1Several hypotheses have been proposed to explain the structure of multi-species assemblages. Among these, abiotic environmental factors and biotic processes are often favoured. Several recent studies examining anuran communities identified environmental factors to be only of minor importance in the composition of leaf-litter and canopy assemblages in pristine forests. Instead, spatial effects and spatially structured environments were considered more important.
2In this study, we investigated whether these findings could also be confirmed for very heterogeneous stream habitats in the primary rainforest of the Ulu Temburong National Park, Brunei Darussalam. We thus investigated anuran assemblage compositions on 50 stream sites with regard to environmental and spatial influences.
3Cross-product correlations indicated that both factors (spatial and environmental parameters) determined assemblage composition of anurans. Environment itself may be spatially structured, yet this interrelation did not contribute to the explainable variation of frog community compositions within the study area.
4Detailed analyses of the environmental parameters with nonmetric multidimensional scaling revealed that community structure was mostly affected by three major environmental characters: stream turbidity, river size and the density of understorey vegetation. Based on these habitat characteristics, we assigned species to three distinct habitat guilds.
5The results underline the importance of riparian habitat heterogeneity in pristine forests in structuring anuran assemblages. We conclude that different anuran assemblages, that is, leaf litter, canopy and stream communities, follow different assemblage rules and thus are not directly comparable.
Spatial effects may result from biotic processes such as dispersal strategies and abilities (Clements 1916; Woodward 1983; McCarthy 1997; McCarthy & Lindenmayer 2000). Hence, habitats may contain similar species because they are in close proximity. Ignoring space may lead to false ecological conclusions as the environment itself may be spatially structured (Borcard et al. 1992), for example, mites being dependent on humidity, which in turn is a function of distance to water (Borcard & Legendre 1994). Detection and analysis of spatial autocorrelation is a new paradigm in ecology (Lichstein et al. 2002). Separating these effects and their relative contribution to structuring species assemblages thus is an ecological and methodological challenge.
Studies that examined amphibian assemblages concerning spatial, environmental and spatially structured environmental effects yielded contrasting results. On regional scales Parris (2004), detected major environmental and minor spatial effects on anuran species composition. However, a significant proportion of variance was explained by spatially structured environment. Comparative studies of anuran assemblages in pristine and disturbed forests of the afro- and neotropics revealed that in both states, spatial structure is of major importance, whereas pure environmental effects were governing assemblages only in degraded habitats (Ernst & Rödel 2006, 2008; Hillers et al. in press). Within pristine forests, predictions of assemblage compositions on a species-specific level were not possible with habitat parameters (Ernst & Rödel 2005). This led the authors to conclude that in pristine forests, priority effects and lottery recruitment were of greater importance than species-specific responses to the environment, although species have, for example, specific breeding habitats. The environment is not a limiting factor as the matrix is freely crossable and specific habitats are regularly available (Ernst, Linsenmair & Rödel 2006; Ernst & Rödel 2008). Menin et al. (2007) support this view as terrestrially breeding frogs showed little beta diversity explainable by habitat variation.
Inger & Colwell (1977) suggested that climatically unpredictable environments tend to prevent formation of distinct guilds, while predictable environments support guild formation and result in higher species richness. In this study, we expand this by examining if environmental predictability determines community structure. Most aforementioned studies described relatively homogeneous landscapes with environmental differences only detectable on very large or very small scales (geographical regions and microhabitats, respectively). Herein we hence predict that environmental factors will be important predictors of local assemblages if the landscape is very heterogeneous.
Vasudevan, Kumar & Chellam (2006) argue that in hilly terrains, amphibian diversity varies largely between streams due to very diverse habitats. In the hilly heart of Borneo, a patchwork of riparian habitats occurs (Inger, Voris & Frogner 1986; Inger & Voris 1993; Grafe & Keller in press). Thus, investigating mountainous streams in Borneo's pristine forests might be promising to understand the forces structuring riparian anuran guilds. Investigation of riparian anurans as a separate entity is also in agreement with Ernst & Rödel (2008), who call for separate investigation of functional groups.
In this study, we investigated, whether strong environmental heterogeneity results in distinctly different anuran assemblages. We especially addressed the following questions: (i) Do environmentally similar sites bear similar species assemblages? (ii) Are proximate sites inhabited by more similar assemblages than distant ones? (iii) Are environmental parameters related to spatial distance (and vice versa) and are thus spatially structured? Finally, we investigated environmental influences on species and assigned species to guilds.
Materials and methods
This study took place within the Ulu Temburong National Park (UTNP; Brunei; 115°09′ E, 4°33′ N; described in Earl of Cranbrook 1993; Grafe & Keller in press). We conducted plot sampling from May–October 2006. Mean air temperature during fieldwork (19·00–03·00 h) was 24·2 °C (± 0·9 °C SD). Mean water temperature of streams was 24·1 °C (± 0·4 °C SD). Rainfall is slightly seasonal with drier periods from January to March and August to September. Precipitation during the study period was high, with irregularly alternating periods of low and heavy rainfall.
We selected ten streams between 0·8 m and 3·5 m mean widths on almost equal elevations (50–150 m a.s.l.). They represented a cross-section of streams with a diverse set of environmental riparian characteristics. We paired streams such that two streams were comparable in connectivity to the overall stream system. Streams of one pair were feeders of the same larger stream (see Fig. 1). Within each stream, we established five plots of 5 × 10 m with the longer sides parallel to the centred stream. Plot locations were not selected systematically (e.g. regular distance between plots), but rather at random with irregular intervals to cover the heterogeneity (for descriptions see Supplementary Table S1). Average distance between neighbouring plots was 86 m (range: 20–200 m).
We visited each plot eight times with 4–18 days between visits. Within plots, we intensively searched for frogs visually and acoustically so that even well-camouflaged species were detected with reasonable certainty. Except for very fragile species (Leptobrachella mjobergi, Leptobrachella parva and Philautus tectus; for author of scientific names see checklists of Das 2007 and Grafe & Keller in press), we marked frogs individually by toe clipping. We ablated maximally one tip of a toe for each leg excluding thumbs and toes necessary for moulting (i.e. longest toe). Recaptures were excluded from analyses. For the three unmarked species, all encounters were used. Species abundances of all sampling occasions were summed for each site and square root transformed. Mean time spent sampling was 20 min (± 3 min SD).
We characterized plots by various habitat parameters. Density of vegetation was measured in three strata (1–10 cm, 11–100 cm and 101–200 cm) by counting all individual plants with their apex in a stratum. Above 200 cm height, branches usually extended into plots. We determined canopy openness (Lemon 1956) with digital pictures of a spherical Densiometer Model-A (Forest Densiometers by R. E. Lemon, Bartlesville, OK, USA), which was a gridded convex mirror reflecting a 180° view of the canopy.
To determine physical characteristics of the soil, we ranked three ground types (solid rocks, gravel and leaf litter) for dominance of both stream and environment. Based on these, we calculated a factor (rockiness) that was used for analysis (see Supplementary Table S2). To describe the surface area, we decreasingly ranked by eye flowing stream, pools with stagnant water, bare soil or rock, ground covered with vegetation and ground covered with fallen logs for dominance of ground cover.
All distinct altitude differences (≥ 20° slope change and ≥ 5-cm height difference) within a plot were treated as individual gradients. Areas between gradients were considered as flat areas. All gradients were measured in length and height to the nearest 1 cm. Flat areas were only measured in lengths. We used two measures in evaluating a stream's slope (Supplementary Fig. S1): (i) The height of the gradient with the strongest altitudinal difference was the representative for the strength of waterfalls. (ii) The general topography was determined by calculating the mean of heights of all gradients including flat areas (heights set as 0 cm). This mean was weighted by the gradient's lengths. Thus, heights contributed to the mean with respect to their horizontal expanses.
We recorded parameters fluctuating due to varying climatic situations during each visit. The stream water volume was calculated by multiplying a stream's average width (mean of minimum and maximum width) with its maximum depth. Within each plot, we measured water currents with a JDC Flowatch flow meter (accuracy of 0·1 m s−1; JDC Electronic SA, Waadt, Switzerland) in nine equally sized grid cells. For each visit, all cells were averaged (excluding cells without water). For volume and water current, we calculated means and standard deviations over all visits as plot parameters.
We checked for cross-product correlations between distance matrices (species: Bray–Curtis distance, environmental parameters: Euclidean distance and spatial distance: aerial distance) with Mantel tests (R package vegan; Oksanen et al. 2006; R Development Core Team 2006). Distances based on stream courses (as corridors) were not used, as they explained less variance than the aerial distances (data not shown here). For the aerial distance, altitudinal differences were disregarded. We also applied partial Mantel tests for spatial and environmental correlation with species composition, which remove the effect of the respective other. Mantel tests have been criticized in their analytical power (Cliff & Ord 1973; Thioulouse, Chessel & Champely 1995). Hence, we additionally tested environmental parameters for spatial autocorrelation with Moran's I for all plots and within streams (r package ade4; Thioulouse et al. 1997). P values base on randomization tests with 1000 permutations.
When studying assemblages complexity due to multidimensionality is a major problem. Each species is one dimension in which plots may be ordered. Assemblages with more than three or four dimensions loose interpretability. Ordination techniques reduce this multidimensionality (see McCune, Grace & Urban 2002). We herein decided to apply nonmetric multidimensional scaling with the Bray–Curtis index as distance measurement (NMDS; r package vegan; Venables & Ripley 2002; Oksanen et al. 2006). This method is well suited for ecological data, since it allows non-normal distributed and ranked variables (Kruskal 1964; Kenkel & Orloci 1986; McCune et al. 2002). We accepted three axes at a maximum instability of 0·0001 with medium stress (Supplementary Table S3; pcord, MjM Software Design, Gleneden Beach, OR, USA; Monte Carlo simulations with 1000 iterations,). With Mantel tests, we estimated the partition of community variation explained with each axis by correlation of Euclidean distance matrices of the respective axis with the community matrix. Finally, we correlated environmental vectors onto these ordination axes (r package vegan; Venables & Ripley 2002). According to the species’ projected coordinates within the ordination, we clustered species into guilds.
We encountered 27 out of 66 amphibian species (Grafe & Keller in press) of five families (species): Bufonidae (6), Megophryidae (6), Microhylidae (1), Ranidae (12) and Rhacophoridae (2). Species accumulation curves for streams flattened after three to six visits. We assume all are riparian breeders (except Philautus tectus) since we encountered tadpoles or mating individuals within streams or species are reported to use streams as oviposition sites (Inger et al. 1986; Manthey & Grossmann 1997; Malkmus et al. 2002). P. tectus is a direct developer, but was exclusively found at streams and not in the surrounding area. Thus, it might be dependent on streams for reproduction as is true for the congener Philautus hosii (Inger & Stuebing 2005).
space, environment and spatially structured environment
The results of the Mantel tests indicated, that species composition was influenced by both environmental (r = 0·3034, P < 0·001) and spatial distance (r = 0·2087, P < 0·01). No spatial autocorrelation of environmental structure was detectable (r = 0·039, P > 0·1). With partial Mantel tests, explanatory power of the respective matrices hardly differed (environmental: r = 0·2953, P < 0·001; spatial: r = 0·1961, P < 0·01), confirming the lack of spatial autocorrelation. Moran's I indicated spatial autocorrelation for only two characters (Supplementary Table S4), out of 170 combinations of streams and parameters (stream 1/water current: P < 0·05, I = –0·98; stream 10/area occupied by vegetation: P < 0·05, I = –0·88). All other combinations were not significant (all P > 0·05) and spatial distance did not explain much of environmental variance (mean I =–0·28 ± 0·13 SD). Over all streams, no parameter was autocorrelated (all P > 0·1, mean I = 0·05 ± 0·04 SD).
The ordination of plots in species space with three NMDS axes cumulatively explained 80·2% of the variance of species composition (Mantel tests; NMDS 1: r2 = 0·243, P < 0·001; NMDS 2: r2 = 0·478, P < 0·05; NMDS 3: r2 = 0·081, P > 0·05). For five environmental parameters, we found strong correlations (Table 1 and Fig. 2). Stream volume and area of stagnant waters correlated positively with NMDS 1. This indicated preferences for habitats in a transition from small streams to large water bodies. Species positively correlating with NMDS 1 (and thus with stream size) were Ansonia leptopus, Bufo juxtasper, Chaperina fusca, Limnonectes leporinus, Meristogenys jerboa, M. poecilus, Pedostibes hosii, Rana picturata, R. signata and Staurois latopalmatus (Table 2). The remaining species showed a clear negative tendency or at least no strong positive correlation (e.g. Rhacophorus belalongensis).
Table 1. Correlation of environmental parameters with the anuran community ordination and ordination projection for each axis. Estimations of P values base on 1000 permutations and are listed as ***P < 0·001, **P < 0·01 and *P < 0·05
Table 2. Correlation of species with ordination of anuran community structure and vectors of ordination projections. P values base on 1000 permutations and are listed as ***P < 0·001, **P < 0·01 and *P < 0·05
Morphological and behavioural features supporting guild affiliation are designated as†large clutches (> 500 eggs),‡unpalatable tadpoles,§ultrasound calls,¶tadpoles with well-developed sucker organs and††foot-flagging and‡‡spawn glued to rocks. Families: BU = Bufonidae; ME = Megophryidae; MI = Microhylidae; RA = Ranidae; and RH = Rhacophoridae.
Ordination axis NMDS 2 was correlated with the two slope measurements, hence, indicating similar importance of strong waterfalls and general steepness (Fig. 2). Thus, NMDS 2 represented a transition of stream dynamics from stagnant or moderately flowing streams to turbulent streams with strong slopes. Species with positive association were Ansonia albomaculata, A. longidigita, A. platysoma, Leptobrachium montanum, Limnonectes kuhlii, Megophrys nasuta, Rhacophorus belalongensis, Staurois guttatus and S. tuberilinguis (Table 2). Species negatively correlating with NMDS 2 were Ingerana baluensis, Ingerana sp., Leptobrachella parva, L. mjobergi, Leptolalax dringi, L. gracilis, Philautus tectus and Rana raniceps (Table 2).
Ground level vegetation correlated with NMDS 1 and NMDS 2 (Table 1), but was an inseparable covariate with low correlation coefficient.
Based on the high explanatory power of NMDS 1 and NMDS 2, we clustered species into three specific guilds (Table 2 and Figs 2 and 3). Species with strong positive correlation with NMDS 1 were clustered as ‘large stream species’ (LSS). The second and third group both had negative association with NMDS 1, but were antithetic in their correlations with NMDS 2. Species with positive correlations were grouped as ‘waterfall stream species’ (WSS). Species avoiding such habitats, apparent by negative association, were termed ‘calm stream species’ (CSS).
However, a by far larger part of variation in assemblage composition was explained by the environment, that is, similar species inhabiting similar sites. This is in contrast to arboreal and leaf-litter assemblages investigated by Ernst & Rödel (2006, 2008) and Menin et al. (2007). In the first study, predictability by environment was only detected in habitats altered by humans, whereas in the second, space exclusively determined species compositions. Menin et al. (2007) identified terrestrial breeding amphibians to be mostly habitat generalists, independent of environmental parameters. As in our study, frog and fish assemblages of mountainous streams in Queensland, Australia were mostly explainable by the environment (Pusey, Arthington & Read 1995; Parris 2004). This indicates that stream habitats within a pristine rainforest may generally be heterogeneous enough to enforce differentiation of assemblages by environmental parameters, whereas those of the forest leaf litter and canopies may not. We thus believe that anuran assemblage patterns within these three types of habitats, all within pristine forests, are subject to different structuring forces.
Partial Mantel tests as well as Moran's I yielded that the environment was to almost no amount spatially structured, and thus spatially structured environment did not play a relevant role in explaining species composition. The habitats were more patchily distributed than forming a continuous gradient. It is likely that the hilly topography created such an almost random mosaic of habitat patches that consequently contributed to high overall anuran species richness and species turnover between sites (Grafe & Keller in press). Additional evidence for this interpretation comes from studies that associate high species diversities with diverse riparian habitats (e.g. Pusey et al. 1995; Vasudevan et al. 2006).
Ordination of the anuran assemblages in the UTNP revealed that species composition altered almost completely with changes in three characteristics of streams and their environment: density of understorey vegetation, stream size and presence of waterfalls. Understorey vegetation is generally assumed to be an important environmental characteristic for riparian frogs (Parris & McCarthy 1999). Since several of the UTNP species use lower vegetation as calling or resting sites (e.g. A. albomaculata, A. longidigita, L. mjobergi, Staurois spp.), it is an important structural component. Parris & McCarthy (1999) identified importance of palms, because they indicate mesic conditions of the understorey. In our study, understorey vegetation was a better indicator than rare palms, but presumably for the same reason, that is, a highly humid microhabitat.
Stream size was the second factor structuring anuran assemblages. This follows other studies, which show that larger streams differ in their assemblages from smaller ones (Inger & Voris 1993; Parris & McCarthy 1999; Eterovick & Barata 2006). Principal advantages of larger streams may be increased incidence of light and thus increased growth of algae to forage and higher water temperature for faster larval development, at least in shallow edges (McDiarmid & Altig 1999). Relative constancy and predictability of climatic fluctuations of large water bodies may also be of importance (Parris 2004).
The third and major environmental factor was stream turbidity (presence of waterfalls and general habitat steepness) at smaller streams. In contrast to extensive waterfalls and cascades at small streams, large streams in the study area were slower flowing and hardly interrupted by waterfalls, wherefore differences in stream turbidity were only of minor relevance. Inger & Voris (1993) also identified slopes as major determinants in structuring anuran species assemblages at streams. Most species encountered at waterfalls show morphological and/or behavioural adaptations to a torrential, noisy environment. For example, in the vicinity of waterfalls, Staurois species use visual signals to communicate (Hödl & Amezquita 2001; Grafe & Wanger 2007) and Meristogenys males likely call using ultrasound like Huia cavitympanum (Arch, Grafe & Narins 2007). Reciprocal influences on communities of adult and larval stages have been documented (Inger et al. 1986), whereby morphological features of tadpoles may determine the presence of adult frogs e.g. A. albomaculata tadpoles have abdominal suckers to adhere to rocky surfaces (Malkmus et al. 2002). However, many species showed clear negative associations to slope. These lack special adaptations to the torrent and noisy environments and they are less likely adapted to the predators of large streams (e.g. possessing small clutches and sizes). Consequently, they are dependent on small stagnant water.
Based on the two most important environmental characters (stream size and turbidity) alone, we were able to define three discrete guilds, namely ‘waterfall stream species’, ‘calm stream species’ and ‘large stream species’. Known behavioural and morphological characters of the larval as well as the adult stages of the various species back up this guild assignment. The pronounced heterogeneity of the differing stream environments lead to differing habitat-specific adaptations. These habitat differences are strong enough to result in distinct re-occurring assemblages at specific habitats almost independently of spatial distance. In comparison with other anuran studies (Inger & Voris 1993; Parris & McCarthy 1999; Parris 2004; Ernst & Rödel 2005, 2006, 2008; Eterovick & Barata 2006; Menin et al. 2007; Hillers et al. in press), this leads to the conclusion that riparian habitats in mountainous pristine rainforests may be generally more heterogeneous than the leaf litter and canopy habitats. This heterogeneity seems to result in environmental predictability and this in turn to predictability of anuran species compositions by habitat type (Inger & Colwell 1977).
In summary, the biotic control model (spatial effects) was not able to explain the full variation of anuran community composition in the UTNP alone nor was the individual concept (environmental determinants). Both factors had significant influences, yet questions about spatial limitations remain open. Thus, riparian anurans of the UTNP are fascinating objects for further studies of assemblage rules. Furthermore, keeping the current dramatic decline of this group of vertebrates in mind (Blaustein, Wake & Sousa 1994; Stuart et al. 2004; Pounds et al. 2006), this investigation provides a baseline for future comparisons with anthropogenically degraded areas in Borneo. It is not yet clear how loss or alteration of habitats affects riparian anurans on the island, but it is obvious that it will dramatically unhinge ecosystem processes as described for other habitats and regions (Gillespie et al. 2005; Ernst et al. 2006; Gardner et al. 2007b).
We gratefully acknowledge the support of the Kuala Belalong Field Studies Centre (KBFSC) staff, its director Hjh Kamariah Binti Abu Salim and supervisors Md. Salleh Abd. Bat and Rodzay Bin Abd. Wahab. Furthermore, we thank Jan Beck, Thiemo Braasch and Matthias Siegle for assistance in the field. A.K. appreciates support of the German Academic Exchange Service (DAAD). T.U.G. received financial support from a University of Brunei Darussalam Research Grant (UBD/PNC2/2/RG/1(58)).