S. Mc. J. Bot. 1999, 65( I): 89- 96 89

Predicting the distribution offynbos and succulent karoo biome boundaries and plant communities using generalised linear models and geographic information systems

R.G. Lechmere-Oertel and R.M . Cowling* Institute for Plant Conservation, Department of Botany, University of Cape Town, Rondebosch, 7701 Republic of South Africa

Received 28 Alay 1998: revised 18 NOl'ember /998

Vegetation survey data from Matjiesrivier Nature Reserve, in the eastern Cederberg Mountains of the fyn bos biome, were used to model the distribution of fynbos and succulent karoe biomes and plant communities. Predictions were based on generalised linear models (GLMs) with the presence or absence of a biame or community as the response variable. Geology, altitude, rad iation and landtype, which were derived from remotely-sensed sources (maps, aerial photographs and digital terrain models), were used as explanatory va riables in the GLMs The boundary between fynbos and succulent karoo could be predicted accurately based on geology and altitude alone. Fynbos occurred on sandstone above 800m; succulent karoo occurred on sandstone below 800m, or on shale at any altitude. Community distributions were less accurately predicted, but could be used to generate a geographic information system (GIS) map similar to a vegetation map derived from ground survey. These models can be used to map vegetation over large areas in the fynbos/succulent karoo ecotone with relatively little effort or cost.

Keywords: Biome boundary, fynbos, GIS, growth forms, modelling, succulent karoo.

*To whom correspondence should be addressed. E·mail: rmC@botzoo .uct .ac.za

Introduction Cape. He concluded that topo-climatic factors and geology were Static and non -predictive vegetation maps are problematic for a the primary determinant of this boundary. Cowling el al. (1997) number of reasons. Firstly, vegetation is dynamic: composition have reviewed the literature on the fy nbos/succulent karoo and physiognomy change over varying scales in response to boundary, and concluded that moist ure availability, rather than environmental change, biotic interactions and alterations in di s· geology, is the primary determi nant. However, despite the turbance regimes. Secondly, such maps cann ot be used to predict importance of biome boundary determinants for predicti ng the changes in vegetation boundaries, either in response to impacts of global change on biome distributions, no attempt has short-term fluctuations (e.g. drought effects or changes in the fire been made to model the tyn bo S/succu lent karoo boundary in regime) (Austin 1991; van der Rijt el al. 1996), or lo nger-term terms of easil y measurable explanatory variables. Thi s study changes induced by, for exampl e, climate change (N ielson attempted to model the di stribution of communities and th e fyn­ 1993). Thi rd ly, ' snapshot' maps cannot be ex trapolated to large bos/sllccu lent karoo bound ary in the study area using several eas­ areas of unsurveyed terrai n (Franklin 1995). This increases the ily measurable environmental variables. We asked three time, and hence the cost, of vegetation mapping, since each new questions: area must be re-mapped from scratch (Margules & Austin 199 1). I. Can the distribution and environmental determinants of th e On the other hand, predictive vegetati on maps, especially major communities be predicted accurately? when derived from models based on easily measurable environ­ 2. Can th e geographical positi on and environmental determinants mental vari ables, have many advantages (Frankl in 1995). They of the boundary between fynbos and succul ent karoo be pre­ can be extrapolated over large areas of remote terrain, thereby di cted accurately? reducing th e survey effort for checking predictions, and can be 3. Can the models derived from the first two questions be used in used for monitoring and predicting vegetation change in a GIS to produce realistic maps of th e vegetation of th e study response to global change. Generalised linear models (G LMs) area? have com mon ly been used to derive predicti ons on the environ· mental contro ls of species' (Margules el 01. 1987~ Austi n et al. Materials and Methods 1990; see review in Franklin 1995) and community (Austin el at. Study area 1983, 1984; Valverde & Montana 1996) di st ributi ons. Maps are usually created in a geographical information system (G IS). Our Matji esrivier Nature Reserve (MNR) (32°25'S. 19°17'E) straddles interest in this paper is the distribution of plant communit ies the boundary between fynbos and succulent karoo bi omes in the associated with the transition between the fy nbos and sllcculent Western Cape, South Africa. MNR experiences a mediterra­ nean-type climate and has a gradient of decreasing an nual rai nfall karoo biomes in the eastern Cederberg Mountains. This region from th t! west (300 mm) to the east ( J 00 mm) (Lechmere-Oertel includes the Matjiesrivier Nature Reserve (th e study area) and a 1998). The geology of MNR comprises three groups in the Cape large area of inhospitable terrai n that is being incorporated into Supergroup. The shale and silty sandstone of the Bokkeveld Group the conservati on system as 'conservancies'. Th erefore, there is a lie between the sand stone and quartzite of the Table Mountain and need to derive a vegetation map for conservati on planning in th e Wittebe rg Grou ps. The conformity between the rocks of these diffe r· area, specifically to assess the extent to whi ch different plant enl geo logic groups is abrupt. giving rise to steep edaphic gradien ts. communities are represented in the system. These climatic and edaphic gradients are reflected in a transition We were also interested in assessing the determinants of the from fynbos into succu lent karoo over distances of a few meters to a fy nbos/succul en t karoo boundary in the region. Euston-Brown f\!w kilometers. Overall, the vegetation of the MNR is floristically ( 1995) used GLMs to pred ict th e distribution of families and and structurally complex. since it compri ses elements fro m two di s­ growth forms diagnostic fo r fynbos and thicket biome vegetation tinct and typically diverse vegetation types: mountain fynbos and in the Kouga-Baviaanskloof mountain complex in the Eastern lowland succulent karoo (Low & Rebelo 1996) (Table t). 90 S. Mr. J. Bot. 1999. 65( I )

Table 1 The major vegetation types in the Matjiesrivier categories (Table 2). After preliminary modelling, gravel plains were grouped with rocky slopes to reduce model instability. Nature Reserve (MNR) and their relationship to biomes and broad-scale vegetation types Model development Vegdatinn typc in MNR (ex momc and vcgcttltion t) pe (sensu The data used for model derivation "ere not distributed normally Lcc hnere~Ocrtel 1998) Low & Rebel o 1996) and included both categorical and continuous variables. Generalised AstcmCC()llS Fynbos Matrix Fynbos biome linear models (GLMs) extend the regression framework to three situ~ ations \vhere ordinary regression would not be appropriate. These RestiOld Sandy fynbos MOllntain Fynbos arc: (I) where the data are not normally distributed; (2) \vhere the data need to he transformed (using a link function) bcfor!; a li near Shale SUccllknt Ktlroo Succulent Karon nimHe model can be titted: and (3) \vhere the data comprise both categori~ Sandy Succuknt KtlnJO Lowland Succulent Karon cal and continuous variables (Genstat 5 Committee 1987: Crawley 1993: Trexler & Travis 1993). There are three main issues that need SucCtlkny Karon Matrix to be considered when using GLMs to predict the distribution of communities: These arc: (1) the selection of an appropriate statistical model; (2) the selection ofsuitahle variables to be used as predictors: Vegetation data and (3) the critical evaluation of the titted regression model for out­ Data comprise 125 lOin x 10 m sites samplLd for species and liers and intluential observations (N icholls 1991). gnl\\"th form composition and subsequently classified into commun i ~ We modelled each of the biomes and major communities Cfable tics using standard multivariate methods (st.:t.: Lechmere~Oerlel 1998 1) identified in the MNR, except for azonal types (Dwarf Bedrock fo r details). The sites were positioned through MNR using an Shrub land and Kloof Thickt:t: see Lechmere-Oertel 1998). sepa­ approach similar to that employcd by gradst.:ct sampling (Gillison & ratdy agains.t the explanatol}' environmental variables. The response Brewer 1985: Austin & Heyligers 1(91). variable indicated the presence ( I) or abs!;ncc (0) of the relevant attribute (community or biome) in a site. Therefore. in each of the Explanatory variables models, the distribution of the response variable was assumed to be We chose environmental explanatory variables thtlt \vere rt.:adily binomial; therefore, a 10git link function \vas used (McCullagh & available from remotdy ~sens ed or computcr~generated data sets, and Nc1der 1989). The sites used in these analyses were assumed to be that rdlected current hypotheses on the determinants boundaries independent observations of the vegetation. be tween fynhos and adjacent biomes ~ principal ly soil moisture sta~ Significant explanatory variables were selected for the models tus and soil nutrient status (Cow·ling et a/. 1997). Rainfall. evapora­ using a forward step-wise process (GENSTAT 5 Committee 1987, tion and soil texture (amongst other hlctors) control moisture Crmvley 1987). The process of model building was iterative. A ll the availability whereas soi l nutrient status depends largely on geologi­ explanatory variables were added to the model singly on the tirsl cal parent material. While it is possible to derive cl imatic data from pass. The variable that accounted for the highest significant change accessible dawbases, most soil factors n:quire intensive sampling. in deviance was added to the model. All the remaining variables l-l owever. in the fynbos biomt.:. geology has a strong intluence on a were then added to the new model singly and the one with the high­ number of soil variables (Cowling & I·lolines 1991) and has been est significant change in deviance retained. The null hypothesis ~ found to bl.! an adequate explanatory surrogatt.: for these variables that there \\-·ould be no change in the likel ihood ratio (deviance) for (Ellston~l3nl\\n '995). each addition ~ was tested by comparing the change in deviance with Five explanatory variables \veft~ chosen that renect so il moisture the l! statistic (Crawley 1993). This was repeated until no more sig­ and nutrient availability: mean allnllal rainfall, altItude. radiation. nificant variables remained. In each case, the interactions between geology and landtype Crable 2). Although rainfall and altitude arc the main explanatory variables were tested first. The main variables closely correlated in the study area (Lechmere-Oertel 1998). both were only tested if there were no significant interactions. The contin~ were included in the initial model to determine wh ich had better uous variable, radiation. \vas litted both as a linear and quadratic explanatory power. Annual rainfall data were derived from a surface funct ion to test for possible curvature in the response bel\veen the interpolation model (CCWR 1996) that is based on averaged rainfa!l transformed probability and the variable (Nicholls 1991). The tinal measurements (from climate stations with over 20 years of data) tlnd topographic variables such as altitude. aspect. shading and slope angle. The interpolated rai nfall data \vere incorporated into a GIS on Table 2 The final categories for explanatory variables a Illinllle ~ by~minute grid. The nearest data point to each site was used in the generalised linear models (GLMs) to predict used as the mean annual rainfall value for that site. Preliminary anal­ ) st.'s indicated that rainfall cou ld be collapsed into t\VO categories, the distribution of community and biome types in the ahove or helow 200 mm (Table 2). without a significant decrease in Matjiesrivier Nature Reserve e'\pianatory power. Variable Category Variable description Topographic data (altitude. a..'ipect and slope angle) were derived from a surface elevation model (SEM) generated in a GIS from digital Table Mountain and Witteberg Group sand~ 1:50000 map sheets 32 1 9CB, 321 9AD (Surveyor General. Mow­ Geology I stones bray. Cape Town). Preliminary analyses indicated that altitude could Geology2 Bokkeveld GrouB:shales be collapsed into t\','o categories. above or below 800 m Cfable 2), with almost no loss in explanatory power. Solar radiation data for Rainl < 200 mm yr· ! summer. \~ int!;r and the equinox were calculated for each site from 1 aspect and slop!; angle regression equations (Schulze 1975) for the lat~ Rain2 > 200 mm yr· itudc 32°S. Both aspect and slope angle \\lerc derived from the SEM. Altitudel < 800m Three geological categories (Table Moulltain Group. Witteberg Group and 130kkcvdd Group) were digitised from a 1:250 000 geological Aititude2 >800m sheet (32 19 Clanwi1Jiam) and \:10 aerial photographs. Prelimi~ aoa Landtypel Rocky talus slopes and flat gravel plains nary analyses. indicated no distinction in explanatory power between Table Mountain Group and Witteberg Group sandstones: therefore, Landtype2 Bedrock sheets these were grouped in subsequent analyses(Table 2). Landtype3 Flat sandy plains The same aaial photographs were used to digitise fall I' land type S. Mr. J. Bot. 1999,65( 1) 91

minimum adcquat~ modt!l contained only the signi licant cnviron­ Results mc::nta l variahles. Predicting plant community distributions Preliminary analyses shO\vcd th at sotnl.! of the categories in 1he Overall , the distributions of the five communities (Table I) were environmental variabks iwd a n.:sponsc that wns always zero or one. not \vell explained by the three most significant environmental vor \!:xample. fynbos communitks never occurn.:d on shale (geology variables: altitude, geology and landtype. Although not a reliable 2). and Shale Succulent Karoo was thl! only community t!vcr to measure of goodness-of-fit for models w ith categorical response occur on shale. V:uiablcs with such slructuml leras or ones wen; variables, the low deviance of the m ini m um adequate models n:mo"t:d from the model as the; cilust:d instability (Lindsey 1989). (Table 3) indicated poor explanatory power. In all models, alti~ ye t they were sti ll biologically meaningful. These unstClblc variables tude proved to be a better explanatory variable than rainfall and Wl..!n! incorporated into the !inal models as a sta Lt:ITIcnl of fact, hUI was, therefore, used as a surrogate variable for tem perature and did not in ll uencc the regression equation (Lindsey 1989). moisture regime [rainfall increases and mean temperature decreases with increasing altitude in the study area (Lech~ Critical evaluation of the model mere~Oerte l 1998)]. Radiation did not emerge as a significant Rcgn.:ssion models litted In data for predictive purpost!s must hc explanatory variable in all analyses. evaluated in tenl1 S of some simple diagnostic mcasun:s (Nicholl s The p robability of a comm unity occurring at a sile was calcu~ 1991). In all 111odcls. the fi tted va lucs , .. ere calculated as it function lated by substituting / ( the linear predictor ) fro m each or the fol~ of the observed values and the residuals examined. Also. the error lov,,'ing equations (4-8) into equation 2. A comparison of the variam:e was plotted against the response variabh!. fitted values and the response variable was used as an indication The aim of the moddling was to predict the occurrence of of the goodness -of~fit for all models. c\liJrs~.>sca k ,·cgctation units \\ ithin a landscape llsing il few easily ll1i!asured environmental variables. Bearing this in mind, lillIe effort Astel'GceOliS Fynhos /vlalrix (AFiVl) was madc to adjust the models onCI! a suitable sc-t ofprt:dictor varia~ oks had necn chosen, even when the standard errors. residua ls or Only altitude and geology caused a significant change in d ev i~ liul!d values suggested that the model was nol enti rdy stable. ance w hen added to the Illodel. Howeve r, geology 2 (shale) had a very high standard error that caused model instability. This was Model predictions because A FM was never fou nd on shale: its occurrence was a structural zero that distorted tbe model. Therefore, geology was Pour steps ,\ ere used to generate pn.:dictiVt: vegetation models: (1) removed from the model and treated as a statement of fact. T he the gC ll l!ration of the GLMs d(:scribed above: (2) lllcorporation of fina l model (eq. 4) did not predict the occurrence of AFM well. thl.! significant e:-.planatory variables into a spatial framework using Of the 36 sites with fynbos, the m odel correctly predicted 33 GIS: (3) gell!.!ration of predictions on the spatial distribution of the sites. However, the model incorrectly predicted the occurrence of live communitks (Tabh! I ) and the boundary uf the I)'nbos biome. AFM in 39 sites (Table 4). The final model for AFM was: bast.:d un these variables: and (4) a comparison of the pr!.!tiicted vcg~ !.!Iation mup wilh a map c.lerived from a ground survey ( Lec h~ On sandstone / ~ -2.686 + (2 .375 * al titude2); !11t.:re~Oertel 1998). o n shale / = - 0:> (eg.4) Polygon covers of tht.: three categorical environmental variables finally used in the rnodds. namely altitudl.!. geology and landtype. The model predicted that AFM would be fo und o n sandstone were generated in 0. G IS. A composite cover o( all the environmental geology above 800 m (p = 0.4 ± 0.08), irrespective of laodtypc. variables significant 1Il a GLM \\as built hy overlaying the indiv i d ~ Sites below 800 m would not support AFM (p ~ 0.06 ± 0.04). nor ual covers within Ihe GIS. The composite cover for each model thus would sites with shale~derived soils. com prised a Humber o f polygons that had unique combinations of the signi fi cant environmental variables. The output from a G LM was Reslioid Sandy FyI/bas (RSF) a regression cquation: Landtype and a ltitude were the o n ly significant variables in the model (Table 3). However, altitude I « 800 In ) caused instab ility (eg. I ) in the model because R SF was never found below 800 Ill . Thus, where / is the linear predklo~ ; a is the n:gression constant: and hI is a ltitude was treated as a state ment of fact and removed from the the regression coefficient for thl.! explanatory variahle Xl. The proba~ lll odeL Similarly, RSF was never found all bedrock sheets hi li t)' (p) of il vegetation unit occurring in a ]1o lygon was calculated (Iandtype 2) and this was also treated as a statement of fact. The by: final model (eq . 5) did not predict the occurrence o f RSF accu~ rately. Only 9 o f 15 s ites were correctl y classified (Table 4). The p ~ ex p' - (exp' + I) (eg.2) Illode l a lso incorrectly predicted the occurrence o f RSF in fline The standard error of p was calculated by: sites. The final model for RSF was:

III = ~O'J: se ~± 1.96v[p *( I-p) - nJ (eq.3) < 800 or on bedrock / > 800 m and on rocky slopes or sandy plains / ~ -3.481 + (3.481 The values of the environmental vari ables from each polygon in the • landtype3) (eg . 5) composite GIS cover were substituted into the regression equation from the G LM. Thus. the probability of a comm unity occurring in The occurrence of RSF could thus be predicted , a lbeit rather cvcry polygon in the composite GIS cover was calcu lated from the poorly. o n the basis o f land type a nd altitude. RSF was restricted GLM. The polygons wen.! colour codt:d aCl:nrdi ll g to Ih !.! ir p rohabil ~ to high altitude sandy plains (p = 0.5 ± 0.08). Geology was not ity values. thus giving a map showing the areas whert! that particular significant in the model, although RSF never occurred on shale . L:ommunity was most li kdy to Decur. In the cas!.! of tht: coverage with all the major commun it ies. where there was an overlap of po ly~ Shale SucculenJ Karoo (SSK) gons with ditfcrent prohabilities. the polygon with th~ high!.!st proba­ Geology was the only variable to cause a significant change in hility of occurrl!nCt: superseded the others. The predicted deviance when added to th e SSK model (Table 3). Geology 2 distri butions o f communities and the fynbnslsuccull!tlt karao bou nd~ (shale) caused model instability because SSK was the only CO Ill~ ary were mapped in the GIS and compared visually against the munity ever found on shale. T herefore. shale was treated as a actual vegetation map. state ment o f fact and removed from the model . Based on geology S. Afr. J. Bot. 1999. 65( I )

Table 3 The deviance of the maximal (max), full (full) and minimum adequate (min) logistic regression models (see text) for the five major communities and the fynbos biome ve getation in the Matjiesrivier Nature Reserve (MNR). Significant explanatory variables were chosen by a process of forward selection (see text) . The estimate (est), standard error (se) and t-value (t) for each significant variable in the minimum model are given. Values of t > 2 are approximately significant at the 0.05 level. AFM = Asteraceous Fynbos Matrix, RSF = Restioid Sandy Fynbos, SSK =Shale Succulent Kara , SaSK =Sandy Succulent Karoo, SKM =Succulent Karoo Matrix Model deviance

Vcgdalion max filII min Variables est sc

AFM 150 .19 21 Com' ranI + -2.686 0.596 -1.5

Altitudc2 2.375 0.639 3.7 RSF s--, 41 24 Constant + -3A81 (I.':'/"' 48 Landtypc3 3.48 1 3AR I 4.1

SS K 144 76 70 COI1SIa/1t -2.03' (1.3117 6.6

SKM 128 55 31 Constant + -0.128 0.292 0.-1 Altitude2 -2 .790 0.589 4.7

SaS K 42 19 8 COI1SfaIU + -3.181 0. 7/8 -19 Landtypc3 1.872 0.957 2.0

Fynbos 170 94 66 COI1Sralll + -3. 83 1.111 38 Altiludc2 4.64 1. 04 4.5 alone, the final model had relatively good predictive power. Only Sandy Succulent Karoo (SaSK) 12 of 33 s ites were incorrectly classified (Table 4), A ll the mis­ Landtype and altitude caused the lll ost significant changes in classifications were due to sites that were incorrectl y classified deviance w hen added to the model (Table 3). Because SaSK was as non-SSK. These sites were all on sandstone geology, but all never found above 800 In, altitude was incorporated into the final had some other feature, slI ch as gravel patches, wh ich gave ri se model (eq. 8) as a statem.ent of fact. As with its fynbos equivalent to s imilar environmental conditions to the shale geology. T he on sandy soils (RSF), SaSK was never found on bedrock sheets: model was un able to account for these s ites. The fi nal model for therefore, this landty pe was also treated as a statement of fact in SSK was: the model. The model was a poor predictor of SaSK, with onl y three out of five sites being correctly classified. Five s ites were O n sandstone 1 = -2.037; in correctl y classified as SaSK (Table 4). The final model for on shnle p = I (e'l.6) SaSK was:

Although SSK was the only vegetation type to occur on < 800 m and on rocky slopes or sand y plains I = -3.48 1 + ( 1.872 sha l e~ d e rive d soil (p = I ± 0), it was also predicted to occur occa­ , landtype3); sionally on sandstone-derived soil (p = 0.1 ± 0.05). > 800 m or on bedrock sheets 1 = -00 (eq.8)

SaSK most frequently occurred on low altitude sandy plains (p = Succulent Karoa Malrix (SKM) 0.2 ± 0.07). The same sandy plains support RSF at hi gh altitude. As with as teraceous fynbos matrix, the model for SKM was best The low probability of predicting the occurrence of SaSK is descri bed by alti tude and geology (Table 3). Once again, geology because the low altitude sandy plains also suppo rt SKM in places was incorporated into the final model, (eq. 7) as a statement of and SaSK is occasionally found on low altitude rocky s lopes (p = fact to reduce model instability. In comparison to the A FM 0.1 ± 0.05). In deed, it was difficult to separate SaSk and SKM at model , th e SKM model was a better predictor of community di s~ a coarse scale. At best, the model predicts where SaSK will defi­ tribution (Tnble 4). The model over-estimated the occurrence of nitely not be located. SKM, incorrectly predicting it in 13 si tes, and under-est: mated it in five sites. However, 22 out 27 sites were correctl y classified Fynbos/SlicCulenl karoo boundary (Table 4). The fin al .m odel for SKM was: Altitude and geology were the only significant variables in the model accounting for the di stribution of fynbos. As with the On sandstone I = -0.128 + (-2 .790 * altitude2); other models, geology was treated as a statement of fact in the on shale I = ~ 'X; (eq.7) model, because fynbas was never found on shale. The model (eg. SKM was associated with low altitude, sandstone sites (p = 0.5 ± 9) had relatively good predictive power. Only one of 54 si tes was 0.08). Thus, on sandstone derived soils, fynbos occurs at higher incorrectly classified as non-fynbos (Table 4). The model did, altitudes and while succulent karoo is found at lower altitudes. As however, incorrectly predict the occurrence of fynbos in 15 sites with AFM, the transitional nature of the fynbos/succul ent karoo (Table 4). The final model for fy nbos was: boundary was responsible for the relatively low probability for On sandstone I = -3 .83 + (4.64' altitude2); predicting the occurrence of SKM and the fact that SKM was on shale 1= -00 (eq.9) occasiona lly found on sandstone above 800m (p = 0. 1 ± 0.05). The occurrence of fynbos could thus be predicted with high S. Afr. J. Bot. 1999,65(1) 93

and Succulent Karoo Matrix (SKM). can be attributed to the Table 4 The number of sites correctly and incorrectly transitional natu re of the boundary - an ecotone between the fy n­ predicted by the regression models (see text and Table 3) . bus and succulent karoo biomes in the eastern Cederberg. Here. See Table 3 for explanation of vegetation abbreviations fynbos and succulent karoo elements are intermingled across a Predicted number of sites broad band on sandstone-derived soils where the annllal rainfall varies between 150 and 200 mm (Ledunere-Oertel J 998). Simi­ Incorn.:ct[y Incorrectly lar patterns occur on the sandy forelands of Nama qual and (Cowl­ Vcgt!talion Nil. tlfsites Corn.!Cl absent present ing el (1 /. 1999) and elsewhere in the western part of the fynbos AFM 36 :1 3 3 36 biome (R.M. Cowling, pers. obs.). With in this ecotone, wherever there are locally mesic or xeric sites, the balance shifts to fynbos RSM 15 <) (, <) or succulent karoo, respectively. None of the explanatory varia­ SSM 33 21 12 II bles retained in the models accounted for such fine-scale envi­ ro nmental and associated vegetation heterogeneity. Radiation SKM 27 22 5 13 regime, a function largely o f slope aspect and inclination. was SaSK 5 3 0 5 expected to account in part for this variation, but was not signifi­ cant in any of the models. The reason for this is unknown. espe­ Fynbos 54 53 15 cially since Holland and Steyn (1975) found marked vegetation differences on north and south aspects in the fynbos biome. accu racy on the basis of two easi Iy meas urable environmental which they attri buted to radiation loads. Simil a rl y, Levyns variables - altitude and geology. Generally. fynbos biome vege­ ( 1950) attrib uted the boundary between succulent karoo and fyn­ tation in the study area was associated with sandstone geology at bos vegetation in the Little Karoo to differences in slope aspect. high altitude (p = 0.7 ± 0.08). when geology was held constant. Thus, fy nbos biome communi­ ties were associated with cooler, south-facing aspects. Critical evaluation of the models RSF and SaSK were poorly predicted, probably becau se these The residuals frolll all the models were nO( rand om and the error communities were modell ed largely on th e basis of landtype3 variance was not constant. The models for fynbos and all COI11- (flat sandy plains). However, both A FM and SKM were occa­ munities. except SSK, overestimated the occurrence of the vege­ sionally found on sandy plains. The classification of landtypes tation type. Although these features were problematic for making into mutually exclusiv e categories is problematic. For example, accurate pred ictions and extrapolations, we be li eve [hat the mod­ this classification did not account for depth of sand above the els were adequate for the purpose they were designed to fulfil l, bedrock, a factor that is crucial in differentiating between res­ especial ly considering the coarse scale of the explanatory varia­ tioid and asteraceous fy nbos types on the arid margin o f the tyn­ bles used in this study . In all cases, the models were biologically bos biome (Campbell 1986). Nontheless. considering the original requirements of the GLMs, we felt the cu rrent models were suffi­ meaningful and gave an estimate of the probflbility with which cient for making biologically relevant predictions of community the occurrence of a community or the fynbos/succulent karoo boundary could be predicted. The models co uld be improved by distributions, based on three very crude environmental variables. including further explanatory vari ables, but this was not desira­ Determinants of the fynbos/succulent karoo boundary ble in terms of the initial model requirements of a few easily measured variables. The occurrence of fynbos in a site could be predicted with a probability of 0.7 on the basis of two very crude environmental A comparison of the predicted and actual vegetation maps variables: altitude (> 800 In or < 800 m), and geology (sha le or sand stone). Both of th ese variabl es indi rectly cont rol moisture There was good overall correspond ence between the predicted availability to plants. In the mountains of the fynbos biome, rain­ and actual vegetati on maps (Figure I ). The occurrence of SSK fall increases with increasing altitude, which means that more and RSF was accurately predicted. The model was unable to moisture enters the soil at higher than lower altitudes (Campbell account for the transitional nature of the AFM- SKM boundary, 19 83). Furthermore, temperature decreases with increasing alti­ and predicted that AFM, and hence fynbos biome vegetation, tude, resulting in lower potential evaporation at higher altitude would occur throughout this ecotonal area. si tes. Soil texture also controls moistu re availability to pl ants. Shale-derived soils, which are fine-textured, are effectively more Discussion arid than coarse-textured, sandstone-derived soil (Lech­ Predicting the distribution of communities mere-Oertel 1998). However, shale-derived soils are more fertile The aim of this study was to model the environmental correlates than those derived from sandstone, both in the study area {Lech­ of community distribution in the MNR and to generate a predic­ mere-Oertel 1998) and elsewhere in the fynbos biome (Campbel l tive vegetation map derived from these models. The results of 1983; Cowling & Holmes 1992). What the n are the relative roles the GLMs showed that the occurrence of communities could be of soil moisture and ferti lity in explaining the fynbos/succulent predicted with probabilities between I and 0.2. Weak signifi­ karoo biome boundary in th e study area? cance, as we discuss below, may arise for a number of reasons, Of great relevance here is that succulent karoo - both SKM including the requi rement for more or better explanatory varia­ and SaSK - were associated with edaphic habitats (sandstone bles (Brown 1994). However, ou r intention was to li se explana­ bedrock and sand plains, respectively) that supported fynbos veg­ tory variables that did not requi re tedious data collection, thereby etation at higher altitudes. This suggests that moisture rather than reducing the effort required to generate a set ofpedictions. None­ nutrient availability controls the boundary between the two theless, even weakly significant community-environment models biomes (Ellery el at. 199 I; Cowling et u/. 1997; Milton e/ (1/. do warrant interpretation, since the unexplained variance may 1997). Furthermore, the intermingling of fynbos and succulent provide pointers for other determinants of pattern not consid ered karoo elements in a broad ecotonal band 0 11 in fertile sand­ in the model (Brown 1994). stone-derived soils confirms the role of moisture. The rapid tran­ The low probability of predicting the occurrence of the two sition from fy nbos to succulent karoo on shale-derived soils (see most extensive communities, Asteraceous Fynbos Matrix (AFM) al so Bond 1981; Ellis & Lambrechts 1986; Campbell 1986) is S. Afr. J. Bol. 1999. 65( I)

! Actual Vegetation Map

I E!I Shale Succulent Kareo :::::g Asteraceous Fynbos Matrix _ Succulent Kamo Matrix _ Restioid Sandy Fynbos ~ Sandy Succulent Kareo ~ Fynbos I Succulent Kamo transition

Predicted Vegetation Map

Figure I A comparison of a vegetation map derived from ground survey (Lechmere-Oertel 1998) with the predicted vegetation map based on the generalised linear models for community dist ributions. S. Afr. J. Bot. 1999,65(1) 95

probably a response to changes in soil moisture availability survey design: gradsect sampling. In: Nature Conservation: Cost rather than nutrient status. The performance and survival of seed­ EITective Biological Survcys and Data Analysis. eLls. C.R. MilrguJes lings of fynbos and succulent karoo species growing in the study & M.P. Austin. CSIRO. Canberra. area, indicate that fy nbos species are unable to to lerate the drier AUSTIN, M.P., NICHOLLS. A.O. & MARGULES, CR. 1990 Meas­ conditions at lower altitude sites 011 sandstone-derived soi Is urcment of the realised qualitative niche of plants species: examples (LechmereOerteI 1998), At wetter sites, succulent karoa species orlhe environmental niches of live EUCa~l'pIUS spt!c ics. Ecol. .\4ollogr. are probably outcompeted by fy nbos plants. 60: 16 t- t 77. Although the model could not predict the locati on and extent BOND. W.J. 1981. Vegetation GraLliems in the SOlilhern Cape Muun· tains. Unpublished M.Sc. Thesis. University of Cape Town. of the fynbos/succulent karoa transition, it did predict climatic and edaph ic limits beyond wh ich fynbos was never found. Of BROWN, D.G. 1994. Predicting vegetation types at treeline using topography and biophysical disturbance variables. J. '.1::[:. SCI 5: 641 - particular interest for climate change monitoring is the topocli­ 656. matic boundary of 800 m altitude. Should there be an increase in CAMPBELL, B.M. 1983. Montane plant environments in the 1)!I1bns aridity or temperature that corresponds to a shift in altitude, then biome. Bo(hafia 14: 283- 298. the lower li mit of fynbos shoul d shi ft to a higher altitude. CAM PDELL, 13.M. 1986. Montane plant communiti es of the t)."n hos Euston-Brown 's ( 1995) GLMs suggested that a warmer, drier biomc. Vegelatio 66: 3- 16. climate would restrict fynbos components to south-racing slopes CCWR 1996. Computing Centn.: fo r Water Research. Un iversity o f and high altitudes in the Kouga-Baviaanskloof mountain com­ Natal, Pietcrmaritzburg. KwaZulu-Natal. plex in the Eastern Cape. COWLING. R.M .. ESLER. K.J . & RUNDEL. P.W . 1999. Namayua­ land, South Africa - a uniqu(! winter-rainlall desert ecosystem. Planl Applications of predictive mapping Ecology (in press). There are many management appl ications for vegetation map­ COWLING, R.M. & HOLM ES. P.M. 1992. Flora and Vcgt!tation. In : ping using predicti ve models and GIS (e.g. N icholls 1991; Good­ The Ecology of Fynbos. Nutrients. Fire and Oivt!rsity. ed. R.M. eO\\ l­ child 1994; Franklin 1995; Valverd e & Montana 1996). In many ing. Oxford Un iver.;i ly Press. Cape Town. cases the costs of vegetation surveys, in terms of money and COW LING. R.M.. RtCHARDSON. O.M. & MIJSTART. P.J . t<)97. manpower, will prevent inventories of regions ever being com­ I'ynbos. In: The Vegetation of Southern Africa. eds. R.M. Cowling, pleted, or even undertaken, before management decisions need to [),M. Richardson & S.M. Pil!rcc. Cambridge University Press, Cam­ bridge. be made (Nicholls 1989). Using GLMs in a GIS is an important CRAWLEY. MJ. 1993. Ml!thods in Ecolog): GUM for Ecologists. first step to aiding decision making. Furthermore, th e flexibility Blackwell Scientific Puhl. of predictive models within a GIS means that they can be ELLERY. W.N., SCHOLES. IU. & MENTIS. M.T. 1991. An initial adj usted easily shoul d the environment change or the vegetation­ approach to predicting the sensitivity of the South A lrican grassland envi ronment algorithm be improved. biome to cl imatc change. S. Afr. J. Sci. 89: 449- 503. The fact that the fy nboslsucculent karoo boundary and occur­ ELLIS. F. & LAMBRECHTS. J.J .N. 1986. Sui Is. In : Tht! Karon Biome: rence or the major communities in MNR can be predicted within a prcliminary synthesis. ~ds. R.M . Cowling. P.W. Raux & AJ.II. P i~ ­ a GIS environment has implicati ons for management and moni­ terse. Part I - thc physical environment. South African National Sci­ toring projects. In the case of MNR, the neighboring properties entific Programmes Report No. 124. CSI R. Pretoria. that have joined the Cederberg conservancy can be mapped with EllSTON-BROWN, D. 1995. Environmental and dynamic dl!tl!rminants relative ease and at low cost, based on a few easi Iy measured ofvegelation distributi on in the Kuuga and Ba"HlanskloofMountains. environmental variables. Such maps can then be used for man­ Eastern Cape. Unpublished M.Sc. Thesis. University ofCapcTnwn. agement and conservation, and may form the basis for more FRANKLIN, J. 1995. Predictive veg.(!tat iol1 mapping: geographic mod­ intensive mapping (Nicholls 1989). For example, an extrapolated elling of bi ospatial patterns in relation 10 environmental g.radients. map for the conservancy area can be used to calcul ate the areas frog. Phys Geog. 19: 474-499. of different vegetation types fa ll ing inside and outside th e GENSTAT 5 COMMITTEE. t987. GENSTAT 5 Reference Manual. reserve system, and this information can be used to formulate Clan!ndon Press. Oxford. priorities on th e basis of representation goals for conservation GILLISON, A.N. & BREWER. K.R.W. 1985 . The lISI! of gradient (Presley of al. 1993). cJirl!ctcd transects or grad sects in natural resource surveys. J. Ellv. Manage. 20: 103-127. Acknowledgments GOODCHILD, M.F. 1994. Integrating GIS and remotc sensing fix veg­ etation and modelling: methoLiological issues. J. Veg. Sci. 5: 615-f126. Thi s research was supported by fu nds from th e Foundation HOFFMAN, T. 1996. Lowland Succulent Kama. In : Vl.!gelation o f Research and Development, Wo rldwide Fund for Nature: South South Africa, Lesotho and Swazil and. I.!ds. A.D. Low & A.G. Rebelo. Africa and Western Cape Nature Conservation. We acknowledge Department o f Environmental Affairs and Tourism, Pretoria. the use of a 4 x4 vehicle, courtesy of th e Mazda Wildlife Fund. l·tOLLAND. P.G. & STEYN. D.G. 1975. Vegetation responses to latitu­ Thanks to June Juritz for assistance with the generalised linear dinal variations in slope angle and aspect..J. Biogeog. 2: 179- 183. models, and Dave Richardson for comments on an earlier ver­ LECHMERE-OERTEL R.G. 1998 . The environmental deternlinants of sion of this manuscript. the (ynbos/succulent karoo blHmdary Ln the Matjiesrivier Nature Reserve. Western Cape. Unpubl ished M.Sc. thesis. University of References Cape Town. AUSTI N. M.P. 1991. Vegetation theory in relation to cost-efficient sm· LEVYNS. M. R. 1950. The relations of the Cape and Karoo noms near veys. In: Nature Conservation: Cost Effective l3iological Surveys and Ladismith. Cape. Trans. Roy. Soc. S. Afr. 32: 235- 246. Data Analysis. cds. C.R. Margules & M.P. Austin. CS IRO, Canberra. LlNDSEY,1.K. 1989. The Analysis of Categorical Data using GLiM. AUSTIN. M.P .. CUNNINGHAM. R.B. & FLEMING. P.M. 1984. New Springl.!rVeriag, New York. approaches to direct gradient analysis using environmental scalars and LOW, A.B. & REBELO. A.G. 1996. Vegetation of South Ali·lca. statistical curve fitting procedures. Vegetatio 55: 11 - 27. Lesotho and Swaziland. Dl!partment of Environmental Affairs imd AUSTIN. M.P .. CUNNINGHAM. R.B . & GOOD. R. 1983. Altitudinal Tourism, Pretoria. di stribution of several £uca~vplus species in Southern New South MARGULES. C.R. & AUSTIN. M. P 199 1. Nature Conservation: Cost Wa les.AusJ. J.Ecol. 8: 169- 180. Effcclive Biological Surveys and Data Analysis. CSIRO. Canherra. AUSTIN. M.Il. & HEYLIGERS. P.C. 1991. New approach to vcgetation MARGULES. CR .. NICHOLLS. A.O. & AUSTIN. M.P t987. Di ver- 96 S. Afr. J. Bot. 1999, 65( I)

s ily of ElI(: a~\lp llls species predicted by a multi -variable environmen­ REBELO, A.G. 1996a. Ccntral Mou ntain Renosterveld. In : Vegetation lu i gradient. OecoJogta 7 1: 229- 232. o f South Africa. Lesotho and Swaziland. eds. A.B. Low & A.G. McCULLAGH, P. & NELDER. .I .A. 1989. Generalized linear Models Rebelo. Department of Environmenwi AfTairs and Tourism, Pretoria. (2 Edition). Chapman and Hall . London. REBELO, A.G. 1996b. Mou ntain Fynbos. In: Vegetation of South MILTON. S,J .. YEATON. R.I .. DEAN. W.lU . & VLOK. J.H,J. 1997. Afri ca, Lesotho and Swaziland. eds. A.B. Low & A.G. Rebelo. Succulent Karoa. In : Vegetati on of Southern Africa. (!ds. R.M. Cowl­ Department of Envirormlental Affairs and Tourism, Pretoria. ing, O.M. Rkhurdson & S.M. Pierce. Cambridge Un iversi ty Press, SCHULZE, R.E. 1975. Incoming radiation fluxes on sloping terrain: a Cambri dge. general model for usc in South A lricll. Agrochemophysika 7: 55-61. NICHOLLS. A.O. 1989. How to make biological surveys go further TREXLER, J.C. & TRA VIS, 1.1993. Non-traditional Regression Analy­ with gCIll!raliscd lincar models. Bio!. Cons. 50: 5 1-75. ,e,. Ecologv 74 : 1629-1637. NI CHOLLS . 1\.0. 1991. Examples of the use of generalised linear !nOt1· VALVERD E, P.L. & MONTANA, c. 1996. A rapid methodology for els in analysis of survey data for conservation evaluation. In : Nature vegetation survey in Mexican arid lands. J. Arid Environ. 34: 89- 100. ('om:crvation: Cnst EtTcctive Biological Surveys and Data Analysis. VAN DE RUT, C.W.C,J., HAZELHOFF, L. & BLOM, C. W.P.M.1996. C.R. M;'lrgult:s & M.P. Austin (Eds). CSIRO, Australia. Vegetation zonation in a fomler tidal area: a vegetation-lype response NIE LSON. R.P. 1993 . Transient ecotone response to cl imatic change: model based on DCA and logistic regression using GIS. J. Veg Sci. 7: some l: OIu:t:ptuaJ and modcll ing approaches. £('01. App. 3: 385-395 . 505- 518.