahari woodland of the southeast coincide with areas of declining mean and variability of annual precipitation (Fig. 6), implying that post-shift NDVI may be better explained as a result of lower interannual variability (Fig. 6b, d and f). These metrics of change in NDVI (declining/increasing) prove statistically significant for a portion of
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
397
Fig. 7. Temporal changes in vegetation amount and heterogeneity shown (a) conceptually across the four quadrants available and (b) the observed annual data and as 7-year running means of NDVI mean and NDVI variance over the landscape. Mean and variance are proxies for biomass quantity and heterogeneity respectively.
pixels (Figs. 8 and 9) though clearly when considering change with regard to the initial state of the landscape (1982) more of the study is characterized as having changed significantly (Fig. 8b and c). The trends coupled with the identification of statistically significant declines and increases in the metrics suggest that changes in interannual variability of precipitation may be equally (thresholds running more diagonally in Fig. 1) or more important (thresholds becoming more vertical) in explaining observed patterns of spatial persistence or declines of primary production in southern African savannas than simple MAP (the equivalent of horizontally oriented thresholds). Despite questions concerning the changing numbers of stations employed in the gridding of precipitation, the high spatial association between the NDVI patterns, which are not interpolated, and the regions experiencing changes in inter-annual variability add credibility to the broadly derived spatial changes in precipitation. Field verification (June 2011) indicated that areas of NDVI decline were typified by little grass or trees despite local narratives that the areas had previously been heavily vegetated.
Discussion The IPCC (2007) anticipates that southern Africa will experience an increase in variability and a decrease in MAP. Time-series analysis of NDVI indicates that at a landscape scale, the shift in climate of the late 1970s has impacted this savanna ecosystem. The patterns of change in NDVI show initial decreases in NDVI, followed by a recovery period of increasing NDVI values (Fig. 7). Mean NDVI and spatial heterogeneity increased during the last three decades despite sustained diminished precipitation relative to pre-1975 conditions. The increasing spatial heterogeneity at this scale is a consequence of different ecosystems responding to shifts in climate in varied ways. In the southern Kalahari woodlands and dry savannas (green in Figs. 8e10), for example, recovery (increasing NDVI) is associated with decreases in both MAP and inter-annual variability. By contrast, the miombo woodlands of the northeastern and northwestern catchment (purple in Figs. 8e10) are not recovering, and small changes in MAP along with sizeable increases
398
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
Fig. 8. Results from the application of the spatially explicit persistence analysis of NDVI from a 28-year time series for (a) Directional persistence relative to 1982, (green represents increases in the metric, yellow areas of little to no change and purples decreases), (b) Regions of statistical significance at 0.10 significance level relating to scores of greater than þ/ 8 and (c) those significant at 0.005 level indicated by scores of greater than þ/ 10.
in variability as evidenced by continued decreases in NDVI. Returning to the conceptual diagram (Fig. 1), we hypothesize that dryland savannas (those portions where the relative human impact through fire, grazing and settlement is not severe) have maintained or increased their productivity despite lower MAP whereas some northern miombo woodlands may have crossed an ecological threshold towards declining NDVI associated with increased precipitation variability. Although water availability is a limiting factor for plant growth and productivity throughout terrestrial ecosystems, the degree to which vegetation is sensitive to its change varies by biome (Huxman et al. 2004). The counterintuitive response of vegetation to decreases in MAP may in part be due to changes in the frequency and magnitude of precipitation totals. For example, despite decreases in MAP, changes in the frequency of extreme totals and/or increased variability in precipitation, has been shown to modify soil moisture and increase soil water stress in biomes accustomed to moderate quantities of water (Knapp et al. 2008). Conversely, in arid regions, those systems adapted to limited total quantities and high variability, could benefit from increased precipitation
variability and a relative decrease in evaporative water loss (Knapp et al. 2008). The integral role of soil in the relationship between NDVI and precipitation has been examined in this region, and findings indicate that soil is central in mediating the impact of changes in precipitation (Farrar, Nicholson, & Lare, 1994), and that biophysical functioning of the ecosystem may result in counterintuitive responses of vegetation to precipitation decreases. Decreased MAP accompanied by increased variability, potentially modifies vegetation by altering the physical medium between precipitation and vegetation e the soil (Weltzin et al. 2003). Additionally, the efficiency with which vegetation utilizes water varies as precipitation availability, changes. Huxman et al. (2004) show that increases in precipitation result in decreased efficiency of water utilization, thus, vegetation with limited access to precipitation use water more effectively. In this study area research combining the use of remote sensing and field observations indicate changes in vegetation composition occurring within this study time period (e.g. Ringrose, Matheson, Wolski, & Hunstman-Mapila, 2008; Ringrose, Vanderpost, & Matheson, 1996). These hypothesized reasons underlying the observed changes in NDVI, are
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
399
Fig. 9. Results from the application of the spatially explicit persistence analysis of NDVI from a 28-year time series for (a) Directional change in NDVI in each year compared to the previous year (green represents increases in NDVI values, yellow identifies areas of little to no change over the time period, while purples represent decreases in NDVI values) and (b) the statistically significant regions for this change at 0.10 confidence interval relating to > þ/ ±8.
cautiously made give recent work highlighting mechanisms other than biophysical change that could contribute to increases in NDVI values. Morton et al. (2014) show that sun-sensor geometry yield changes in recorded near-infrared reflectance thereby impacting vegetation indices and patterns derived from these indices. At a regional scale this suggests that ecosystems have differing resilience characteristics, and response pathways, to climatic perturbations. Similar work relying on the use of NDVI to examine the
Fig. 10. Results from the application of the spatially explicit persistence analysis of NDVI from a 28-year time series showing Massive Persistence. (Green represents increases in NDVI values, yellow identifies areas of little to no change over the time period, while purples represents decreases.)
relationship between climate and vegetation, highlights greening trends in northern Africa since the 1980s, and like this study finds spatial variation in the greening trends (Herrmann, Anyamba, & Tucker, 2005). Time-series analyses provide a significant improvement in methods for detecting, measuring and evaluating landscape resilience. One limitation of this work is the use of 1982 as the benchmark for comparison of change. Precipitation records do not consistently identify 1982 as a particularly dry year; however, as with the use of all benchmarks the spatial analysis of change within the study area is relative to the status of the landscape in 1982. Future work could consider utilizing multiple benchmark years, however this would reduce the value of the temporal extent of the data, a characteristic which, with 27 years of standardized observations, is a distinctive and highly valuable characteristic of this data. Changes in climate are driving change in some landscapes, if not in others, at scales that human managers cannot perceive easily. At more local scales, however, land use and management decisions are also relevant, and, perhaps of even greater importance, are the changes in variability. These observations, especially the widely noted decrease in tree and herbaceous cover and increases in dense shrub cover (Ringrose et al. 2008), lend credence to concepts of ‘state-and-transition’ or ‘multiple stable state’ models associated with the resilience framework (Gunderson & Holling, 2002; Holling, 1973). At the local scale, ecosystems are moving over thresholds through drivers like grazing, fire and management. However, analyses of the mean and variance of NDVI over time and space (Figs. 8e10), suggest that threshold crossings may also be occurring differentially on much larger spatial and temporal scales than previously assessed. This is potentially linked to a shift in MAP and more importantly, interannual variability of precipitation. Within a state-and-transition context, the amount and homogeneity of vegetation is shifting between productive-patchy, and less productive-homogenous states. The nature of this shift varies across vegetation types and in response to both changes in MAP and precipitation variability. Ultimately, savannas are adapted to high inter-annual variability of precipitation; however, changes in that variability itself, combined with prolonged drought,
400
C. Gibbes et al. / Applied Geography 53 (2014) 389e401
potentially produce spatially differential responses. Despite a decline in MAP across the study area following the late 1970s, we see an overall trend of increasing biomass in drier savannas as the landscape recovers/adapts to this landscape-scale perturbation to the system. However, at the wetter margins of savannas where miombo woodlands dominate, habitats appear to be vulnerable to the increases in variability of precipitation. In both instances the response of the vegetation to the new climate states defined by mean and variability may be analogous to the effect of future climate change on this landscape (IPCC 2007). Conclusion The response of vegetation to changes in climate patterns explored over a 28-year time period at the spatio-temporal scale employed here serves as an example of the combined use of the conceptual model (Fig. 1) and the developed methodology to operationalize resilience concepts both for this particular system and more generally. The conceptual model presented is based on non-equilibrium concepts, which consider the response of a system to a perturbation. When applied to this semi-arid landscape, one in which precipitation is a major determinant of ecological composition and functioning, the perturbation to the landscape is characterized by a shift in climate and changed in climate variability, as measured by changes both MAP and in precipitation variability. The corresponding methodology developed and applied to this model examines the trends and statistical significance of changes in a quantifiable system characteristic e vegetation as measured using NDVI e with regard to major changes in precipitation patterns. Climate variability and climate shifts are explored in concert with an examination if changes in vegetation patterns. The findings indicate that changes in MA and in precipitation variability within this landscape correspond to increases in mean NDVI and increases in spatial heterogeneity at the landscape scale. The underlying framework which intertwines concepts of system shifts, the use of remotely sensed data to measure system characteristics, and the development of methods to identify statistically significant changes in the landscape is globally applicable and holds the potential to advance the ways in which system dynamics are studied. Acknowledgements This study was funded by NASA LCLUC Project # NNX09AI25G. Titled “The Role of Socioeconomic Institutions in Mitigating Impacts of Climate Variability and Climate Change in Southern Africa” PI: Dr. Jane Southworth. The authors thank the reviewers and editor for thoughtful and constructive guidance which improved the quality of this article. References Carpenter, S. R., & Brock, W. A. (2006). Rising variance: a leading indicator of ecological transition. Ecology Letters, 9, 311e318. Chavez, F., Ryan, J., Lluch-Cota, S., & Niquen, M. (2003). From anchovies to sardines and back, multidecadal change in the Pacific Ocean. Science, 299, 217e221. Cui, X., Gibbes, G., Southworth, J., & Waylen, P. (2013). Using remote sensing to quantify vegetation change and ecological resilience in a semi-arid system. Land, 2(2), 108e130. Dublin, H. T., Sinclair, A., & McGlade, J. (1990). Elephants and fire as causes of multiple stable states in the Serengeti-Mara woodlands. Journal of Animal Ecology, 59, 1147e1164. Ebbesmeyer, C., Coomes, C., Cannon, G., & Bretschneider, D. (1991). Linkage of ocean and fjord dynamics at decadal period. In D. H. Peterson (Ed.), Geophysical Monographs of the American Geophysical Union: 55. Climate variability on the Eastern Pacific and Western North America (pp. 399e417). Ellis, J. E., & Swift, D. M. (1988). Stability of African pastoral ecosystems: alternate paradigms and implications for development. Journal of Range Management, 41, 450e459.
Farrar, T. J., Nicholson, S. E., & Lare, A. R. (1994). The influence of soil type on the relationships between NDVI, rainfall, and soil moisture in semiarid Botswana. II. NDVI response to soil oisture. Remote Sensing of Environment, 50, 121e133. Fauchereau, N., Trzaska, S., Rouault, M., & Richard, Y. (2003). Precipitation variability and changes in southern Africa during the 20th century in the global warming context. Climatic Change, 29, 139e154. Fekete, B. M., Vorosmarty, C. J., Roads, J. O., & Willmott, C. J. (2004). Uncertainties inprecipitation and their impacts on runoff estimates. Journal of Climate, 17(2), 294e304. Fuller, D., & Prince, S. (1996). Precipitation and foliar dynamics in tropical southern Africa: potential impacts of global climatic change on savanna vegetation. Climatic Change, 33, 69e96. Gaughan, A. E., & Waylen, P. R. (2012). Spatial and temporal precipitation variability in the Okavango-Kwando-Zambezi catchment, South Africa. Journal of Arid Environments, 82, 19e30. Graham, N. E. (1994). Decadal-scale climate variability in the 1970s and 1980s: observations and model results. Climate Dynamics, 10, 135e159. Gunderson, L. H., & Holling, C. S. (2002). Panarchy. understanding transformations in human and natural systems. Washington, D.C: Island Press. Hanan, N., & Lehmann, C. (2010). Tree-grass interactions in savannas: paradigms, contradictions, and conceptual models. In M. Hill, & N. Hanan (Eds.), Ecosystem function in savannas. Boca Raton, FL: Taylor and Francis Group. Herrmann, S. M., Anyamba, A., & Tucker, C. J. (2005). Recent trends in vegetation dynamics in the African Sahel and their relationship to climate. Global Environmental Change, 15, 394e404. Hill, M. J., & Hanan, P. H. (2011). Ecosystem function in savannas: Measurement and modeling at landscape to global scales. Boca Raton, FL, USA: CRC Press. Hirota, W., Holmgren, M., Van Nes, E. H., & Scheffer, M. (2011). Global resilience of tropical forest and savanna to critical transitions. Science, 334, 232e235. Holling, C. S. (1973). Resilience and stability of ecological systems. Annual Review of Ecology and Systematics, 4, 1e23. Huxman, T. E., Smith, M. D., Fay, P. A., Knapp, A. K., Shaw, R., Loik, M. E., et al. (2004). Convergence across biomes to a common rain-use efficiency. Nature, 429, 651e654. Intergovernmental Panel on Climate Change [IPCC]. (2007). Climate change 2007: The physical science basis. Summary for policymakers. Contribution of working group I to the fourth assessment report. The Intergovernmental Panel on Climate Change. Knapp, A. K., Beier, C., Briske, D. D., Classen, A. T., Luo, Y., Reichstein, M., et al. (2008). Consequences of more extreme precipitation regimes for terrestrial ecosystems. Bioscience, 58(9), 811e821. Lanfredi, M., Simoniello, T., & Macchiato, M. (2004). Temporal persistence in vegetation cover changes observed from satellite: development of an estimation procedure in the test site of the Mediterranean Italy. Remote Sensing of Environment, 93, 565e576. Linard, C., Gilbert, M., Snow, R. W., Noor, A. M., & Tatem, A. J. (2013). Population distribution, settlement patterns and accessibility across Africa in 2010. PLoS One, 7(2), e31743. Lunetta, R. S., Knight, J. F., Ediriwickrema, J., Lyon, J. G., & Worthy, L. D. (2006). Landcover change detection using multi temporal MODIS NDVI data. Remote Sensing of the Environment, 105, 142e154. Mantua, N., Hare, S., Zhang, Y., Wallace, J., & Francis, R. (1997). A Pacific interdecadal climate oscillation with impacts on salmon production. Bulletin of the American Meteorological Society, 78, 1069e1079. Martineu, C., Caneill, J., & Sadourny, R. (1999). Potential predictability of European winters from the analysis of seasonal simulations with an AGCM. Journal of Climate, 12, 3033e3061. ~ o, climate change, and southern African climate. EnviMason, S. J. (2001). El Nin ronmetrics, 12(4), 327e345. Matsuura, K., & Willmott, C. J. (2007). Terrestrial precipitation: 1900e2006 gridded monthly time series. Retrieved February 18, 2009, from http://climate.geog.udel. edu/wclimate/html_pages/Global_ts_2007/README.global.p_ts_2007.html. Meyer, K. M., Wiegand, K., Ward, D., & Moustakas, A. (2007). The rhythm of savanna patch dynamics. Journal of Ecology, 95, 1306e1315. Morton, D. C., Nagol, J., Carabajal, C. C., Rosette, J., Palace, M., Cook, B. D., et al. (2014). Amazon forests maintain consistent canopy structure and greenness during the dry season. Nature, 506, 221e224. Namias, J. (1978). Multiple causes of the North American abnormal winter 1976e77. Monthly Weather Review, 106, 279e295. Nicholson, S. E. (2000). The nature of precipitation variability over Africa on time scales of decades to millenia. Global Planet Change Letters, 26, 137e158. ~ ones, M. L., Ve lez, I. D., Mantilla, R. I., Ruiz, D., et al. Poveda, G., Rojas, W., Quin (2001). Coupling between annual and ENSO timescales in the malaria-Climate association in Colombia. Environmental Health Perspectives, 109(5), 489. Richard, Y., & Poccard, I. (1998). A statistical study of NDVI sensitivity to seasonal and interannual precipitation variations in southern Africa. International Journal of Remote Sensing, 19, 2907e2920. Rind, D., Goldberg, R., & Reudy, R. (1989). Change in climate variability in the 21st century. Climatic Change, 14, 5e37. Ringrose, S., Matheson, W., Wolski, P., & Hunstman-Mapila, P. (2008). Vegetation trends along the Botswana Kalahari transect. Journal of Arid Environments, 54, 297e317. Ringrose, S., Vanderpost, C., & Matheson, W. (1996). The use of integrated remotely sensed and GIS data to determine causes of vegetation cover change in southern Botswana. Applied Geiography, 16, 225e242.
C. Gibbes et al. / Applied Geography 53 (2014) 389e401 Rodriguez-Iturbe, I., D'Odorico, P., Porporato, A., & Ridolfi, L. (1999). On the spatial and temporal links between vegetation, climate and soil moisture. Water Resources Research, 35, 3709e3722. Sankaran, M., Hanan, N. P., Scholes, R. J., Ratnam, J., Augustine, D. J., Cade, B. S., et al. (2005). Determinants of woody cover in African savannas. Nature, 438(7069), 846e849. Scholes, R. J., & Walker, B. H. (1993). An African savanna: Synthesis of the Nylsvley study. Cambridge, UK: Cambridge University Press. Shi, G., Ribbe, J., Cai, W., & Cowan, T. (2007). Multidecadal variability in the transmission of ENSO signals to the Indian Ocean. Geophysical Research Letters, 34, L09706. http://dx.doi.org/10.1029/2007GL29528. Staver, C. A., Archibald, S., & Levin, S. A. (2011). The global extent and determinants of savanna and forest as alternative biome states. Science, 334, 230e232. Townshend, J., & Justice, C. O. (1986). Analysis of the dynamics of African vegetation using the normalized difference vegetation index. International Journal of Remote Sensing, 7, 1435e1445. Trenberth, K. E. (1990). Recent observed interdecadal climate changes in the Northern Hemisphere. Bulletin of the American Meteorological Society, 71, 988e993. Trenberth, K. E., & Hurrell, J. (1994). Decadal atmosphere-ocean variations in the Pacific. Climate Dynamics, 9, 303e319. Vicente-Serrano, S. M., Beguería, S., Gimeno, L., Eklundh, L., Giuliani, G., Weston, D., et al. (2012). Challenges for drought mitigation in Africa: the potential use of geospatial data and drought information systems. Applied Geography, 34, 471e486. Washington-Allen, R. A., Ramsey, R., & West, N. E. (2004). Spatiotemporal mapping of the dry season vegetation response of sagebrush steppe. Community Ecology, 5, 69e79.
401
Washington-Allen, R. A., Ramsey, R. D., West, N., & Norton, B. (2008). Quantification of the ecological resilience of drylands using digital remote sensing. Ecology and Society, 13, 33. Washington-Allen, R. A., Van Niel, T. G., Ramsey, R. D., & West, N. E. (2004). Remote sensing based piosphere analysis. GIScience and Remote Sensing, 41, 136e154. Washington-Allen, R. A., West, N. E., Ramsey, R., & Efroymson, R. A. (2006). A protocol for retrospective remote sensing-based ecological monitoring of rangelands. Rangeland Ecology and Management, 59, 19e29. Weltzin, J. F., Loik, M. E., Schwinning, S., Williams, D. G., Fay, P. A., Haddad, B. M., et al. (2003). Assessing the response of terrestrial ecosystems to potential changes in precipitation. Bioscience, 53(10), 941e952. Westman, W. E., & O'Leary, J. (1986). Measures of resilience: the response of coastal sage scrub to fire. Vegetatio, 65, 179e189. White, F. (1983). Vegetation of Africa e A descriptive memoir to accompany the Unesco/AETFAT/UNSO vegetation map of Africa. Natural Resources Research Report XX. 7 Place de Fontenoy, 75700 Paris, France: U. N. Educational, Scientific and Cultural Organization, 356 pages. Yang, L., Wylie, B. K., Tieszen, L. L., & Reed, B. C. (1998). An analysis of relationships among climate forcing and time-integrated NDVI of grasslands over the US northern and central Great Plains. Remote Sensing of Environment, 65, 25e37. Zen, N., & Neelin, D. (2000). The role of vegetation-climate interaction and interannual variability in shaping the African savanna. Journal of Climate, 13, 2665e2685.