This study was done in the SMLE, northern Botswana (Figure 1), an extensive region of woodlands and open grasslands wedged between two major wetland systems – the Okavango Delta and the Linyanti Swamps. The SMLE is characterised by a semi-arid climate with an annual rainfall of around 500 mm on the western side (Okavango Delta) and increasing eastwards to around 600 mm in the Chobe Enclave region (Botswana Meteorological Services). The climate is characterised by a wet season (December–April), cool, early dry season (May–August) and a hot, late dry season (September–November), where the daily maximum temperatures are regularly between 35 °C and 40 °C during this time (September–November) (Fynn et al. 2014). Water originating from the Angolan highlands into the ecosystem through the Okavango Delta, Linyanti Swamps, Selinda Spillway and Savuti Channel leads to the development of extensive floodplain grasslands and swamps adjacent to the extensive woodland systems (Mendelsohn et al. 2010). Alluvial clays and aeolian sands in the ecosystem are occupied by Colophospermum mopane and Philenoptera nelsii–Terminalia sericea woodlands (dryland woodlands), respectively (Wolski & Murray-Hudson 2006), and low-lying drainage systems are dominated by woodlands of mixed communities (Combretum imberbe and others) (Mendelsohn et al. 2010). A large paleolake system known as the MD (Figure 1), characterised by clays of lacustrine origin, occurs between the Okavango Delta and the Chobe region (Teter 2007). The extremely high clay soils of the MD provide key habitat heterogeneity in the general landscape of the SMLE, which is dominated by the Kalahari sands. Vegetation on the clay soils of the MD is characterised by open Senegalia/Vachellia (Acacia) spp. savanna grasslands and higher forage quality (Fynn et al. 2014).

Vegetation composition was sampled using 40 m × 20 m (800 m²) plots during the wet seasons of 2014 and 2015, with sampling conducted from January to the end of March to ensure that plants had attained inflorescences for easier identification. Certain remote habitats of the MD on water-logged heavy clay soils were inaccessible during the wet season and could only be sampled during the early dry season (April to mid-May) once the soils had dried out. Plots were stratified within homogenous vegetation units (determined visually by extensive field surveys of the region) to ensure adequate sampling of all key vegetation types within the study area (Figure 1). A total of 801 plots were located using two random numbers between 20 and 100, with the first random number taken along an access route such as a road and the second perpendicular from the road into the plant community. Plots were generally spaced at least 500 m apart in each plant community. The GPS coordinates of each plot were recorded using a Garmin GPS Map 62s. All plants (trees, grasses and forbs) rooted in the plots were identified, recorded and their percentage cover estimated. Unknown plants were pressed in the field and brought to the Peter Smith Herbarium (PSUB) collection at the Okavango Research Institute (Maun, Botswana) for identification, with nomenclature for all species following The Plant List (http://www.theplantlist.org). Five soil samples collected from each plot were mixed to form a composite sample, which was brought to the Okavango Research Institute laboratory for texture and nutrient analysis.

The data for all 801 vegetation plots (% cover abundance of species) were standardised using relativisations by maximum in Principal Component Ordination 6 (PCORD) (McCune, Grace & Urban 2002). These data were subjected to hierarchical cluster analysis (β linkage, β = -0.25, Sorensen distance) in PCORD 6 (McCune et al. 2002), comprising species (505 species) occurring at more than one site based on species distribution across 801 plots. Indicator species analysis (ISA) (Dufrene & Legendre 1997) was used to identify levels to define ecologically meaningful communities, and indicator values (IVs) were examined for statistical significance using Monte Carlo technique in PCORD 6 (McCune et al. 2002). Differences between communities were examined using multiresponse permutation procedure (MRPP) using Sorensen distance measure (McCune et al. 2002). In addition, non-metric multidimensional scaling (NMS) was used to plot the communities in ordination space in relation to key orthogonal gradients in the vegetation data using PCORD 6 (McCune et al. 2002). Such gradients are likely related to key environmental drivers and thus provide insights into how plant communities are structured in the ecosystem.

Owing to the vast size of the SMLE, we were unable to purchase detailed RapidEye imagery for the entire ecosystem and thus purchased RapidEye imagery to map the core area of the ecosystem, where we needed most detail for several herbivore studies currently being conducted there and then used Landsat 8 imagery to map the remainder of the ecosystem. A RapidEye composite image composed of 34 tiles and wet season Landsat Enhanced Thematic Mapper scenes from Earth Explorer USGS of the study area were radiometrically and atmospherically corrected in ENVI 4.8 (ENVI 2010). The Landsat images were mosaicked using seamless mosaic, and maximum likelihood classifier (supervised classification) in ENVI 4.8 (ENVI 2010) was used to map a subset of the 15 vegetation classes identified in the vegetation classification (reasons for mapping of a subset rather than all 15 vegetation classes is given in the results). A corrected RapidEye image was also mapped using the maximum likelihood classifier in ENVI 4.8 (ENVI 2010). The two classified images were seamlessly mosaicked together, and an area of interest was extracted by mask in ArcMap 10.2 (ESRI 2010). The area of each habitat was computed and converted into a minimum mapping unit in ArcMap 10.2 (ESRI 2010).

For mapping purposes, we required a habitat map that presented functionally distinct classes for herbivores. Thus, we grouped functionally similar classes (which, in addition, were generally difficult to distinguish from each other through remote sensing and thus difficult to map as separate communities), described by the cluster analysis (Figure 1-A1) and shown in the NMS (Figure 3). For example, the six sandveld types are not easily distinguished from each other through remote sensing and are functionally similar for herbivores, thus they can be mapped as one unit. By contrast, the communities on alluvial soils were very different functionally (e.g. wetland vs. mopane or riparian forests vs. dry floodplains) and could be easily distinguished through remote sensing and thus were mapped separately. With regard to the communities on heavy lacustrine clay soils, two distinct, functionally different communities occur, (1) the short grasslands with acacia spp. on silty soils and (2) the tall, open grasslands on vertisols, thus, functionally for herbivores, we chose to map them as two classes.

A coincident minimum of ISA p-value and maximum number of significant indicator species were found in seven and nine communities, respectively. Another coincident minimum of ISA p-value and maximum number of significant indicator species were found in 15 communities and were considered a meaningful ecological level of communities for a detailed vegetation classification. Thus, hierarchical cluster analysis done on all 801 vegetation plots recognised 15 main vegetation communities within which many sub-communities occurred (Figure 1-A1). MRPP tests of 15 communities suggested significant differences between communities (p < 0.000, Table 1-A1) with chance-corrected within group agreement, A = 0.265. Pairwise comparisons between communities suggested significant differences (p < 0.000, Table 1-A1).

The first axis of the NMS1 analysis appears to be a gradient of soil texture and fertility, with communities on the most sandy soils having the most negative values, communities on loam soils having intermediate values and communities on extremely high clay soils having the most positive values on NMS1, respectively (Figure 3; Table 1). The second NMS axis (NMS2) appears to be a weak gradient of wetness, with communities having the most positive values on NMS2 being those that receive some sort of seasonal flooding from the annual flood pulse into the Okavango and Linyanti systems (Setaria sphacelata – Gomphocarpus fruticosus community) or seasonal rainfall (Setaria incrassata – Dichanthium annulatum community) or occur near permanent water with perhaps a shallow water table (Tribulus terrestris – Senna obtusifolia community) (authorities for plant species names are according to http://www.theplantlist.org).

The NMS (Figure 3) delineated several vegetation communities in the SMLE that occur on deep Kalahari sands, and they appear to be differentiated according to subtle variation in the silt and clay content of the soil (Table 1). In the vegetation or habitat map, we refer to these communities on deep sands as sandveld communities (Figure 2). Although Baikiaea plurijuga – Baphia massaiensis is also found on deep sands (Table 1), we refer to it in the map as Baikiaea forest, not as sandveld (Figure 2).

These communities occur at intermediate levels of silt and clay (as compared with the lower levels of silt and clay of the sandveld communities) as a result of ancient alluvial deposition of sediments.

The vegetation of the SMLE was clustered into 15 major vegetation communities characterised by different herbaceous and woody species (Figure 3), with the primary axis of variation in plant community composition (NMS1) appearing to be driven by variation in soil texture and fertility (Figure 3; Table 1). The secondary axis of variation in plant community composition (NMS2) appeared to be driven by wetness, although it was not a clear effect as with the texture gradient probably because we focused mainly on dryland communities. The landscape template, which provides the basis of the vegetation heterogeneity of the region, was formed by a variety of processes. The mosaic of sand-filled paleoriver channels among alluvial deposits that supports the mopane–sandveld woodland mosaic between the Okavango Delta and the Linyanti Swamps was formed by ancient wetlands similar to the current delta, with the river channels subsequently becoming infilled by Kalahari sands of aeolian origin (McCarthy, Gumbricht & Mccarthy 2005). Similarly, the deep sands supporting the Baikiaea forests are of aeolian origin (McCarthy et al. 2005). The vast 3000 km² MD originates from Paleolake Mababe (Teter 2007), which has a central sump of about 70 km × 20 km, where lacustrine clays and sediments deposited, giving rise to the deep vertisols of the MD. Between the beachhead of the lake and the sump zone is a zone of soils where it appears that deposition of lacustrine clays declined with increasing distance from the sump zone. Thus, the inner sump has soils of the highest clay content supporting the Bothriochloa insculpta – Rhynchosia minima and Cenchrus ciliaris – Senegalia mellifera communities, followed by the edges of the sump with more silty soils supporting the Chloris virgata – Boerhavia coccinea community, then a zone of silty sands supporting mopane woodland and between the mopane and the beachhead of the lake are sandy soils but with higher clay and silt than the typical aeolian sands, which support the Boscia albitrunca – Dactyloctenium giganteum and Brachiaria nigropedata – Combretum hereroense sandveld communities (Figure 2).

Outside the MD, the typical aeolian sands have the lowest clay and silt contents because they received no lacustrine sediment deposition. The extremely low clay sands support the Ipomoea chloroneura – Oxygonum alatum, Eragrostis pallens – Ochna pulchra and Commiphora angolensis – Combretum collinum sandveld communities. Species such as C. angolensis, C. collinum, B. plurijuga, B. massaiensis, P. nelsii, O. pulchra, B. africana and T. sericea which dominated communities growing on sandy soils in this study were also found as indicator species for sandy regions in other studies (Coetzee et al. 1976; Gertenbach & Potgieter 1978; Tedder 2012; Tedder et al. 2013).

The mopane community was associated with alluvial soils of slightly higher clay than the sandveld communities (Table 1), as observed elsewhere (Tedder 2012; Tedder et al. 2013; Van Voorthuizen 1976; Wolski & Murray-Hudson 2006). The mopane and sandveld communities are widely distributed across the ecosystem, forming a woodland mosaic of mopane woodland on alluvial soils alternating with sandveld woodland on paleoriver channels infilled with Kalahari sands (Figure 2). The mopane–sandveld mosiac (as well as Baikiaea forest) provides key habitat for herbivore species favouring medium and tall grasses, such as buffalo, roan and sable antelope and elephant (Bennitt, Bonyongo & Harris 2014, 2015; Fynn et al. 2014; Sianga 2014; Taolo 2003), because of the abundance of digestible, leafy forage of high-quality grass species dominant in sandveld woodland (e.g. D. eriantha, Panicum maximum and Schmidtia papophoroides), as well as D. milanjiana and Panicum maximum in mopane woodland (Sianga et al. 2017b). Importantly, these high-quality grasses are most abundant far from permanent water (>15 km – 20 km) (Sianga et al. 2017b), which explains why buffalo tend to favour the woodlands furthest from permanent water during the wet season (Bennitt et al. 2014; Sianga 2014; Sianga, Fynn & Bonyongo 2017a; Sianga et al. 2017b), as do roan and sable antelope (Havemann 2014; Hensman et al. 2014). In addition, the numerous ephemeral waterholes of the mopane woodland allow herbivores to remain far out from permanent water in these woodlands during the wet season where they are able to avoid high concentrations of predators (Harrington et al. 1999). Once the waterholes dry up during the dry season, herbivores are forced to move closer to the permanent water sources of the Okavango Delta and Linyanti Swamps, where buffalo forage mainly in the wetlands (Bennitt et al. 2014, 2015; Sianga 2014; Sianga et al. 2017a), whereas roan and sable antelope visit the wetlands only every 3–4 days to drink and then return back to the safety of the woodlands far from water (Havemann 2014; Hensman et al. 2014). Thus, these vast woodland systems provide key functional habitat heterogeneity for provision of high-quality forage far from water during the wet season and low predation risk all year round.

The Croton megalobotrys – Setaria verticillata community, a riparian woodland (riverine) community, was correlated with silt–clay soils and occurred along edges of watercourses (Witkowski & O’Connor 1996). This community was dominated by species such as Croton megalobotrys, P. violacea, Combretum mossambicense, D. mespiliformis and S. nigrescens, which are adapted to obtaining soil moisture through lateral ground-water discharge from higher water tables (Ellery, Ellery & Mccarthy 1993; Hamandawana 2011; Ringrose et al. 2007). The Setaria sphacelata – Gomphocarpus fruticosus community, a floodplain grassland type found along watercourses, had grass (S. sphacelata) and sedge (C. longus) or forbs (G. fruticosus) as indicator species suggesting extensive wetness, as these species are mostly abundant in wetlands or swamp margins (Heath & Heath 2009). This community probably experiences periods of dryness over the annual cycle as indicated by the presence of opportunistic dryland species (C. platysepala, M. balsamina, S. triphyllum and B. hispidula). Variation in wetland community composition is driven by gradients of flood depth and duration, with C. dactylon often characterising the parts of the gradient with the lowest depth and duration of flooding, Panicum repens and S. sphacelata often characterising areas with intermediate depth and duration of flooding and tall sedges and grasses such as Oryza longistaminata and Vossia cuspidata characterising areas of the gradient with the largest depth and duration of flooding (Fynn et al. 2015; Murray-Hudson et al. 2011, 2014). This variation in composition and phenology on flooding gradients provides important variation in green forage supply for herbivores from the early to late dry season, owing to variation in availability of soil moisture for growth, allowing for adaptive foraging over the dry season (Fynn et al. 2015). Our sampling focused on the dryland communities, and we did not attempt to sample across the flood depth and duration gradient in the wetlands; thus, the Setaria sphacelata – Gomphocarpus fruticosus community represents only a small part of the variation in plant community composition that would occur in the region mapped as wetland in Figure 2. The spatial location of various wetland community types is not a constant and will shift location within the bounds of the area mapped as wetland (Figure 2) according to variation in flooding regimes over time. From a conservation management perspective, however, it should be recognised that the overall wetland community shown in the map (Figure 2) represents gradients of wetness and composition that provide critically important forage and adaptive foraging options for many herbivore species from the early to late dry season right across Africa (Fynn et al. 2015) and in the SMLE (Bartlam-Brooks, Bonyongo & Harris 2013; Bennitt et al. 2014; Fynn et al. 2014; Sianga et al. 2017a). Consequently, linkages between large wetland systems and adjacent dryland systems must be maintained to ensure that functional habitat heterogeneity is maintained (Fynn et al. 2015; Hopcraft et al. 2010).

Communities on the high clay soils of the MD, especially the Chloris virgata – Boerhavia coccinea and Cenchrus ciliaris – Senegalia mellifera communities, are critical wet season habitats for the large zebra migration in the region, as well as for wildebeest, tsessebe and impala (Fynn et al. 2014; Joos-Vandewalle 2000; Sianga 2014; Sianga et al. 2017a). This is because of the high clay soils and accumulation of a high concentration of minerals in the soil when it was a lake system (Teter 2007). The P-rich soils give rise to high P content in grass leaves (as well as other minerals) (Fynn et al. 2014; Joos-Vandewalle 2000; Sianga 2014). Thus, pregnant and lactating herbivores can obtain sufficient intake of nutrients to meet their high demands for nutrients during the wet season, a key functional aspect of wet season ranges for herbivores (Hopcraft et al. 2010; Owen-Smith 2004). In this regard, the Chloris virgata – Boerhavia coccinea community on the edge of the MD appears to be particularly important for P, having the highest soil P levels (Table 1), while the vertisols of the Cenchrus ciliaris – Senegalia mellifera communities deeper into the MD have the highest concentrations of potassium and calcium. This may explain why zebra are observed to switch their foraging bouts between Chloris virgata – Boerhavia coccinea community and the Cenchrus ciliaris – Senegalia mellifera community over the day (Sianga 2014), which may be a mechanism to maximise overall intake of key minerals, protein and energy during the wet season (Owen-Smith 2002). Also, the open grasslands of the MD provide better visibility, which reduces predation risk and is thus suitable as a calving ground. In fact, selection for low predation-risk habitat for the calving period may dominate the hierarchy of habitat selection decisions made by ungulates (Rettie & Messier 2000). During the day, zebra appeared to make use of short, open grasslands with high visibility (mainly the Chloris virgata – Boerhavia coccinea community) but at night, they moved further into the MD to the S. mellifera woodlands perhaps as an adaptive strategy to elude predators – reliance on sighting predators during the day and hiding from them at night (Sianga 2014).

Dominance of much of the southern half of the MD by C. ciliaris is informative as this species is often found in areas with elevated levels of P in soils (Blackmore, Mentis & Scholes 1990), which demonstrates why the MD is an important wet season range for herbivores. The Bothriochloa insculpta – Rhynchosia minima community is dominated by B. insculpta, which is adapted to seasonally flooded vertisols which form hard surfaces during winter (Cook & Clem 2000). The well-developed root systems of these large perennial grasses (B. insculpta and C. ciliaris) also likely promote access to soil moisture from deeper water tables when surface water dries out during the dry season (Cook & Clem 2000; Hamandawana 2011). The Setaria incrassata – Dichanthium annulatum community was dominated mainly by grass species such as S. incrassata, D. annulatum, P. coloratum, C. dactylon and D. milanjiana and occurred on seasonally flooded vertisols, with seasonal flooding maintaining this community as an open grassland with no trees (Cook & Clem 2000). Species such as S. incrassata and D. annulatum are well-known dominants of seasonally flooded, heavy clay soils in southern Africa (Cook & Clem 2000). This community type was found only in the north-eastern part of the MD and occurs in one of the most remote and inaccessible areas of northern Botswana, rarely ever accessed by people because there are no roads there and seasonal flooding of the heavy clay soils makes access impossible during the wet season. Our many sightings of roan antelope in this vegetation type while we were sampling suggest that it is regularly used by roan antelope, which is a species known to favour seasonally flooded grasslands and prefers areas with little disturbance by people.