Vegetation structure in water-limited systems is to a large degree controlled
by ecohydrological processes, including mean annual precipitation (MAP)
modulated by the characteristics of precipitation and geomorphology that
collectively determine how rainfall is distributed vertically into soils or
horizontally in the landscape. We anticipate that woody canopy cover, crown
density, crown size, and the level of spatial aggregation among woody plants
in the landscape will vary across environmental gradients. A high level of
woody plant aggregation is most distinct in periodic vegetation patterns
(PVPs), which emerge as a result of ecohydrological processes such as runoff
generation and increased infiltration close to plants. Similar, albeit
weaker, forces may influence the spatial distribution of woody plants
elsewhere in savannas. Exploring these trends can extend our knowledge of how
semi-arid vegetation structure is constrained by rainfall regime, soil type,
topography, and disturbance processes such as fire. Using
high-spatial-resolution imagery, a flexible classification framework, and a
crown delineation method, we extracted woody vegetation properties from
876 sites spread over African savannas. At each site, we estimated woody
cover, mean crown size, crown density, and the degree of aggregation among
woody plants. This enabled us to elucidate the effects of rainfall regimes
(MAP and seasonality), soil texture, slope, and fire frequency on woody
vegetation properties. We found that previously documented increases in woody
cover with rainfall is more consistently a result of increasing crown size
than increasing density of woody plants. Along a gradient of mean annual
precipitation from the driest (< 200 mm yr
African savannas are complex tree–grass systems controlled by combinations of climate, soil, and disturbance processes such as fire and herbivory (Sankaran et al., 2008). In dry savannas, water availability determines the establishment, growth and survival of plants and competitive plant traits are often of a water-saving nature (Chesson et al., 2004; Pillay and Ward, 2014). Abiotic environmental factors, such as the rainfall regime, soil type, and topography, impact ecohydrological processes by controlling infiltration rates, runoff generation, and available water capacity, which in turn impact the growth and survival of woody plants in the landscape (Ludwig et al., 2005). Climate, both rainfall patterns and temperatures, could change in many parts of Africa (Gan et al., 2016), and the effect on vegetation will depend on how those pressures interact with other abiotic and biotic factors. In addition to ecohydrological factors, savannas are heavily influenced by the frequency and intensity of fires (Bond, 2008), as well as herbivore regimes (Sankaran et al., 2008), which often combine to suppress woody cover to levels well below its climatic potential (Sankaran et al., 2005). A thorough understanding of the underlying processes that influence savanna vegetation structure is key to assessing the future resilience and productivity of these ecosystems.
Across environmental gradients we expect to see variation in woody vegetation properties, including individual-level characteristics (mean crown size) and population-level characteristics (crown density, woody cover, and the spatial distribution of plants in the landscape). Woody cover is fundamentally a function of crown sizes and crown density and by studying these components individually, it is possible to attain important insight into the function of ecosystems and what ecosystem services they provide. Two landscapes with similar woody cover but different sizes of individual trees will sequester different amounts of carbon (Shackleton and Scholes, 2011), harbor different fauna (Riginos and Grace, 2008), and differ in biogeochemical dynamics (Veldhuis et al., 2016a). The level of spatial aggregation among woody plants can help us understand facilitative and competitive processes determining survival of seedlings and saplings. Woody plants increase water infiltration and local accumulation of soil and nutrient resources, as well as altering sub-canopy microclimates (Barbier et al., 2014; Dohn et al., 2017; Gómez-Aparicio et al., 2008). These short-range facilitative effects usually operate at spatial scales of a few meters but may increase the degree of aggregation among woody plants at larger scales (Scanlon et al., 2007; Xu et al., 2015). Overland flows of water can be especially effective at redistributing resources over longer distances, in some conditions leading to the emergence of periodic vegetation patterns (PVPs; Rietkerk and van de Koppel, 2008; Valentin et al., 1999). Contrasting infiltration rates between bare and vegetated patches lead to redistribution of water and soil resources, which reinforces an organized pattern. While soil texture type has been weakly associated with the occurrence of PVPs (Deblauwe et al., 2008), the impervious conditions of the bare patches are generally caused by shallow soil depths, hardpans, or soil crusts (McDonald et al., 2009). On flat ground, PVPs take the form of spotted, labyrinthine, or gapped patterns depending on soil water availability. On a gentle slope, they develop into vegetated bands that run parallel to contour lines (Valentin et al., 1999). While PVPs have been studied extensively, their formative processes are seldom linked to ecohydrological processes in other types of savanna landscapes.
Methodological workflow showing datasets (rounded boxes) and methods (square boxes) used to estimate woody vegetation structure and analyze relationships with environmental variables.
To analyze how woody cover, crown size, crown density and the spatial
pattern of trees vary with environmental gradients, we need to map the
landscape at the level of individual trees. Satellite-based high-spatial-resolution (HSR; < 4 m) sensors have the necessary degree of detail
for this task. Papers delineating individual trees from HSR in African
savannas have shown promising results (Karlson et al., 2014;
Rasmussen et al., 2011), but these studies are generally restricted to small
geographical areas. In this paper we present an analysis of woody properties
sampled across the diverse water-limited savannas of Africa using a
combination of WorldView, Quickbird, and GeoEye satellite data (
Location of the 48 study areas, containing 876 study sites, spread
out over African rangelands. The rangeland areas are from the anthropogenic
biomes product (Ellis and Ramankutty, 2008), and symbol size for study areas
is proportional to the number of study sites in each. The map to the right
shows a study area on the border between Somalia and Ethiopia and exemplifies
the sampling strategy for study sites (white rings). The placement of sites
was guided by a 0.04
Our methodological approach included a flexible classification approach based on unsupervised classification, tree crown delineation, and boosted regression tree analysis (Fig. 1).
We used data from WorldView-2, WorldView-3, GeoEye-1, and Quickbird-2
satellites, with varying ground resolutions (
Once the locations of sites were established, each site was preprocessed
using IDL scripts in ENVI 5.2. This included Gram–Schmidt pan sharpening of
the blue, green, red, and infrared bands, and orthorectification using
embedded rapid positioning coordinate (RPC) information and an SRTM v2 DEM (Farr et al., 2007). The orthorectified images
were resampled using a nearest-neighbor method to a standard 0.6 m ground
resolution creating a 400
After the 240
Crown delineation steps for a woodland site in Zambia.
False-color imagery of periodic vegetation patterns identified among
the sites:
The delineated crowns played an important role in this analysis because they
were used for calculating crown density, crown sizes, and woody plant
aggregation. Our analysis of the derived woody properties did not focus on
absolute numbers but on how they vary across environmental gradients under
the assumption that errors were propagated consistently over space. A visual
inspection of all sites indicated that the crown delineation consistently
produced crown layers that looked realistic when overlaying the imagery. We
recognize, however, that it is extremely difficult to accurately delineate
tree canopies in areas where crowns overlap. In some cases, a large tree
crown may be falsely divided into small canopies or a cluster of shrubs may
be grouped together into one crown (Rasmussen et al., 2011). It is important
that the rate of falsely divided and falsely grouped crowns is balanced since
excessive division of large trees into smaller leads to higher estimates of
both crown density and aggregation. We evaluated the performance of the
classification and delineation methodology using field data from sites in
Kenya (Appendix A). This showed that crowns smaller than
The rainfall data were extracted from the Tropical Rainfall Measuring
Mission (TRMM) 3B42 v7 product (0.25
We derived four statistical properties of woody vegetation from each image:
mean crown size (m
Frequency distributions of mean crown size, crown density, woody cover and aggregation calculated for different MAP ranges.
The dataset includes several sites with PVPs, which often are treated as a special case because of their striking appearance (Fig. 4). It is of interest to examine the environmental conditions associated with the occurrence of PVPs as well as those associated with aggregated woody populations in savannas without PVPs. We therefore separated sites with periodic vegetation from the rest and generated an additional set of models. The category with periodic vegetation contained 149 sites situated in Somalia, Senegal, Chad, Mali, Niger, Namibia, and Sudan. The identification was based on visual inspection, and all sites with traits of periodic patterning (spotted, labyrinthine, gapped, or banded) were put in the PVP category. We created one model for predicting aggregation among all sites, one for predicting aggregation among sites with no PVPs, and a third for predicting the occurrence PVPs. In the latter model, all PVP sites were given the value 1 and the rest 0, and the BRT family parameter was set to “bernoulli”, appropriate for binomially distributed data.
Boxplots of estimates of crown size, crown density, and woody cover
along a rainfall gradient. Red points denote the means. Between the driest
(< 200 mm yr
Modeled BRT responses (“partial dependencies”) of woody canopy
properties to each environmental variable when accounting for the average
effect of the other four variables. The red lines are smoothed
representations of the responses, with fitted values (model predictions based
on the original data) for each of the 876 sites shown as grey dots. The
We started by calculating frequency distributions of the four woody
properties divided into three rainfall categories (Fig. 5). The more arid
savannas (< 400 mm yr
Modeled BRT responses (“partial dependencies”) for predictions of
under what conditions PVPs occur (top), and woody aggregation (
Boxplots with woody properties divided into MAP bins (Fig. 6) show that
woody cover and crown sizes increased more sharply with increasing rainfall
than crown densities. Along the rainfall gradient from the driest (< 200 mm yr
Relative influence of each environmental variable and the
cross-validated
Our estimates of aggregation were based on the
Additional insight was drawn from analyzing aggregation along distances,
with the data categorized into PVPs and subdivisions based on MAP and soil
texture (Fig. 9). All categories were dispersed at short distances because
each crown takes up space and there is bound to be a short distance between
the center points of adjacent plants. Sites with PVPs had the highest levels
of aggregation reaching a maximum at around 25 m (Fig. 9). The combination
with wetter climes (
Relative influence of each environmental variable and the
cross-validated
Level of aggregation among tree crowns calculated using Ripley's
Numerous authors have investigated how woody canopy cover varies across African savannas in response to variation in environmental variables (Good and Caylor, 2011; Sankaran et al., 2005; Staver et al., 2011). Given that tropical savannas cover about an eighth of Earth's land surface (Scholes and Archer, 1997) and contributes heavily to the global carbon cycle (Poulter et al., 2014), it is important to understand the makeup of these variations in terms of crown sizes and tree densities. By separating woody cover into mean crown size and density, we were able to analyze whether they respond differently to environmental factors and how they combine to drive landscape-scale woody cover across the continent. Our results suggest that crown sizes respond more strongly to rainfall than crown density (Fig. 6). This indicates that the commonly observed relationship of increasing woody cover with MAP in African savannas (e.g., Sankaran et al., 2005) is more a result of increasing size of trees than increasing tree density, at least in savannas with MAP < 700 mm. We also found a unimodal relationship between crown sizes and soil texture that was not present in the results for crown density (Fig. 7). Soil properties have a considerable effect on the water cycle and a few studies have noticed that woody growth is suppressed on clayey soils in drylands (Lane et al., 1998; Sankaran et al., 2005; Williams et al., 1996). Recently, Fensham et al. (2015) showed that the effect is likely due to the higher wilting point on clays, which limits the soil moisture available for plants to extract. A combination of low rainfall and fine-textured soils can lead to very low soil water potentials and impact the vegetation in a way reminiscent of even drier conditions. In our results, the relationship appears unimodal with suppression on both the clayey and the sandiest end. Woody growth is then controlled by available soil moisture, which can be limited by either a high wilting point on clayey soils or low field capacity on sandy soils. Our results suggest these constraints affect the size of woody plants and not their abundance. Crown density was most strongly influenced by rainfall seasonality and appeared to have a unimodal response function (Fig. 7). The sites with very low rainfall seasonality (< 0.8) were all situated in the western part of East Africa (Serengeti, Masai Mara, and northern Uganda) in a region with bimodal rainfall distributions and far lower seasonality that further east. Many of these sites had low woody densities and cover but likely for other reasons than rainfall seasonality. Elephant densities are thought to be a key driver of woody cover in the Mara-Serengeti ecosystem (Morrison et al., 2016). Browsing, especially by elephants, has a great impact on woody structure (Sankaran et al., 2013) and is a key factor we did not capture in this analysis. If we focus on sites with rainfall seasonality above 0.8, there is a more linear relationship with lower crown density and cover in areas with high rainfall seasonality which could be associated with the long periods of high water stress in more seasonal systems. Lehmann et al. (2014) found that high rainfall seasonality can constrain canopy closure and is an important predictor for the presence of savanna. Overall, the estimated woody properties were more strongly influenced by rainfall amounts and seasonality than by soil, slope, and fire. Fire frequency had a weak negative association with both woody cover, crown sizes, and densities. Fire has, however, an interactive relationship with vegetation structure (Archibald et al., 2009) and this analysis cannot separate the effect of fire on vegetation from impacts of vegetation structure on the fire regime.
In accordance with previous research (Deblauwe et al., 2008), we found that the formation of highly aggregated PVPs is associated with specific environmental conditions. Periodic patterns are most likely to occur in areas with high rainfall seasonality low mean annual rainfall, fine-textured soils, and flat or gently sloping terrain (Fig. 8). These are factors that influence ecohydrological processes such as the propensity to form overland flows during rainfall events (Valentin et al., 1999). The results are in agreement with a global study on the biogeography of PVPs by Deblauwe et al. (2008), who found similar effects in regions with strong seasonal variation in temperature and more constant rainfall (Australia and Mexico) and in regions with distinct rainfall seasonality but more constant temperatures (Africa). Our analysis further shows that the same factors that contribute to PVP emergence are associated with higher levels of aggregation among woody plants elsewhere in African savannas. PVPs thus appear under conditions that naturally favor local facilitation and patchiness. However, the vegetation at many sites with these conditions does not exhibit highly organized periodic patterns, which could be related to soil properties other than texture. The dominant process in the formation of PVPs is a significant overland flow from bare to vegetated patches which requires near impervious soils. This property is typically associated with shallow soil depths, physical crusts, or hardpans (Leprun, 1999; McDonald et al., 2009), and is not strongly dependent on soil texture.
Previous studies have linked local aggregation and patchiness in savannas
to fire frequency (Veldhuis et al., 2016b), seed dispersal (Pueyo et al.,
2008), runoff–erosion processes (Ludwig et al., 2005), and short-range
facilitation through modified microclimate close to nurse plants (Holmgren
and Scheffer, 2010). With increasing abiotic stress, we expect stronger
tree–tree facilitation in accordance with the stress gradient hypothesis (He
et al., 2013). In our analysis, the most influential predictor for modeling
aggregation was rainfall seasonality (Table 2), a factor that could influence
plant dynamics in more than one way. The pronounced dry season associated
with highly seasonal systems exerts a strong abiotic pressure, especially on
juvenile trees with less developed root systems. Juvenile survival through
the dry season is likely higher in the shelter of nearby trees. Over time, a
bias in survival rates may lead to higher aggregation among adult trees. Once
the wet season arrives, it often comes in heavy downpours which can quickly
saturate the top soil leading to overland flows. This leads to both
redistribution of water resources to woody patches with higher infiltration
rates, as well as redistribution of litter and soil resources (Ludwig et al., 2005).
The more concentrated rains may also alleviate competition for water during
the growing season leading to facilitation being the dominant force in highly
seasonal drylands. There was also a clear relationship between fine-textured
soils and higher aggregation. Fine-textured soils increase runoff through
lower infiltration rates, and may also amplify stress during the dry season
through their higher wilting point. Sites with the combination of
coarse-textured soils (
Using high-spatial-resolution imagery, a flexible classification framework, and a crown delineation methodology, we estimated several key woody vegetation properties in African savannas and analyzed how these vary with local environmental conditions. We find that woody cover, crown sizes, and woody plant densities are more strongly influenced by rainfall amounts and seasonality than by soil texture, slope, and fire frequency. Of specific interest is that mean crown size responded more strongly to mean annual rainfall than plant densities, indicating that the commonly observed relationships between woody cover and rainfall (e.g., Sankaran et al., 2005) is more a result of increasing crown sizes than variation in crown density. Larger crown sizes were associated with mid-textured soils and appeared suppressed on both clays and very sandy soils. The level of aggregation among woody plants was most strongly related to rainfall seasonality, as was the occurrence of PVPs. Similar processes that influence patchiness in savannas also contribute to the formation of PVPs, with impermeable soil conditions being the possible difference between a patchy savanna landscape and highly organized periodic vegetation.
The estimates of woody vegetation properties are available in the Supplement of this paper.
This appendix describes a validation analysis of estimated mean crown size,
crown density, and woody cover using field data collected in southern Kenya
during September–October 2015. Plots were established in five protected
areas: Tsavo West NP, Tsavo East NP, Amboseli NP, Ol Pejeta wildlife
conservancy, and Il Ngwesi group ranch (Fig. A1). In total, we established
28 plots with at least four plots in each protected area. The size of plots
varied with the density of trees and shrubs, ranging from
350 to 8000 m
Our analysis of detection ratios (Fig. A2) indicated a detection threshold
of
When calculating the relationship between estimated and field-measured woody
properties (Fig. A3), we excluded all field-measured trees and shrubs with
a diameter less than the 2 m detection threshold. Estimates of the woody
properties then fall relatively close to the one-to-one line. The four sites
in Amboseli were dominated by large umbrella thorn acacias (
Map of the five protected areas in southern Kenya where field work was conducted. The positions of individual plots are marked with blue triangles.
Detection ratios of woody plants in classified imagery at field work sites. The values were calculated as mean detection ratios for trees divided into bins of width 40 cm.
Validation of estimated mean crown size, crown density, and woody
cover. The Amboseli sites and one site in Ol Pejeta were excluded when
calculating
The authors declare that they have no conflict of interest.
The satellite data were provided through a NASA agreement and under NextView license. We thank Jamie Nickeson, Njoki Kahiu, Sujan Parajuli, and Dinesh Shrestha for assistance with data retrieval, field work, and image classification. We also thank the Kenya Wildlife Service (KWS), Ol Pejeta Conservancy, and Ilngwesi group ranch for access to field sites in Kenya. The project was funded by the National Science Foundation (Coupled Natural-Human Systems Program) and the NASA Terrestrial Ecology Program. C. R. Axelsson was also supported by a graduate research fellowship through the Geospatial Sciences Center of Excellence at South Dakota State University. Edited by: Kirsten Thonicke Reviewed by: Penny Mograbi and Sytze de Bruin