**3. Results**

#### *3.1. Woody Vegetation Surveys*

We identified 44 woody plant species across the 78 transects where taxonomic information was recorded. We recorded the highest diversity at Kuke (site 7), with 21 species found in all six transects in the region (Figure 2a), and 13 species found along transect four during the wet season. In general, both richness and abundance decreased as data collection moved southwards which follows the precipitation gradient, although Kuke is the notable exception. We recorded the lowest total abundances for all six transects at NG5 (Site 6) and Bokspits (Site 15) and the highest abundances at Quangwa (Site 4) and Kuke (Site 7) (Figure 2b). Supplementary Information provides the data which includes geographic location (WGS 1984), a list of species recorded, and their morphological classification. We recorded eight species in Morphological Group I (bipinnate leaf structure), fourteen species in Morphological Group II (tall dense canopies), fifteen species in Morphological Group III (small dense canopy species), five species in Morphological Group IV (tall open canopies), and two species in Morphological Group V (small open canopies). Due to the low number of species recorded in Morphological Group V, these were withheld from the statistical analysis to prevent any generalization or over-fitting of the models.

**Figure 2.** Total (**a**) woody vegetation species richness and (**b**) woody vegetation species abundance summed across the six transects at each of the 15 zones across the Kalahari Transect, plotted against annual precipitation (mm). Sites listed (1) Shakawe, (2) Tsodilo, (3) Gumare, (4) Quangwa, (5) Drotsky's Caves, (6) Ng 5, (7) Kuke, (8) Ghanzi, (9) Ghanzi South, (10) Bere, (11) Tshane, (12) Tshane South, (13) Mabuasehube, (14) Tsabong, and (15) Bokspits.

#### *3.2. Regression Analysis*

Our results indicated a number of important drivers of woody vegetation species richness and abundance (Table 2). Precipitation was the most important environmental variable when we considered all species together for both richness and abundance with borehole density and fire included in the final models. When species were deconstructed into morphological groups, we observed a variety of significant environmental drivers and interactions between these variables (\*\* significant at α < 0.01, \* significant at α < 0.05). The relative importance of each environmental driver often changed when we compared the regression models for richness and abundance of the same morphological group, indicating that the processes that determine diversity are different from those determining abundance. Boreholes were the most important driver for morphological groups II abundance and III richness, while livestock was the most important driver for morphological group IV abundance.

**Table 2.** Regression output for species richness (SR) and abundance (AB) for the four Morphological Groups (MG). \*\* significant at α < 0.01, \* significant at α < 0.05. Tobit (T) and Poisson (P) regression analysis undertaken based on distribution of data.


Precipitation and borehole density were included in all final models for every morphological group, while livestock was not important when all species were considered together, but included for all morphological groups (both richness and abundance) with the exception of morphological group I richness. Similarly, fire frequency was included for most morphological groups, with the exception of morphological group III and total species richness. Several two-way interactions were returned across the different models, and these were often significant. Morphological groups I and III abundance had the most interactions among all variables, suggesting these species have a complex and dynamic relationship with the environment.

Precipitation was a significant variable in all final regression models for species richness and abundance for all but three morphological groups, and it was the most important variable for total species richness and abundance, and morphological group I richness (Table 2). Precipitation had a positive relationship with species richness for morphological groups II and III, and abundance for morphological group II. This relationship was expected since these groups are characterized by dense canopy broad leaf species resulting in higher Leaf Area Index (LAI) and hence higher water requirements [68]. A negative relationship for morphological group I (bipinnate species) richness and rainfall was identified. This could be due to the fact that in xeric environments such species outcompete the majority of broad-leaved vegetation due to their general morphological characteristics and ecological traits (such as long root traps [69]), meaning the diversity of these species increases in arid areas where other water-dependent species simply cannot survive.

Livestock density was not included in the final model as selected by AIC when all species were considered together, but it had a positive relationship with morphological groups II and III (Table 2). Small dense canopy species such as *Grewia spp, Rhus tenuivirus* and *Ziziphus mucronata* notably have relatively low palatability [70]. Thus, if these species were already established when grazing increased in the area, they would not be a ffected by livestock. It was also the most important variable in determining abundance of morphological group IV (tall open canopy), reporting a negative relationship. Borehole density also had the most influence in determining both abundance of morphological group II and richness of morphological group III, forming a negative relationship with both ecological indicators. Boreholes also had a negative relationship with all response variables with the exception of morphological group I abundance. However, the interaction between boreholes and livestock density was significant for morphological group I abundance, indicating a negative relationship. This interaction was also significant for morphological groups II and III (although positive). These findings contradict previous research, and indicate that broad leaf species thrive in locations where there are more cattle and boreholes, while bipinnate species decrease. Fire had a negative influence on both richness and abundance at a regional scale (Table 2). Fire was generally negatively correlated to the overall abundance of woody species, but had a positive relationship with abundance of morphological groups III and IV, albeit not significant.

## **4. Discussion**

Following the global trend in the conversion of savanna landscapes to woodier landscapes [7,27], the aim of this research was to investigate the variables responsible for woody vegetation composition in the western Kalahari, in particular those that cause high diversity and abundance of these species. We identified a variety of environmental drivers that are responsible for high diversity and abundance of woody vegetation, most notably precipitation, borehole density, grazing, and fire.

Our results generally agree with the observation that the rainfall gradient of the Kalahari is associated with an increase in woody vegetation [16–20]. Interestingly, the highest species richness was recorded at Kuke (Figure 1—Site 7), where the annual precipitation is 450 mm (in the middle of the rainfall gradient). The substantially higher species richness at Kuke can be explained by the site being located in an area bu ffering the Ghanzi farm-block to the south and the wildlife areas to the north. Both livestock and wildlife numbers are low here, and furthermore, fires have not occurred in this area due to both fire prevention strategies and the existence of the veterinary cordon fences acting as fire breaks. Therefore, our results indicate that while rainfall has a strong influence on woody vegetation, other factors also contribute significantly.

Our findings corroborate the positive association of bipinnate abundance (morphological group I) in areas close to boreholes [38], as well as an overall reduction in woody vegetation cover [71]. The negative relationship with small dense species is intuitive, as trampling loosens the soil and prevents these species from rooting. However, when grazing is high, the significant negative interaction between borehole density and grazing with bipinnate abundance contradicts the existing theories behind woody vegetation patterns. This relationship is a result of the fact that a higher number of boreholes and cattle represent more managed commercial ranches where cattle are routinely rotated between fields, and the regular use of multiple boreholes by the livestock negates the impact of trampling on the soil. This subsequently reduces the rate of bush encroachment by the unpalatable and thorny bipinnate species, and a positive relationship with other morphological groups is observed.

The negative relationship between fire frequency and woody vegetation corroborates observations from other dryland ecosystems [9,41] and supports a mechanistic understanding of the e ffect of fires in mixed tree-grass plant communities [40,72–74]. These findings support the observations at Kuke, that absence of fire does increase vegetation diversity and abundance (particularly for smaller species), and that the removal of fire from a landscape could increase bush thickening [49]. However, when fire was included in the models, it was seldom the most important variable (Table 2), with the exception of a positive interaction between livestock and fire when modelling morphological group I abundance

(albeit not significant). While diversity and abundance did decrease, the lesser impact compared to the other environmental variables suggests that frequent fires may not have such severe implications on the ecosystem's biodiversity as proposed [45]. However, the MODIS MCD64A1 product used in this study ([57]; Appendix A) does not account for fire intensity which could still negatively impact the landscape.

The deconstruction of species into morphological groups that are internally homogenous provided an opportunity for an improved understanding of the processes that underlie the patterns [50]. Despite this, in savanna ecosystems, research has focused on individual species (e.g., [21,24,45]) where findings are generally not always scalable to the wider ecosystem as species do exhibit idiosyncratic responses to the environment [75]. Subsequently, we feel that our analysis has related the importance of environmental drivers on the structure and physiological properties of the species, while it is not so specific that we cannot generalize processes to a scale that is useful for land managers.

It should also be noted that other factors may influence woody vegetation patterns. Topographic heterogeneity [76], atmospheric carbon [46], and harvesting [77] have all been found to influence woody vegetation communities. These factors were excluded due to the topographically homogenous landscape under study, and the fact that regional data on carbon and harvesting are di fficult to obtain; however, future research should continue to explore the impact of these factors. We also investigated time since last fire as a variable in the regression analysis; however, fire frequency was found to have more influence on woody vegetation patterns and was subsequently the only fire variable retained in the final models to prevent any issues of multicollinearity. Similarly, we measured grazing as density of cattle recorded from aerial surveys, although grazing could be represented using intensity (e.g., quantification of herbaceous tissue removal or an assessment of high, medium, or low). However, available data on such features was not available to this study. Recently, the statistical e ffects of spatial autocorrelation have been noted [78] and methods to incorporate and explore this into regression models have become more common [79–81]. However, we made the decision not to incorporate spatial autocorrelation in our analysis so that discussion could focus specifically on the environmental factors across the transect.

We used a combination of generalized linear models with Poisson error distributions and Tobit regression models to analyze our data. Biodiversity indicators such as species richness and abundance often exhibit distributions that are unsuitable for a number of statistical techniques. The literature surrounding the use of statistical analyses that do not account for lower limits to explore ecological questions is perhaps part of the reason we still have ambiguity surrounding the drivers of woody vegetation in savanna ecosystems. While our results corroborate the existence of well-established biodiversity-environment relationships (e.g., positive relationship with MAP), we also identified several novel biodiversity-environment relationships from the Tobit models (e.g., positive relationships with livestock). Subsequently, research should continue to explore more suitable statistical methodologies with which to analyze ecological data so that any managemen<sup>t</sup> strategies implemented from findings are better informed.
