Savanna fire and the origins of the ‘underground forests’ of Africa


  • Olivier Maurin,

    1. African Centre for DNA Barcoding, Department of Botany & Plant Biotechnology, University of Johannesburg, Auckland Park, Johannesburg, Gauteng, South Africa
    Search for more papers by this author
    • These authors contributed equally to this work.
  • T. Jonathan Davies,

    Corresponding author
    1. African Centre for DNA Barcoding, Department of Botany & Plant Biotechnology, University of Johannesburg, Auckland Park, Johannesburg, Gauteng, South Africa
    2. Department of Biology, McGill University, Montreal, QC, Canada
    • Author for correspondence:

      T. Jonathan Davies

      Tel: +1 514 398 8885


    Search for more papers by this author
    • These authors contributed equally to this work.
  • John E. Burrows,

    1. Buffelskloof Herbarium, Lydenburg, South Africa
    2. Department of Plant Sciences, University of Pretoria, Hatfield, Pretoria, South Africa
    Search for more papers by this author
  • Barnabas H. Daru,

    1. African Centre for DNA Barcoding, Department of Botany & Plant Biotechnology, University of Johannesburg, Auckland Park, Johannesburg, Gauteng, South Africa
    Search for more papers by this author
  • Kowiyou Yessoufou,

    1. African Centre for DNA Barcoding, Department of Botany & Plant Biotechnology, University of Johannesburg, Auckland Park, Johannesburg, Gauteng, South Africa
    2. Department of Environmental Sciences, University of South Africa, Florida, Gauteng, South Africa
    Search for more papers by this author
  • A. Muthama Muasya,

    1. Department of Biological Sciences, University of Cape Town, Rondebosch, Western Cape, South Africa
    Search for more papers by this author
  • Michelle van der Bank,

    1. African Centre for DNA Barcoding, Department of Botany & Plant Biotechnology, University of Johannesburg, Auckland Park, Johannesburg, Gauteng, South Africa
    Search for more papers by this author
  • William J. Bond

    1. Department of Biological Sciences, University of Cape Town, Rondebosch, Western Cape, South Africa
    2. South African Environmental Observation Network, Newlands, Western Cape, South Africa
    Search for more papers by this author


  • The origin of fire-adapted lineages is a long-standing question in ecology. Although phylogeny can provide a significant contribution to the ongoing debate, its use has been precluded by the lack of comprehensive DNA data. Here, we focus on the ‘underground trees’ (=geoxyles) of southern Africa, one of the most distinctive growth forms characteristic of fire-prone savannas.
  • We placed geoxyles within the most comprehensive dated phylogeny for the regional flora comprising over 1400 woody species. Using this phylogeny, we tested whether African geoxyles evolved concomitantly with those of the South American cerrado and used their phylogenetic position to date the appearance of humid savannas.
  • We found multiple independent origins of the geoxyle life-form mostly from the Pliocene, a period consistent with the origin of cerrado, with the majority of divergences occurring within the last 2 million yr. When contrasted with their tree relatives, geoxyles occur in regions characterized by higher rainfall and greater fire frequency.
  • Our results indicate that the geoxylic growth form may have evolved in response to the interactive effects of frequent fires and high precipitation. As such, geoxyles may be regarded as markers of fire-maintained savannas occurring in climates suitable for forests.


Savannas, ecosystems with continuous grass cover and variable woody cover, are one of the world's major biomes, reaching their greatest extent in the seasonally dry tropics. Over large parts of their climate range, savannas occur in mosaics with forests, leading to the idea that they are products of anthropogenic fire and deforestation. However, there is accumulating evidence that, worldwide, savannas are of ancient origin. Dated molecular phylogenies point to the origins of the C4 grasses that dominate savannas in the Oligocene (earliest at 32 million yr ago (Ma)) with considerable diversification already occurring by the Miocene (Christin et al., 2008, 2014; Edwards et al., 2010; Bouchenak-Khelladi et al., 2014). The savanna biome first became a prominent component of tropical vegetation from the late Miocene (c. 8 Ma) according to isotopic evidence from palaeosols and fossil teeth (Cerling et al., 1997). Thus, savannas were a major component of tropical vegetation millions of years before humans began to fell and burn forests. Studies have indicated that, once savannas began to spread, they expanded rapidly, reaching their maximum extent during Pleistocene glacial periods. Today, they cover c. 20% of the world's vegetated land surface (Collinson, 1988; Bond, 2008).

The causes of this rapid spread and the long delay between the origins of C4 grass lineages and their expansion into the savanna biome have been topics of intense research interest over the past 20 yr (Cerling et al., 1997; Sage, 2004; Keeley & Rundel, 2005; Beerling & Osborne, 2006; Osborne, 2008; Edwards et al., 2010; Scheiter et al., 2012). Ehleringer et al. (1997) were the first to suggest a general hypothesis for the appearance of C4 grasses, arguing from photosynthetic considerations that C4 grasses would first have outcompeted their C3 precursors in low latitudes with warm growing seasons and when atmospheric CO2 dropped below 500 ppm. They suggested that this threshold was passed in the late Miocene, explaining the rapid global expansion of savannas from that time. However, subsequent studies, using a variety of proxies, have shown that CO2 dropped below the 500 ppm threshold in the Oligocene, much earlier than the rise of the savanna biome, but consistent with the origin of C4 grasses and other lineages with CO2-concentrating mechanisms (Pagani et al., 2002; Arakaki et al., 2011; Beerling & Royer, 2011).

Although photosynthetic advantage may explain how C3 grasses were outcompeted by C4 grasses, it does not explain how grasses outcompeted trees, and therefore why grasslands replaced ancestral forests. C4 grasses are intolerant of shading and are rare or absent in closed forest understories (Ehleringer, 1978; Sage, 2001). For the savanna biome to have expanded, forests would have had to retreat. Increasing aridity is one potential pathway to forest retreat. Phytolith studies in central North America have shown that forests were replaced by C3 grasslands, which were, in turn, replaced by C4 grasses during the late Miocene (Strömberg, 2005). The mechanism for forest retreat has been attributed to increasing aridity from the Oligocene. Forest retreat as a result of growing aridity has also been invoked to explain the spread of grasses in Pakistan and Europe (Strömberg, 2011).

However, a climate-based hypothesis for the distribution of savannas does not explain why many contemporary C4 savannas occur as alternative states to closed forests in tropical landscapes (Hirota et al., 2011; Lehmann et al., 2011; Staver et al., 2011a). The presence of forests indicates climatic conditions that can support closed vegetation. Fire is increasingly being recognized as a key factor maintaining contemporary grasslands in a climate that can support forests (Bond et al., 2005; Lehmann et al., 2011; Staver et al., 2011b; Hoffmann et al., 2012). C4 grassy biomes currently account for c. 80% of the world's burnt area per year (Chuvieco et al., 2008; van der Werf et al., 2010), and many humid savannas burn several times in a decade and some burn twice within a year (Chuvieco et al., 2008; Archibald et al., 2013). The combination of high grass productivity given sufficient moisture, low decomposition rates of C4 grasses and a dry season suitable for burning every year provides the essential ingredients for the frequent fires characteristic of C4 savannas (Bond et al., 2003; Bond, 2008). Savanna fires can penetrate forest margins and, depending on the rates of postburn forest recovery, facilitate savanna advance into forest habitat (Kellman, 1984; Hoffmann et al., 2009, 2012; Murphy & Bowman, 2012). Because tree recovery from injury, such as fire damage, is particularly slow at low CO2, there may additionally have been synergies between fire-maintained savannas vs closed forests and atmospheric composition (Bond & Midgley, 2000, 2012; Beerling & Osborne, 2006; Kgope et al., 2010; Scheiter et al., 2012).

The demonstration of an empirical link between fire and the advance of savanna into forest has been a challenge. Charcoal records from marine cores show an exponential increase in fire activity coincident with the expansion of the savanna biome (Herring, 1985; Morley & Richards, 1993; Keeley & Rundel, 2005; Hoetzel et al., 2013). Unfortunately, there is no equivalent terrestrial fossil record. The landscapes in which fires are currently most prominent are deeply weathered and not conducive to fossil preservation. Thus, the fossil evidence is currently weighted towards the aridity route for forest retreat and savanna expansion (Strömberg, 2011). Dated molecular phylogenies provide an alternative tool to explore the origins of fire-adapted lineages and biomes (e.g. Bytebier et al., 2011; He et al., 2011; Midgley & Bond, 2011).

Simon et al. (2009) and Simon & Pennington (2012) used phylogenetic methods to infer the history of fire activity in Brazilian savannas (cerrado), and estimated the origin of woody plants restricted to the cerrado as < 10 Ma, with most savanna lineages dated as < 5 Ma, consistent with isotopic evidence for the timing of the spread of C4 grassy biomes. Thus, far from being an ancient vegetation type (e.g. Cole, 1986), cerrado is more likely to be a recently evolved biome. Savanna lineages differ from their forest relatives in a suite of fire-adapted traits: thick bark, reduced height and large underground storage organs (Coutinho, 1982, 1990; Hoffmann et al., 2003, 2004, 2009, 2012). One of the most distinctive growth forms characteristic of savannas is the geoxylic suffrutex (White, 1979; ‘geoxyle’ of Simon & Pennington, 2012). These are functionally herbaceous plants with ‘woody xylopodia underground but only limited and often short lived aerial shoots’ (Simon & Pennington, 2012; see also Coutinho, 1982, 1990; Appezzato-da-Glória et al., 2008). Xylopodia are underground structures consisting of ‘a lignified complex of root and shoot tissue with a high capacity to resprout and produce new shoot buds’ (Appezzato-da-Glória et al., 2008), and may provide an alternative adaptive escape route from fire. White (1979) described this growth form for African savannas and likened these plants to underground trees with branches buried and only the shoot tips and leaves emerging – he referred to them eloquently as Africa's ‘underground forests’. We use the term ‘geoxyles’ for underground trees hereafter.

Here, we locate the origins of the geoxyle life-form on the first comprehensive phylogeny of African woody plants to explore the origins of African savannas. First, we test whether woody species with this putatively fire-adapted growth form emerged at similar times to those of South American cerrado. Second, we explore the origins of the geoxyle life-form and consider whether these species provide a marker for the appearance of humid savannas through the fire mechanism of forest retreat. Last, we contrast the diversity of lineages contributing to the woody flora of African savannas with the flora of the South American cerrado.

Materials and Methods

Our study area includes the Zambezian region between 8°5′S and 34°5′S latitude and 11°7′W and 40°9′E longitude. It encompasses 12 countries, including Angola, Botswana, Democratic Republic of the Congo, Lesotho, Malawi, Mozambique, Namibia, South Africa, Swaziland, Tanzania, Zambia and Zimbabwe, with the Atlantic and Indian Oceans delimiting the region in the west and east, respectively.

Taxon sampling

A total of 1400 (of the c. 2200) woody plant species, comprising 117 families and 562 genera (Coates Palgrave, 2002; Schmidt et al., 2007; Boon, 2010; Van Wyk et al., 2011; Germishuizen & Meyer, 2013), were sampled over a period of 6 yr in southern Africa (Supporting Information Notes S1), including 53 of the c. 200 geoxyle taxa recorded for the region (Notes S2). We follow White's (1976) definition of geoxyles as plants that have a perennial below-ground woody root/stem, flowering and fruiting on seasonal and short-lived (resprouted) stems that do not exceed 1 m tall, and occur in areas receiving annual rainfall above 750 mm. As construed by White (1976), the term ‘geoxylic suffrutex’ is confined to those woody underground trees that belong to genera whose species are mainly otherwise trees and shrubs. He excluded from his definition those genera that have similar subterranean growth forms but which lack large tree relatives, for example Acalypha, Eriosema, Gnidia, Hypericum, Indigofera, Phyllanthus, Syncolostemon, Tephrosia and Vernonia. White's definition includes the most striking examples of closely related tall trees and functionally herbaceous ‘underground trees’ (Figs 1-3).

Figure 1.

Examples of geoxyles. (a) Lannea edulis (Anacardiaceae); (b) Combretum platypetalum ssp. oatesii (Combretaceae); (c) Morella pilulifera (Myricaceae); (d) Parinari capensis ssp. capensis (Chrysobalanaceae); (e) Elephantorrhiza elephantina (Fabaceae); (f) Dichapetalum cymosum (Dichapetalaceae); (g) Ziziphus zeyheriana (Rhamnaceae). Photographs courtesy of John Burrows (a) and Olivier Maurin (b–g).

Figure 2.

Geoxyles and their tree relatives. (a) Gardenia subacaulis (Rubiaceae); (b) Gardenia ternifolia; (c) Protea paludosa ssp. secundifolia (Proteaceae); (d) Protea roupelliae ssp. roupelliae; (e) Cussonia corbisieri (Araliaceae); (f) Cussonia arborea. Photographs courtesy of John Burrows (a, b, e, f) and Olivier Maurin (c, d).

Figure 3.

Three habitats rich in geoxyles. (a) Open woodland in north-west Zambia; (b) seasonally waterlogged grassland in north-west Zambia; (c) highveld habitat in Lydenburg, illustrating the misty conditions characteristic of the South African Highveld. Photographs courtesy of Olivier Maurin.

Phylogeny reconstruction

DNA extractions from leaf material, polymerase chain reactions and sequencing for the two plant DNA barcoding regions (rbcLa and matK) (CBOL Plant Working Group, 2009) were conducted using standard protocols (Hajibabaei et al., 2005; Ivanova et al., 2008). Complementary strands were assembled and edited using Sequencher v.4.8 (Gene Codes, Ann Arbor, MI, USA). The matK alignment was performed using Transalign (Bininda-Emonds, 2005). The combined dataset comprised 552 and 942 base pairs for rbcLa and matK regions, respectively. Voucher specimen information and GenBank accession numbers are listed in Notes S2 and on the BOLD DataSystem ( The phylogeny was reconstructed using 1400 taxa representing 117 families and 562 genera of Gymnosperms and Angiosperms. A total of 53 geoxyle taxa, representing 22 Angiosperm Phylogeny Group (APG) families, were included in the matrix.

A maximum likelihood (ML) analysis was performed on the combined dataset using RAxML-HPC2 v.7.2.6 (Stamatakis et al., 2008) on the CIPRES cluster (Miller et al., 2009), and enforcing topological constraints assuming the APG III backbone from Phylomatic v.3 (Webb & Donoghue, 2005). The phylogeny was rooted using representatives of Acrogymnospermae (Callitris, Cupressus, Cycas, Encephalartos, Juniperus, Pinus, Podocarpus, Stangeria, Widdringtonia and Zamia) (Cantino et al., 2007; Soltis et al., 2011). Branch lengths were then calibrated in millions of years using a Bayesian Markov chain Monte Carlo (MCMC) approach implemented in BEAST v.1.4.8 (Drummond & Rambaut, 2007), keeping the tree topology fixed. First, the RAxML starting tree was adjusted so that branch lengths satisfied all fossil prior constraints, using PATHd8 v.1.0 (Britton et al., 2007). Second, we assumed an uncorrelated log-normal (UCLN) model for rate variation among branches and the GTR + I + Γ model of sequence evolution for each partition based on the Akaike information criterion evaluated using Modeltest v.2.3 (Nylander, 2004). Third, we used 28 fossil calibration points from Bell et al. (2010) (Table S1) as minimum age constraints on the stem node of each group, except for the root of the Eudicots, which was set at 124 million yr, with a log-normal distribution following Bell et al. (2010). We performed four independent runs of MCMC, each for 100 million generations, sampling every 1000 generations. We assessed the MCMC log files for convergence using the effective sample size (ESS) statistics in Tracer v.1.5 (Drummond & Rambaut, 2007). The BEAST analysis reported ESS values > 100, indicating that the posterior estimates were not unduly influenced by autocorrelation. We combined the resulting tree files from the four runs in LogCombiner v.1.7.5 (Drummond & Rambaut, 2007), downsampling 1 in 20 000 trees, and discarding the first 25% trees as burn-in. The maximum clade consensus tree, with means and 95% highest posterior density (HPD) intervals, was generated with TreeAnnotator v.1.7.5 (Drummond & Rambaut, 2007).

Statistical analysis

To explore the divergences between African forest and savanna trees, we categorized each species as occurring predominantly in one of savanna, forest or the fynbos of the Cape Floristic Region (CFR). Savannas were characterized by the presence of a C4 grassy layer which forests lack (Ratnam et al., 2011), and were defined as tree–grass mixtures where C4 grasses form a near-continuous herbaceous layer (Ratnam et al., 2011). Forests, by contrast, were defined as closed woody vegetation casting too much shade to support a continuous C4 grassy layer. We did not distinguish between the different forest types, such as rainforests, dry deciduous forests or Afro-temperate evergreen forests. The biomes of the CFR are open shrublands, including fynbos, a fire-prone heathland and succulent shrublands (Mucina & Rutherford, 2006). We then used the phylogeny to identify sister species pairs falling exclusively within one biome type, and extracted their times of divergence. The phylogeny for southern African trees represents a regional sample of arborescent species, that is a particular small sample of angiosperms, thus making ancestral state reconstruction problematic because of the very large number of missing species (i.e. non-tree species and trees found outside of southern Africa). Importantly, our estimates of divergence times are conservative because nodes subtending sister pairs are less likely to be split by missing taxa than are nodes deeper in the tree. For example, it will be less common for sister species found within the African savanna biome to have a closer relative outside southern Africa – which would bias us towards the overestimation of divergence times. We assume that the sampling is equal among biomes (forest, 44.3%; savanna, 48.4%; fynbos, 7.3%), a reasonable assumption given the sampling protocol. We note that the shrubland vegetation of the fynbos is naturally tree species poor.

We compare the distribution of evolutionary ages between African savanna and forest sister taxa using a Wilcoxon rank sum test. For the reasons described above, we do not attempt to reconstruct ancestral ecologies directly, but we suggest that if sister taxa occur within the same biome, it is probable that they diverged within this biome. The oldest sister divergence within a biome may thus provide an approximate minimum age for the biome. Next, we examine evolutionary splits between geoxyles and their tree sisters. Here, we include only unambiguous independent shifts to a geoxyle life-form, where a geoxyle species is nested within a more inclusive tree clade (Table 1), thereby allowing us to infer directionality. The geoxyle life-form has apparently arisen multiple times in a few clades, possibly accompanied by reversals to a tree life-form (e.g. Fadogia, Clerodendrum, Elephantorrhiza and Salacia) (Fig. 1). However, taxon sampling within these lineages was insufficient to accurately reconstruct the location of evolutionary transitions, and we therefore restricted our analysis to the unambiguous shifts in life-form identified between sister groups.

Table 1. Ecological and environmental data characterizing southern African geoxyles and their tree sister groups
Scientific name (voucher)SistersMax height (m)SpinesPoisonousFruit type coding (0 = dry; 1 = fleshy)Fruit volume (cm3)aSeed size (cm)bFire return intervalMAP (mm)MAT (°C)Mean elevation (m)
  1. MAP, mean annual precipitation; MAT, mean annual temperature; NA, not available.

  2. a

    Assuming spheroid fruits.

  3. b

    Seed size is measured as the longest axis of the seed.

Carissa praetermisa (OM2650)23101NA0.53.396824.160
Carissa tetramera (RBN210)231010.52NA2.575921.8NA
Caesaria sp. (BB12551)110004.19NA1.7108124.8220
Casearia sp. nov. (Abbott9191)1200004.189NA2.292417.8630
Combretum engleri (OM1025)3400014.14NA2.649921.81120
Combretum platypetalum (OM2092)3300028.51NA4.098621.31070
Dichapetalum cymosum (OM2117)40.500013.09NA2.649620.91070
Dichapetalum barbosae (OM2374)4NA0000.79NANANANA90
Dissotis canescens (BB12691)51.80000.000130.753.6101522.0780
Dissotis princeps (OM3806)530000.00045NA2.893220.0790
Erythrina abyssinica (OM2095)610100NA1.22.977920.3920
Erythrina acanthocarpa (OM3916B)621000.36NA1.764715.71030
Erythrina caffra (BS0057)620110NA0.81.873718.3410
Erythrina humeana (OM741)74100NA0.82.580719.1680
Erythrina zeyheri (OM1589)70.51000.0421.72.575016.61450
Eugenia albanensis (BB7021)80.40011.77NA2.284520.1390
Eugenia capensis A (BB12289)940011.77NA2.477823.370
Eugenia capensis (Abbott9225)940014.19NA2.686021.2NA
Eugenia verdoorniae (Abbott9122)830014.19NA2.3102717.7470
Ficus capreifolia (OM2566)1070016.28NA3.074322.7420
Ficus pygmaea (MWC20237)1030011.15NA4.7111121.4990
Gardenia cornuta (OM2241)1150019.42NA2.383019.9470
Gardenia resiniflua (OM1272)1170010.00140.352.661921.5800
Gardenia subacaulis (BB12202)110.300185.080.53.7106321.41070
Jasminum fluminense (OM273)129001NANA2.160220.7830
Jasminum quinatum (T416)120.4001NANA2.577216.31440
Lannea discolor (RL1235)13150010.26NA2.664820.4950
Lannea edulis (OM1991)130.30010.40NA4.1100821.31020
Leptactina benguelensis (BB11158)140.40011.510.34.3112421.31090
Leptactina delagoensis (OM1598)1440010.0080NA2.773623.3220
Leucospermum gerrardii (MWC26648)150.4000NANA3.197417.2900
Leucospermum saxosum (BB12687)152000NANA2.499518.21090
Lopholaena coriifolia (OM & MvdB41)162000NANA2.470918.31270
Lopholaena disticha (OM3909)161000NANA2.683717.81080
Maerua andradae (LT1802)170.30NA1NANA3.6105325.3290
Maerua juncea (OM1592)1750NA114.73NA2.061121.5570
Maerua rosmarinoides (OM1476)1750NA1NANA2.584118.7770
Milletia makoudensis (LT1723)181.20008.25NA1.9105924.9250
Millettia usaramensis (OM2433)181000022.160.63.388922.7520
Morella brevifolia (OM3812)190.40010.014NA2.883316.41050
Morella serrate (Abbott9173)19100010.014NA2.671618.0960
Ochna arborea (CS03)20120010.26NANANANA810
Ochna confuse (OM3828)2020010.34NA3.0101520.51090
Ozoroa sp. (BB8074)210.60010.11NA2.796120.3590
Ozoroa albicans (BB8988)211001NANA1.569421.1580
Ozoroa laetans (BF12423)211.50010.28NA1.870620.5670
Ozoroa paniculosa (OM1948)2160010.37NA2.052319.91050
Parinari capensis ssp. incohata (OM3613)222.50010.79NA2.279622.6100
Parinari excels (BB10672)223500113.09NANANANANA
Paropsia braunii (BB10672)34100000.79NA2.674223.9180
Paropsia brazenana (Fishwick sn)3410003.39NA4.6104621.51000
Protea gaguedi (Turpin471)2410000NANANANANA1140
Protea parvula (OM3817)240.16000NANA2.587816.61340
Ritchiea capparoides (LT1805)2540NA1NANANANANANA
Ritchiea pygmae (LT1801)250.40NA1NANA3.0107425.1210
Searsia dentata (OM2251)2850010.0340.25NANANA1250
Searsia leptodictya (RL1655)2790010.0650.252.360919.01110
Searsia pendulina (OM1984)26100010.0140.63NANANA970
Searsia pondoensis (BT10242)2610010.0650.252.689017.6980
Searsia pygmaea (BL7355)270.20010.0590.253.0111417.61050
Searsia tumulicola ssp. meeseana (OM3818)280.70010.110.212.387517.41280
Searsia wilmsii (OM3910)260.50010.0650.252.482618.9940
Tetracera boiviniana (BB9126)294.50000.940.05NANANANA
Tetracera masuiana (BB11174)2910001.330.453.9116922.3990
Ziziphus abyssinica (OM2582)301310114.140.73.477423.0650
Ziziphus zeyheriana (OM3913)301.21010.270.52.363719.31020

For each tree–geoxyle pair, we contrasted differences in ecological traits (spines, toxicity, fruit type, root type and seed size) and environment across their geographical distribution (mean annual precipitation, mean annual temperature, mean elevation and fire frequency) (Table 1). Trait data were recorded from the literature (Coates Palgrave, 2002; Schmidt et al., 2007; Plants of Southern Africa, 2012). Fire frequency, precipitation, temperature and elevation variables were obtained by extracting the mean, minimum and maximum values within each species distribution. Precipitation, temperature and elevation variables were obtained from the worldclim database (Hijmans et al., 2005) and fire return frequency from Archibald et al. (2010). Statistical significance was evaluated using McNemar's test for binary traits, and a Wilcoxon signed rank test for continuous variables (Wilcoxon, 1945). When one of the sisters contained more than one species, contrasts were taken using the average of the species values. Our sister pair approach parallels Felsenstein's (1985) method of independent contrasts and, although it does not make use of all the information contained within the phylogenetic tree, sister comparisons provide a statistically robust method for comparative analyses (Barraclough et al., 1998).

Cerrado vs African savanna

To compare the phylogenetic affinities between the woody vegetation of the Brazilian cerrado and that of the African savanna, we extracted the list of woody plants and shrubs native to African savanna from the barcode phylogeny already described (586 species) and matching life-forms for cerrado listed by Ratter et al. (2012) (; 848 species). We then reconstructed a supertree of African savanna plus cerrado species using the Phylomatic online tool (Webb & Donoghue, 2005), and transformed branch lengths into millions of years using the BLADJ algorithm as implemented in Phylocom version 4.2 (Webb et al., 2008), calibrated with node dates from Wikström et al. (2001). Taxonomic mismatches (i.e. species or genera that could not be matched to higher taxa currently recognized by the APG (APG III, 2009)) reduced the final species set to 548 and 809 species for African savanna and cerrado, respectively.

Phylogenetic structure for the regional floras was first assessed by evaluating the proportion of shared branch lengths between floras using the PhyloSor index of phylogenetic beta-diversity following Bryant et al. (2008), as implemented in the Picante R-library (Kembel et al., 2010). Significance was determined by randomizing the placement of taxa across the phylogeny (999 replicates). Second, we compared the standardized effect size of the mean pairwise phylogenetic distance between species (the net relatedness index (NRI) of Webb et al., 2002) within each flora with a null model in which the same number of species are drawn at random from the phylogeny.


Evolutionary splits between African savanna tree sister species are significantly younger than between forest tree species (1.36 vs 3.15 Ma, median age for savanna splits and forest splits, respectively; W = 4555, < 0.001, Fig. 4), suggesting that speciation events in African savanna have been more recent than in forest, and/or that some forest tree species might either have closest relatives outside southern Africa or non-tree relatives.

Figure 4.

Distribution of evolutionary splits, in millions of years ago (Ma), between savanna (pink) and forest (grey) sister taxa from the dated phylogenetic tree of southern African trees. Areas of overlap are indicated in brown.

The distribution of the geoxyle life-form in Africa is phylogenetically dispersed (Fig. 5), indicating multiple independent evolutionary origins across disparate branches of the angiosperm tree of life, matching observations in cerrado Fabaceae, within the species-rich genus Mimosa, and, more generally, evidenced by the occurrence of cerrado geoxyles across numerous genera and families (Simon et al., 2009). Frequently, a single geoxyle species is nested within a species-rich tree clade, suggesting recent origins of the life-form. Geoxyles have evolved independently in all major lineages (orders) of eudicots, occurring in 30 plant families collected in our study, and arising multiple times in Fabaceae and Rubiaceae. In a few lineages (e.g. Fadogia), we observed multiple origins of the geoxyle habit, suggesting a complex evolutionary history, possibly including reversal to a tree habit, but current taxon sampling does not allow us to accurately resolve independent origins at such fine scales, and we therefore restrict our analysis to single congeneric transitions.

Figure 5.

Phylogeny of southern African woody flora reconstructed on the basis of DNA barcodes using a maximum likelihood approach after transforming branch lengths to millions of years ago by enforcing a relaxed molecular clock and multiple calibrations. Phylogenetic position of geoxyle life-form indicated in red.

We identified 30 independent transitions to a geoxyle life-form occurring across 26 genera (Table 1). Geoxyles have a median divergence date of 2.28 Ma, with a maximum age of 15.15 Ma for Dissotis canescens, which is unusual in that it tends to favour marshy environments. The vast majority of divergences have occurred within the last 2 Ma (Fig. 6). We find no significant ecological differences in geoxyles when compared with their tree sisters (Table 2); they are indistinguishable in fruit type, seed size, root type and herbivore defences. The principal ecological feature that distinguishes geoxyles from their close relatives is their large difference in height (V = 378, < 0.01). However, geoxyles differ significantly in their geographical distribution, being found in regions characterized by higher average rainfall (V = 31, < 0.01) and greater fire frequency (V = 57, = 0.02) than their tree relatives (Table 2). Furthermore, we observe a striking latitudinal gradient in divergence times (Fig. 6). The mean age of evolutionary splits is significantly younger moving towards more southerly latitudes (F = 9.10, = 0.005, r2 = 0.23 for the phylogenetic regression between latitude and log age implemented in the caper R library (Orme et al., 2012) and allowing lambda to take its ML value). However, the correlation is driven by an absence of older splits at high southern latitudes, whereas, at more tropical latitudes, we find both young and old divergences.

Table 2. Sister group contrasts between geoxyle life-forms and their tree relatives for various ecological and climatic variables
VariableTestTest statisticP value
  1. McNemar, McNemar's χ2 test; NA, not available; ns, not significant; Wilcoxon, Wilcoxon signed rank test.

HeightWilcoxonV = 378< 0.01
Fruit typeMcNemarNAns
Seed sizeWilcoxonV = 78.50.78
Mean fire frequencyWilcoxonV = 570.02
Mean annual precipitationWilcoxonV = 31< 0.01
Mean annual temperatureWilcoxonV = 1330.85
Mean elevationWilcoxonV = 1330.85
Figure 6.

Scatter plot of divergence times (million years ago (Ma)) between geoxyle life-forms and their tree sister groups against the latitudinal mid-point of the geoxyle geographical distributions (= 36). Horizontal error bars represent latitudinal range extents of geographical distributions; vertical error bars indicate maximum and minimum divergence time estimates between sister groups from the Bayesian posterior distribution of dated trees.

In the supertree of the woody taxa of the cerrado and African savanna (Fig. S1), we find that, unsurprisingly, there is less phylogenetic overlap between the two floras than expected by chance (PhyloSor index = 0.15 in comparison with the null expectation of 0.266, min = 0.246, max = 0.284). However, although cerrado taxa show phylogenetic under-dispersion (NRI = 1.747, = 0.038), perhaps reflecting some in situ radiations within a few lineages, the reverse is not true; African savanna taxa are not significantly clustered on the phylogeny (NRI = 0.339, > 0.05).


We have shown that the geoxyle habit evolved independently multiple times in Africa; however, it is always associated with the savanna biome. If we can identify the driving forces in the evolution of this life-form, we might, therefore, also provide insights into the origin and spread of savanna.

Previous work has suggested that shifts from forest to savanna were probably frequent in the South American cerrado (Pennington et al., 2006; Simon et al., 2009). Here, we show that African savanna trees represent an even more phylogenetically dispersed set of taxa, suggesting that shifts from the forest to savanna biome in Africa were at least as frequent. A paucity of endemic genera in both cerrado (Simon et al., 2009) and African savanna further supports parallels in the evolutionary origins of these biomes and a history of frequent shifts from forest to savanna with only few radiations within the savanna biome (see also Oliveira-Filho et al., 2013). In addition, we find that evolutionary divergences between African savanna trees are younger than those between forest trees. If we assume that evolutionary splits between extant species both occupying the same habitat type indicate a shared ancestor within that habitat, our results support a relatively young age for the African savanna biome – that is an absence of older splits between African savanna trees reflects limits to the upper age of the biome. Geoxyles have a median divergence time of 2.28 Ma, but with origins of many taxa dated to within the last 2 million yr, perhaps indicative of a more recent expansion of savanna. Further, we find a latitudinal gradient in age of geoxyle origins in Africa, with an absence of older ages in the south. If the evolution of the geoxyle life-form can be used as a marker for the presence of fire-maintained savanna ecosystems, our results would suggest a gradual expansion of African savanna from the equator to higher latitudes over the past few million years.

The evolution of ‘underground trees’

Four major factors have, at various times, been put forward as evolutionary drivers of the geoxyle habit in Africa. Burtt Davy (1922) hypothesized that winter frost on the South African Highveld was the main factor responsible for woody plant species escaping underground; in the South African context, cold temperatures may well have been a contributing factor to the evolution of underground trees. It is also possible that some woody species evolved the geoxyle habit in response to mammal herbivory, particularly on the grassy plains of the South African Highveld (Steenkamp et al., 2001). Africa, with its historically vast herds of antelope and other grazing ungulates, would have had much of its open areas subjected to incidents of severe herbivory. However, the region of greatest geoxyle diversity, the northern areas of the Zambezian Domain (White, 1983), receives little or no frost and the Zambezian Region is dominated by miombo woodlands, a habitat type noted for its paucity of herbivores as a result of the poor nutritional quality of its forage (Frost, 1996), and a long, harsh dry season of almost 7 months (Rodgers et al., 1996). Further, we find no general relationship between the geoxyle life-form and temperature or elevation, and the prevalence of herbivore defences (e.g. presence of spines or poisons) does not differ between geoxyles and their tree relatives. It would seem, therefore, that neither frost nor herbivory has been the major selective pressure in the evolution of the geoxyle habit.

Frank White, in his landmark paper on the underground forests of Africa (White, 1976), attributed the rise of the geoxyle habit mainly to edaphic factors. He pointed out that the open grassy edaphic seasonal wetlands, which are a characteristic feature of the south-central African landscape, support a number of geoxyles, particularly along the margins of these open wetlands or ‘dambos’. These seasonally waterlogged grassy depressions often occur on sandy, oligotrophic soils, which are anaerobic when waterlogged. However, although a handful of geoxyles (Ficus pygmaea, Syzygium guineense ssp. huillense, Erythrina baumii, Protea baumii) are confined to dambos, the majority are not. For example, the region of high geoxyle diversity in north-eastern South Africa (this paper) have no endemic wetland geoxyle species, whereas, in the geoxyle-rich coastal grasslands of the Maputaland region, the geoxyles grow away from the numerous wetland depressions (Matthews et al., 1999). Notably, these grassy dambos are subjected to frequent burning, and it is possible that the open, sunny environment of the dambos would also provide advantages to a dwarf woody plant that would otherwise be shaded in the taller adjacent woodlands. This link with fire brings us to our final putative driver of the geoxyle habit.

Geoxyles in Africa are restricted to savanna habitats or upland grasslands (geoxyles are absent from closed forests). They occur almost exclusively in higher rainfall savannas with frequent fires (White, 1979). Indeed, the seasonal or regular burning of the woodlands and grasslands of Africa is the most often quoted factor driving the evolution of the geoxylic suffrutex habit in Africa (White, 1976 (in part); Vollesen, 1981; Lock, 1998, 2006; Matthews et al., 1999), as well as in the South American cerrado (Simon et al., 2009; Simon & Pennington, 2012). Although the majority of savanna trees have developed features to withstand fire (thick bark, fire-resistant shoots), geoxyles may have escaped fire by developing their woody component below ground, thus minimizing their resource input into annual vegetative growth to the benefit of flower and fruit production. Most conclusive is the fact that it is clear that, wherever geoxylic suffrutices occur in Africa, fire is a regular feature in the landscape. Our results show that the occurrence of the geoxyle life-form correlates significantly with fire and precipitation, suggesting that both may have contributed to driving trees underground.

How would precipitation and fire select for trees developing underground? Areas of high precipitation should favour forest development, and the proportion of African landscapes covered by forest increases along a precipitation gradient (Lehmann et al., 2011; Staver et al., 2011a). However, greater precipitation also correlates with high grass productivity, thus favouring frequent fire occurrence that maintains the savanna state (Bond, 2008; Lehmann et al., 2011; Staver et al., 2011a,b). Fire is a major selective pressure. Population studies have shown that a key demographic bottleneck for savanna trees is the transition from juvenile plants growing below the flame zone to adult plants taller than flame height and resistant to fire damage. Saplings may persist for decades in the flame zone (the ‘firetrap’) without growing into mature trees (Trollope, 1984; Scholes & Archer, 1997; Williams et al., 1999; Higgins et al., 2000, 2007; Werner et al., 2006; Prior et al., 2009; Werner & Franklin, 2010; Bond et al., 2012; Werner & Prior, 2013).

We suggest that the geoxylic growth form is advantageous in areas experiencing the interactive effects of frequent fires and high precipitation. The underground trees of Africa can be regarded as markers of fire-maintained savannas occurring in climates suitable for forests (extensive African savannas also occur in arid climates where fires are rare; Lehmann et al., 2011). Poor growing conditions that reduce growth rates of juvenile trees would also result in reduced probabilities of reaching fire-proof sizes and transitioning to tall mature trees. Geoxyles are common where site conditions reduce growth rates, such as on seasonally waterlogged and/or low-nutrient soils, or, in South Africa, at high-altitude sites with cold winters (White, 1979). Thus, geoxyles appear to be an example of heterochrony (a change in the relative timing and/or rate of developmental processes; Li & Johnston, 2000) with ancestral tree growth forms adopting a dwarf stature from which, unlike their tree ancestors, they are able to flower and fruit.

Fire and the savanna biome

Our prime aim in this study was to use the origin of geoxyles to help date the emergence and spread of fire-dependent African savannas. However, the limited literature on geoxyles, especially in Africa, raises other questions, including the general nature of fire adaptations in savannas. Fire-adaptive traits have been extensively studied in crown fire regimes, especially in Mediterranean-type shrublands (Keeley et al., 2012). Fire-adaptive traits in the surface fire regimes of savannas are not nearly as well studied or understood. As noted by White (1979), geoxyles appear to be a convergent life-form in South American and African savannas. The link with fire in cerrado has been elaborated by Simon et al. (2009) and Simon & Pennington (2012). For African savannas, we have argued above that the growth form has evolved in response to the selective pressures imposed by frequent grass-fuelled fires and soil or climatic conditions that slow woody growth rates. Curiously, geoxyles are absent in northern Australian savannas despite similar fire regimes and soils to those of Africa and Brazil (White, 1979). They have also not been reported, perhaps because they have not been recognized, in the higher rainfall savannas of South and South-East Asia (White, 1979) or the pine savannas of North America (Noss, 2012). Within Africa, White noted a striking difference in the richness and abundance of geoxyles in the savannas of the Zambesian region (this study) contrasting with savannas of the Sudanian region (north of the equator) where they are rare and poor in species. The patchy occurrence of geoxyles in savannas from different geographical regions is perhaps to be expected given the youth of the biome. As noted above, both African savannas and the cerrado have few if any endemic genera, consistent with a young age for the biome in both regions (Simon et al., 2009; this study). The coincidence of a largely Pliocene age for flammable savanna origins in Africa and cerrado is striking. The causes of the abrupt and apparently near-simultaneous assembly and spread of these fire-dependent savannas are unknown, but are an area of active research (Keeley & Rundel, 2005; Scheiter et al., 2012).


We explored the origin of mesic savannas in Africa using geoxylic suffrutices, White's underground forests of Africa, as markers for fire-maintained ecosystems. Our results suggest that these savannas first appeared in the tropics with more recent speciation at lower latitudes in southern Africa. Dates of origin of the geoxyles are mostly from the Pliocene (< 5.3 Ma), consistent with the appearance of woody fire-maintained cerrado plants in Brazil (Simon et al., 2009; Simon & Pennington, 2012). As the taphonomy of fossil savanna sites biases against the detection of fire-maintained savannas, we suggest that phylogenetic approaches are particularly useful in tracing the origin of more humid savannas. As more complete phylogenies become available for clades from the African fire-maintained savannas, and as more species are sampled for fire-adaptive traits, more refined dates of origin should become available. Nevertheless, this study provides the first evidence for dates of emergence of higher rainfall savannas in Africa, and supports the role of fire in their origins. Furthermore, our study suggests independent origins of this growth form from those of Brazilian cerrado. The fire frequencies characteristic of seasonally humid savannas are perhaps the highest in Earth's history. The diverse responses of plants to this extreme selective regime warrant further study.


We thank the Government of Canada through Genome Canada and the Ontario Genomics Institute (2008-OGI-ICI-03), the International Development Research Centre (IDRC), Canada and the University of Johannesburg for financial support, and various local and international authorities who granted plant collection permits. We thank Mike Bingham and Sandie Burrows for valuable discussions during the preparation of the manuscript. We thank two anonymous reviewers for comments on an earlier draft of the manuscript.