SEARCH

SEARCH BY CITATION

Keywords:

  • Bayesian information criteria;
  • Floristic composition;
  • ISOMAP ;
  • Model-based clustering; Monocot herbs;
  • NMDS ;
  • Tropical forest;
  • Tropical soils;
  • UPGMA ;
  • Vegetation classification;
  • Zingiberales

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Questions

Are topography-based forest types floristically consistent between sites in central Amazonia? Do broad landform and geological features control site-specific edaphic and floristic variation and therefore obfuscate the floristic classification based on local topographical classes? Is model-based clustering a useful tool for floristic classification?

Location

Non-inundated forest of central Amazonia, north of the Amazon River.

Methods

We analysed species presence–absence of a group of terrestrial monocot herbs (Zingiberales) in 123 plots (250 × 2 m) concentrated in three sites of non-inundated forests. Distances between plots were 1–140 km. Floristic patterns were extracted by dimensionality reduction using geodesic floristic distance. We applied a model-based cluster analysis (MC) coupled with the Bayesian information criterion to determine the best floristic classification. We used geometric and non-geometric internal evaluators to compare the performance of MC to the agglomerative hierarchical clustering method UPGMA. The floristic clusters were tested for differences in edaphic and topographic features. Landform-geological classes were defined based on geological maps and a digital elevation model. We used the Kappa index and ANOVA to evaluate the agreement between landform–geological classes, floristic clusters and environmental features.

Results

The best MC solution found four floristic clusters. Differences in soil chemical properties, which were linked with lithological classes and broad land-form features, explained abrupt floristic changes and floristic differences between the same topographical habitats of different sites. Within poor soils, floristic classes defined by elevation along the soil catena (upland and valley forests) were fuzzy. Valley sandy forest was not floristically consistent across sites due to subtle edaphic variation. Using a non-geometric internal evaluator, MC coupled with geodesic floristic distance estimation performed better overall than UPGMA.

Main conclusions

Geological classes defined by lithology and broad landform features control the major variation of edaphic and floristic patterns in central Amazonia. MC proved to be a useful method to classify and interpret floristic patterns. Revised vegetation maps that account for lithology, broad land-form features and edaphic conditions would therefore be a better proxy for regional floristic variation than the presently used simple classes based on position along the catena.


Nomenclature
Tropicos.org. Missouri Botanical Garden

(www.tropicos.org; accessed 15 October 2012)

Abbreviations
EOCN

elevation over the channel network

MC

model-based clustering

NMDS

non-metric multidimensional scaling

UPGMA

unweighted pair group method with arithmetic mean

Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Classifying spatial patterns of floristic variation into relatively homogenous units is especially attractive to facilitate decisions in management, zoning and conservation planning. Forest classification schemes, however, are poorly developed in the tropics (Salovaara et al. 2004). Vegetation classifications of Amazonian forest are mainly based on broad-scale physiognomic classes (Veloso et al. 1991), while more fine-scale schemes that incorporate changes in species composition are rare. Local schemes have been proposed (Kahn & Castro 1985; Ribeiro et al. 1999), but field data are lacking for regional extrapolation. Amazonia's vastness, difficult access and mega-diverse flora constrain the rapid collection of reliable field data. In the absence of such a complete inventory, subsets of taxa (Higgins & Ruokolainen 2004) and abiotic environmental layers may be reliable surrogates for general biodiversity patterns (Margules et al. 2002). Validation and improvement of existing classification schemes are therefore urgent tasks for Amazonia.

Non-inundated forest is the predominant vegetation class in the Amazon Basin but this class is not floristically homogeneous. In central Amazonia edaphic and topographic features play a central role in shaping species composition. North of the Amazon River, soils are derived mostly from Cretaceous sediments and are nutrient-poor (Irion 1978; Chauvel et al. 1987; Quesada et al. 2010). The soil catena, i.e. the sequence of changes in soil properties along a local topographic profile, is the main source of edaphic heterogeneity over much of the landscape. A single catena can contain large variation in texture and hydrological properties, but not fertility, going from well-drained clayey soil on the plateau to more water-saturated sandy podzol on the lower slope near water courses (Chauvel et al. 1987; Rennó et al. 2008). Environmental conditions along the catena are mirrored by floristic changes (Kahn & Castro 1985; Costa et al. 2005, 2009; Bohlman et al. 2008). Vegetation along the catena is locally classified into forest classes, such as well-drained upland and seasonal swamp forests (Kahn & Castro 1985) or plateau, slope and valley forests (Ribeiro et al. 1999). Similar habitats at different sites, however, can be floristically distinct. This can be due to historical processes, such as disturbance and dispersal limitation (Zobel 1997; Phillips et al. 2003; Ozinga et al. 2005), or due to underlying site-specific environmental differences (Vormisto et al. 2004b). To date, the floristic consistency of these local topography-based vegetation types in central Amazonia has not been examined across more extensive geographic regions.

About 100 km northeast of Manaus, Zuquim et al. (2012) identified slightly more cation-rich soils than those overlying Cretaceous deposits. The composition of fern species and legume trees (Pansonato et al. 2013) was also distinct from sites on the Cretaceous deposits. This slightly more fertile site straddles the geological boundary between sedimentary Palaeozoic (ca. 425–390 MY) and igneous Precambrian (ca. 1990–1800 MY) formations. Physical and chemical conditions of soils in Amazonia are often strongly associated with geological history and landform evolution (Irion 1978; Rossetti et al. 2010; Higgins et al. 2011). For this reason, the same topographic habitat (i.e. catena compartment) compared between different geological landscapes may be floristically distinct. If this is true, the commonly used catena-based vegetation types will be poor proxies for floristic variation, unless modified to take into account geological and landform features.

Geological and landform features can be easily derived from well-mapped layers, such as geological maps and radar data. This offers an opportunity to predict and map floristic patterns across the extensive inaccessible areas of Amazonia. Geological maps and Shuttle Radar Topography Mission (SRTM) digital elevation data have already been used in western (Salovaara et al. 2004; Higgins et al. 2011, 2012) and eastern (Rossetti et al. 2010) Amazonia to understand and map forest types. Western Amazonia is a more dynamic landscape with more recent evolution than the central sedimentary basin and the Guiana Shield (Hoorn et al. 2010), and it encompasses a broader range of soil fertilities (Tuomisto et al. 2003a; Higgins et al. 2011, 2012). In eastern Amazonia Rossetti et al. (2010) examined the relationship between geology, landform and vegetation types but these types were physiognomically very distinct (e.g. savanna, forest). Additional field data from other geological contexts must be collected to validate the use of landform-based surrogates for classifying floristic patterns across the entire basin or within landscapes completely covered by forest.

Here, we used Zingiberales species, an order of understorey monocots herbs, as a subset of taxa to derive a floristic classification of inventories across ancient geological formations in central Amazonia. Landeiro et al. (2012) showed that high congruence of spatial patterns in central Amazonia between many terrestrial plant groups (including monocot herbs) is related to their common response to edaphic features. This suggests that there is a general environmental driver of floristic patterns and that a subset of taxa can potentially reveal them. Based on field data of 123 floristic inventories concentrated at three sites, we ask: (1) are topography-based vegetation types floristically consistent in central Amazonia, and (2) do broad landform and geological features control site-specific edaphic and floristic variation and therefore obfuscate the floristic classification based purely on local topographical classes?

We bring to bear a powerful but rarely used analytical approach in ecological studies, model-based clustering coupled with geodesic floristic distances. The fuzzy, noisy and non-linear nature of species composition data have led to unsatisfactory outcomes for most clustering algorithms. Data reduced using geodesic distances may circumvent these problems (Schmidtlein et al. 2010). Selecting the best classification solution is a major challenge in numerical classification. Unlike conventional clustering methods, model-based clustering selects the most parsimonious solution, allows overlapping clusters and provides a measure of class membership uncertainty (Fraley & Raftery 1998). To justify the use of these tools, we used geometric and non-geometric internal evaluators to compare the performance of our classification with that of conventional clustering methods and non-geodesic dimensionality reduction.

Methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Study sites

We conducted our study in a landscape of central Amazonia, north of Manaus (Fig. 1). Soils of this region are derived from three major parent rock types, which lie in vertical sequence and are also exposed from south to north in the northern limb of a syncline. Near to Manaus soils are derived from sandstone and unconsolidated fluvial deposits of kaolin clay and quartz sand, constituting the Alter do Chão formation of Cretaceous age (ca. 95–65 MY). From about 100 to 140 km north of Manaus are Paleozoic sandstones of the Trombetas formation (ca. 425–390 MY). North of these is a mosaic of igneous and metamorphic Precambrian rocks (ca. 1990–1800 MY) of the Guiana Shield (Projeto Radambrasil 1978; Fig. 1b). The contact between Paleozoic sandstone and the Precambrian shield coincides with a boundary between latosol and acrisol soils (Ibge-Embrapa 2001). This geological contact also partially matches with a broad landform boundary between the dissected plateaus of a sedimentary basin and the marginal depression found along the edge of the Guiana Shield of northern Amazonia (Projeto Radambrasil 1978; Ross 2005; Fig. 1c). The contact between Cretaceous and Paleozoic formations partially matches with a change in forest type based on elevation above sea level, from lowland to sub-montane rain forests (Ibge 2004). A 3-month dry season (<100 mm·mo−1) occurs near Manaus, changing to an aseasonal zone just 50 km further north (Sombroek 2001). Over the Alter do Chão formation soils are acid and low in cation and phosphorus content, changing from clayey on plateaus to sandy on lower slopes and bottomlands (Chauvel et al. 1987). Valleys and lower slopes also have superficial and shallow water tables, respectively, while soils on plateaus are well drained (Rennó et al. 2008). Soils overlying the Paleozoic deposits are also poor in nutrients but are shallower than soil of Alter do Chão (Irion 1978; Projeto Radambrasil 1978). The Precambrian region includes several igneous geological formations and is overlain by a mosaic of soils, with varying fertility and physical structure (Projeto Radambrasil 1978; Sombroek 2000).

image

Figure 1. Site locations, landform and geological features of the study area, and geographical distribution of the floristic clusters. Map (a) shows the location of study sites in Amazon Basin. Other Precambrian formations in (b) are metamorphic and igneous rock formations. Map (c) shows: the elevation above sea level (a.s.l.) of study areas; four-cluster solution derived from model-based classification of all 123 plots (insets); geological boundary between igneous and sedimentary formation (white line); marginal depression in dark grey to north of the white line and dissected plateaus in light grey to the south. Size of cluster symbols represent sum of cation content. Cluster symbols are the same as in Fig. 3a.

Download figure to PowerPoint

Field data were collected at three sites: Ducke Forest Reserve, reserves of the Biological Dynamics of Forest Fragments Project (BDFFP) and the Uatumã Biological Reserve. The first two are located on the infertile Alter do Chão formation, while the last straddles a contact between sandstone of the Trombetas and igneous rocks of the Iricoumé formation, part of the Guiana Shield. More details concerning Ducke and Uatumã can be found in Costa et al. (2011). The BDFFP sites are described in Zuquim et al. (2008).

Data collection

Plant, soil and topographic field samples

We used the monocot order Zingiberales as subset taxa to investigate floristic patterns for several reasons. Zingiberales species are of accessible height. They are terrestrial herbs, ranging from a few centimetres up to 10 m in height (Costa et al. 2011). Taxonomy is well established. Collecting is easy and identification relatively reliable. Details of ecology, evolution, natural history and the description of Zingiberales species included here can be founded in Costa et al. (2011).

Plant inventory followed the RAPELD protocol of the Program for Biodiversity Research (PPBio; Costa & Magnusson 2010). Zingiberales species were inventoried in 123 plots, 0.05 ha each (250 m × 2 m), with 57 plots at Ducke, 37 at BDFFP and 29 at Uatumã. Within each reserve, plots were arranged in a grid and spaced at least 1 km apart. As noted above, in central Amazonia there is a high correlation between soil properties and position on the catena (Chauvel et al. 1987; Luizão et al. 2004). To minimize the within-plot soil variation, each elongate plot followed the same altitudinal contour. Fieldwork was conducted from 2001 to 2002 at Ducke and from 2006 to 2007 at BDFFP and Uatumã.

We identified and counted every individual with a height >5 cm rooted inside the plot. One clump of clonal species was treated as one individual. Clumps were defined as groups of stems or leaves arising from the soil less than 20 cm from each other, or based on our field experience with each species. We identified the species based on specialized literature and sent fertile material to specialists Helen Kennedy and Paul Maas, to confirm identifications. Voucher collections are deposited in the INPA Herbarium in Manaus, BR, the University of British Columbia Herbarium, CA, and the National Herbarium of the Netherlands at Utrecht University, NL.

We measured four environmental variables for each plot: soil texture, the sum of exchangeable cations (Ca2+, Mg2+ and K+), elevation over the channel network (EOCN) and slope. Sodium was not considered in the sum of exchangeable cations because concentrations were below detection level. Soil samples were taken at a depth of 5 cm at 50-m intervals along the main axis of each 250-m long plot and mixed to produce a composite sample for each plot. Soil collection and analyses followed the same protocol for the three sites, available on the PPBio website (http://ppbio.inpa.gov.br/repositorio/dados), for all three sites. Slope of each plot was the average of six measurements spaced 50 m apart, taken using a clinometer facing perpendicular to the main axis of the plot.

We determined the EOCN of each plot to obtain a topographic variable that is meaningful in terms of position along the soil catena and the associated variation in hydrological conditions. EOCN is very similar to the ‘height above the nearest drainage’ (HAND; Rennó et al. 2008). EOCN requires a hydrologically coherent digital elevation model (DEM) as input, derived from SRTM data. We used a DEM with a pixel resolution of 30 m, interpolated from 90 m SRTM resolution (INPE, http://www.dsr.inpe.br/topodata). We applied sink filling, determined flow direction for each pixel and constructed the drainage network. The headwater threshold for first order streams was 422 pixels at Uatumã and Ducke and 211 for BDFFP. These thresholds were based on field checking of the drainage systems. Finally, the original DEM elevations over sea level were converted to elevations above the nearest streams. Low EOCN values correspond to riparian environments (valleys) and medium to high EOCN values to slopes and plateaus (uplands). EOCN was obtained using the hydrological tools of the SEXTANTE spatial data analysis library (http://www.sextantegis.com/), coupled with the open-source gvSIG v 1.10 software (http://www.gvsig.org/web/). At Ducke and Uatumã, EOCN of each plot was the average of six points taken from 0 to 250 m at 50-m intervals along the plot's long axis. Point locations on the DEM were measured in the field using a Garmin GPS, model 76CSx. At BDFFP, geographic coordinates were taken only at the beginning of each plot with a lower precision GPS, and a single EOCN value was calculated per plot.

We use two thematic layers, geological maps and SRTM elevations, to investigate large-scale landform control of floristic patterns. Geological maps were derived from the Brazilian RADAM Project (Projeto Radambrasil 1978) based on extensive fieldwork and visual interpretation of landforms in 1:250 000 scale X-band radar images obtained in 1970/71. Vector versions were downloaded from the GEOBANK database of the Brazilian Geological Service (CPRM) (http://geobank.sa.cprm.gov.br). For SRTM elevations, we used the same digital elevation model (30-m spatial resolution) used to obtain EOCN.

Data analyses

Our analyses can be summarized as four steps. First, we performed model-based clustering (MC) on scores from a floristic ordination of 123 inventories using geodesic floristic distances. (We used presence–absence data because most species are uncommon in Amazonia and their contributions to floristic structure, using abundance data, therefore likely to be highly diluted or hidden by the overwhelming influence of a few very common species.) Second, using a geometric and a non-geometric criterion, we compared the performance of model-based clustering against that of conventional hierarchical clustering and non-geodesic floristic distances. Third, we used ANOVA to test for edaphic and topographic differences between the model-based clusters. Finally, we used the Kappa index of agreement and ANOVA to test if geological classes mirror floristic and environmental variation. All data analyses were carried out in the R environment (R Foundation for Statistical Computing, Vienna, AT). GIS layers were handled using open-source software, gvSIG v1.10 and QGIS v1.6.0 (http://qgis.org/). The four steps are detailed below.

Step 1: Ordination and model-based clustering
Floristic ordination

To extract and visualize floristic patterns, we used non-metric multidimensional scaling with geodesic floristic distances (NMDS-G; Mahecha et al. 2007). The NMD-G is equivalent to isometric feature mapping (Isomap; Tenenbaum et al. 2000), a non-linear generalization of classical multidimensional scaling (CMDS). Such methods of dimensionality reduction have great promise for extracting non-linear ecological structures from hyperspace (McCune & Grace 2002; Mahecha et al. 2007; Schmidtlein et al. 2010). Conventional ordination methods can produce unsatisfactory representations of object relationships in reduced space, due to curvilinear distortion (Williamson 1978; Minchin 1987). This problem manifests itself mainly in high beta-diversity communities, where many pairs of samples do not share any species (Minchin 1987; De'ath 1999). As the cluster analysis will be performed on objects in a low-dimensional space, it is important to obtain a floristic representation with minimal distortion. The basic idea of ISOMAP is that a simple, single distance measurement is used only for points that are within a threshold distance, while geodesic distance is used between more distant points. Distance is measured by following any non-linear structures or surfaces containing data points, not across empty voids in feature space.

The algorithm is similar to the step-across method (Williamson 1978) and can be summarized as follows. (1) Calculate all pair-wise floristic distances between sites, D. We used presence–absence Sørensen index of similarity transformed into dissimilarities, given by the formula {1 – [2a * (2a + b + c)−1]}, where a is the number of shared species between a pair of sites, and b and c are the number of species restricted to each site. (2) Using Sørensen dissimilarities, define a web of connections between all sites by connecting each site with its k nearest neighbours. (3) Find the shortest path along the strands of this web for each site pair, i.e. create a new matrix of pair-wise geodesic floristic distances D(G). (4) Perform an NMDS ordination using the D(G) dissimilarity matrix. The heuristically obtained optimal value of k nearest neighbours is the smallest possible value that still leads to a completely connected web and that explains more variance in a low dimensional ordination. The process of finding optimal k runs, steps 1 to 4, n – 1 times, where n is the number of samples. With NMDS this requires long computing times, so CMDS was used in steps 1 to 4. After the optimal k was defined, we ran NMDS-G using the D(G) matrix and NMDS using the original Sørensen distance matrix for comparison. In both cases we set to 500 the maximum number of random starts in the search for a best solution. Final solution was set as that with the lowest stress value. NMDS axes were scaled using centring and principal component rotation. We chose to use NMDS-G because it provided satisfactory results when compared with CMDS-G (Mahecha et al. 2007). We used Pearson correlation coefficients to evaluate the environmental associations of NMDS-G axes. Geodesic distances, CMDS and NMDS use the functions isomapdist, cmdscale and metaMDS, respectively, from the vegan package in R.

Cluster analysis

To provide a floristic classification we applied the MC clustering algorithm (Fraley & Raftery 1998, 2002) to the final NMDS-G ordination solution. MC is built on finite mixture models, with each component probability representing a cluster. The mixture model assumes a Gaussian distribution. The mixture parameters, which include the number of clusters and their shape, volumes and orientation, were estimated via the iterative maximum likelihood algorithm expectation–maximization (EM; Dempster et al. 1977). Initial seed values were set from a model-based classical hierarchical clustering (Fraley & Raftery 1998). MC uses Bayesian information criteria (BIC), an approximation of Bayes factor, to set the final clustering solution. BIC searches for the most parsimonious solution, given by a balance of simplicity against complexity (high number of clusters) to avoid over-fitting. BIC is given by:

  • display math

where inline image is the maximized mixture log-likelihood for the model M and mM is the number of independent parameters estimated in the model. The best model has the highest BIC. Competing pairs of models with differences of less than 2 BIC values were taken to indicate weak evidence for selecting one over the other, differences between 2 and 6 to indicate positive evidence, differences between 6 and 10 strong evidence, and differences >10 very strong evidence (Kass & Raftery 1995). The EM estimates uncertainty of cluster membership for each of the 123 samples. These estimates were used as a measure of cluster overlap. MC was performed by the function Mclust in the R package mclust.

Step 2: Evaluating the performance of MC

We compared MC with UPGMA (unweighted pair group method with arithmetic mean), a conventional agglomerative hierarchical clustering method. We chose UPGMA because it is regarded as a high-performance method, because it is widely used in ecology and biogeography (Kreft & Jetz 2010), and because it is frequently applied in tropical forest studies (Tuomisto et al. 2003b; Salovaara et al. 2004; Emilio et al. 2010; Higgins et al. 2011). Cluster method performances were based on a non-geometric and a geometric evaluator (Aho et al. 2008). The non-geometric evaluator was derived from indicator species analysis (ISA; Dufrene & Legendre 1997), which is an intuitive approach using species composition data (McCune & Grace 2002; Aho et al. 2008). Basically, ISA provides information about the fidelity of a species to a predefined floristic cluster. We used a modified version of ISA that allows association of each species with more than one cluster (De Cáceres et al. 2010). We used the overall percentage of significant indicator species recorded in each cluster solution (ISE) to determine the performance of that solution. The higher the number of significant indicator species, the better the classification. For the geometric evaluation, we used the average silhouette width (ASWE; Rousseeuw 1987). Silhouette measures the tightness and separation of clusters. For each observation i, silhouette is defined as:

  • display math

where ai is the average distance between i and all other observations in the same cluster, and bi is the average distance between i and the observations in the nearest neighbouring cluster. The silhouette varies between −1 and +1, and for each cluster solution ASWE is given by the average of S(i) of all observations. The higher the ASWE, the better the classification.

For each method, we plotted ISE and ASWE vs each of six clustering solutions, ranging from two to seven clusters. We evaluated MC using ordinations based on both the geodesic distance matrix and the original Sørensen distance matrix. We performed UPGMA using both reduced and unreduced geodesic and Sørensen distance matrices to also evaluate the effects of dimensionality reduction. Overall, we estimated 36 ISE and ASWE values, 12 for MC (6 solutions × 2 types of distance matrix transformation) and 24 for UPGMA (6 solutions × 2 types of distance matrix transformation × original/reduced distances). UPGMA and ASWE were performed using functions agnes and silhouette, respectively, both from R package cluster. ISE was estimated using function multipatt in R package indicspecies.

Step 3: Environmental association of floristic clusters

Differences in four environmental variables – sum of cations, soil texture, slope and EOCN – were examined pair-wise between floristic clusters using the Tukey honestly significant difference (HSD) post-hoc test, following ANOVA. Sum of cation content was log-transformed because chemical reaction rates increase linearly along logarithmically transformed nutrient concentrations, following the model of Michaelis–Menten for the enzyme kinetics (Lambers et al. 2008). Accordingly, plants presumably perceive these logarithmic changes as if they were linear. EOCN was also log-transformed because it is directly related to water table depths, and above a certain height, increases in depth to groundwater will have a weaker effect on plant available water (Jackson et al. 1995). The log transformation also homogenizes the variance of the data. Percentage sand, which is the complementary measure of clay plus silt concentration, represented soil texture. One plot of the BDFFP and two plots of Uatumã did not have sand and slope data, respectively, and these were excluded only in this analysis. ANOVA was performed with the function aov and Tukey test with the function TukeyHSD, both in the R package stats.

Step 4: Searching for evidence of landform and geological control of regional floristic patterns and environmental conditions

We first mapped floristic clusters over a geological map and an SRTM topographic map to visually check if the main floristic separations were constrained by these land-form properties. This provided an a priori hypothesis of congruent classes. We grouped the geological and the floristic clusters each into two groups that formed a 2 × 2 confusion matrix. Alter do Chão and Trombetas formations were grouped together because they are characterized by the same broad landform feature (dissected plateaus) and type of parent rock (sedimentary rock). The igneous Iricoumé formation found in the marginal depression was the second geological class. We then used Cohen's kappa coefficient (CK) as a quantitative measure of agreement between the two floristic clusters and the two geological classes. CK is given by:

  • display math

where P(a) is the proportion of times that the classes agree and P(e) is the proportion of times that we would expect them to agree by chance. Values vary from 1, perfect agreement, to 0 indicating no agreement (Cohen 1960). We performed ANOVA to test whether the two geological classes differ in relation to the same environmental variables evaluated in step 3. CK was performed with the function kappa2 from R package irr.

Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

General results and ordination solution

We counted 11 397 individuals (mean per plot ± SD = 92.5 ± 82.7) belonging to 46 species (mean per plot ± SD = 8.3 ± 2.0), eight genera and five families (see Appendix S1 in Supporting Information). Only 17 individuals were discarded due to uncertain identification. The ten most common species constituted ca. 80% of all identified individuals. We found 25 species at Ducke, 20 at BDFFP and 34 at Uatumã. About 10% of inventory pairs did not share any species and an arch distortion was visible only between axes 1 and 3 in the final NMDS solution using Sørensen dissimilarity matrix (Fig. 2c). The CMDS with four nearest neighbours (= 4) gave the best association between geodesic floristic distances and Euclidean distances in reduced species space, so NMDS-G was performed with = 4 (see Appendix S2). The arch distortion was eliminated by NMDS-G (Fig. 2d). Otherwise, the NMDS and NMDS-G= 4 had similar results (Fig. 2). Two, three and four clusters were among the best solutions indicated by MC using from one to nine NMDS-G dimensions as optimal solutions (Appendix S2). Four dimensions were then chosen as the best solution, because this is the minimal number of dimensions that revealed four clusters, i.e. the potential ecological pattern in our data. (All four clusters occupied distinct habitats; see below.) Four axes explained 97.8% of the original geodesic distances and showed ca. 7.5% stress. Axes 1, 2 and 4 were associated with at least one environmental variable: axis 1 was correlated only with log sum of cations (= 0.603, < 0.001); axis 2 was correlated with slope (= 0.190, = 0.036); and axis 4 was correlated with sand concentration (= −0.204, = 0.024) and with log EOCN (= 0.220, = 0.015).

image

Figure 2. Arch distortion in 2D projection of 4D NMDS solution. NMDS was performed using Sørensen dissimilarity matrix (a and c) and geodesic distance transformation (b and d). First two axes are displayed in the top panels and axis 1 and 3 are displayed in the bottom panels.

Download figure to PowerPoint

Floristic classification: clustering solution evaluation

MC generally performed better than UPGMA using the non-geometric indicator species evaluator. For MC using geodesic distance, BIC indicated four clusters as the best solution (Fig. 3a), while ISE indicated four and seven clusters as the best solution (Fig. 3c). The difference between the second-best MC solution (three clusters, BIC = 222.44) and the best MC solution (four clusters, BIC = 228.96) was more than 6 units, indicating strong evidence in favour of a four-cluster solution. UPGMA using reduced geodesic distance showed the maximum ISE in a seven-cluster solution (Fig. 3c). The ISA values of all species are presented in Appendix S1. UPGMA often had better performance than MC using the ASWE evaluator (Fig. 3b). For both methods, ASWE indicated two clusters as the best solution. Taking the best cluster solution using both evaluators, the use of geodesic distance improved the overall classification (Fig. 3b, c).

image

Figure 3. Best solutions (i.e. best numbers of clusters) using three different evaluators. In (a) the single curve is derived from the MC method using geodesic distances. In (b) and (c) UPGMA was performed using: original Sørensen distance (UPGMA); original geodesic distance (UPGMA-dG); and reduced Euclidian distance based on Sørensen distance (UPGMAr) and based on geodesic distance (UPGMAr-G). MC was performed using Sørensen distance (MC) and geodesic distance (MC-G).

Download figure to PowerPoint

The best MC classifications using BIC and UPGMA coupled with the ISE criteria, had four and seven clusters, respectively, and showed partially similar patterns (Fig. 4). Objects in the centre and right side of the NMDS-G ordination graph were classified in almost the same way by both methods (Fig. 4a, c). The largest difference was in positioning of objects at the left side of the ordination graph. For these samples, UPGMA created four clusters of very different sizes, including two of very small size (Fig. 4c, d). MC generated two clusters in this area that were more equal in size (Fig. 4a). As BIC and ISE indicated that the four-cluster solution for MC method was a good solution, we concentrated on results based on this solution.

image

Figure 4. Best classifications from MC and UPGMA-rG, uncertainty of the MC classification and dendrogram of UPGMA-rG. MC final classification was based on BIC values (four clusters); UPGMA-rG was based on either ISE (seven clusters) or on ASWE evaluator (two clusters represented by large circles in panel (c). Panels (a), (b) and (c) are 2D projections using the first two of four NMDS-G dimensions.

Download figure to PowerPoint

Model-based floristic classification: uncertainty, environmental association and geographic distribution

We detected both sharp and diffuse boundaries among the four floristic clusters in the MC classification. Cluster four had the most abrupt boundaries and distinct floristic composition (Fig. 4a). All of its inventories had <0.1% classification uncertainty (Fig. 4b). Cluster four shares only 13%, 17% and 49% of species with clusters one, two, and three, respectively. Inventory plots classified into clusters one, two and three had average uncertainties of 14%, 7% and 3%, respectively (Fig. 4b). Clusters one and two clearly overlap with each other in a 2D projection using the first two axes (Fig. 4a), but the final MC solution still indicates the presence of these two clusters. Their separation was more evident in the remaining NMDS-G dimensions (see Appendix S3).

The distinct floristic composition of cluster four was strongly associated with high soil cation content, while clusters one, two and three all had low cation content (Fig. 5). Soils of cluster three were slightly richer in cations than soils of cluster one. Cluster three was more associated with valleys (low EOCN values) and steep slope areas (high slope values) than cluster two. However, after excluding BDFFP plots, for which estimates of EOCN may have large errors due to low precision in GPS measurements, clusters one, three and four had lower values of EOCN than cluster two (Padj = 0.012, Padj < 0.001 and Padj = 0.004, respectively). Furthermore, among clusters one, two and three, log EOCN was strongly negatively associated with sand content (Pearson correlation = −0.78, < 0.001), which reflects pedogenic processes along the catena. Therefore, cluster two was associated with upland clayey soils, cluster one with valley sandy soils and cluster three with valley and steep slope sandy soils.

image

Figure 5. Box plots and Tukey HSD post-hoc test showing the environmental differences between the four model-based floristic clusters (categories 1–4 on x-axis) in paired comparisons (Padj < 0.05). Panel (a) is log scale and (c) is log +1 scale and panels (b) and (d) show the original values. Within each graph, clusters that do not share the same letter have significantly different means.

Download figure to PowerPoint

Cluster four (rich in cations) was concentrated in the northwest corner of the Uatumã landscape (Fig. 1c). This region had low terrain elevation values and partially matched the Iricoumé formation. There was significant agreement between the two consolidated geological classes and the floristic classification when consolidated into two sets of clusters: cluster four comprising one set and clusters one, two and three the other set (CK = 0.70, < 0.001). Cluster four was associated with the igneous Iricoumé formation and all other clusters with the group of sedimentary formations (Fig. 1c). Only the soil chemical variable – cation content – was different between the two geological classes. The soils of the igneous Iricoumé formation had the highest cation content (Table 1). In summary, the spatial distribution of floristic clusters was closed linked with the spatial distribution of environmental variables, which in turn were governed by geological and landform features.

Table 1. Environmental attributes compared between geological formations. First three columns give non-transformed mean values ± 1 SD; last two columns show Fisher value (F) and the probabilities associated with ANOVA (< 0.05) between the two groups of rock types: sedimentary and igneous. For sum of cations and EOCN, ANOVA was performed using log and log + 1 transformations, respectively
Type of rockGeological formationANOVA
Alter do ChãoTrombetasIricoumé
SedimentaryIgneous F P
Sum of cations (cmolc·kg−1)0.34 ± 0.200.49 ± 0.402.58 ± 2.0147.31<0.001
Sand (%)45.89 ± 31.4338.88 ± 23.4536.89 ± 19.090.370.541
EOCN (m)20.01 ± 15.6619.99 ± 18.559.74 ± 8.571.750.189
Slope (°)10.10 ± 7.1811.97 ± 8.4511.33 ± 4.600.010.943

Discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Abrupt floristic changes are linked to soil properties, lithological matrix, landform features and landscape evolution

We find a distinct species composition associated with cation-rich soils in central Amazonia. We also demonstrate that variation in cation content is tightly linked with geological and broad landform features. Variation of soil cation content has been well documented as a main determinant of floristic variation in western Amazonia (Phillips et al. 2003; Tuomisto et al. 2003a,b; Higgins et al. 2011), but rarely in the central region (Pansonato et al. 2013; Zuquim et al. 2012). Soils of both the central Amazonian sedimentary basin and the Guiana Shield have been generally considered infertile (Irion 1978; Stropp et al. 2009). Indeed, the range of cation content found here (0.02–5.21 cmolc·kg−1) is narrower than the range reported for studies in western Amazonia (0.08–24.1 cmolc·kg−1; Tuomisto et al. 2003a; Higgins et al. 2011). In a broader Amazonian context, soils of the Iricoumé formation are rather infertile. The narrow range of variation in soil nutrients found in central Amazonia is nonetheless sufficient to bring about strong floristic variation. Our findings are in agreement with those for western Amazonia: chemical properties of soils play a central role in shaping spatial patterns of floristic composition at landscape and regional scales (Phillips et al. 2003; Tuomisto et al. 2003a,b; Higgins et al. 2011).

The properties of soils in Amazonia are intrinsically related with type of parent rock (Fittkau et al. 1975; Irion 1978). Soils of the igneous Iricoumé formation are clearly richer in cations than those of sedimentary formations. Sandstones and siltstones of the Alter do Chão and Trombetas formations were built from sediments eroded from the peripheral portion of the Guiana Shield (Irion 1984; Santos 1984). During transport and subsequent pedogenesis these sediments were weathered and leached (Chauvel et al. 1987). Soils of Iricoumé are predominantly shallow and rock outcrops are common (F.O.G. Figueiredo, pers. obs.). The exposed rocks supply new pools of nutrients for soils and consequently for plants. Rhyolite, rhyodacite, dacite and andesite are the common rocks found in the igneous Iricoumé formation. These naturally contain minerals richer in soluble nutrients − Ca2+, K+ and Mg2+ – than the sandstones of the Trombetas and Alter do Chão formations (Projeto Radambrasil 1978).

Broad landforms within the Guiana Shield are also linked to the type of rock (Kroonenberg & De Roever 2010) and its resistance to chemical weathering. Granite is one of the most resistant rocks to chemical dissolution (Franke 2009) and possibly more resistant than rhyolite, rhyodacite, dacite and andesite. These differences in resistance generate differences in depth of the weathering front and therefore differential erosive processes create altitudinal differences. The low-elevation Iricoumé formation surrounded by high-elevation granite formations (Projeto Radambrasil 1978; Fig. 1 in our study) is evidence of such processes. Hence, we conclude that long-term processes of erosion and deposition occurring over a lithological matrix explain the variation in relief morphology, soil properties and consequently floristic patterns in central Amazonia.

Topography-based forest types have diffuse floristic boundaries and complex interactions with environmental variables

After the main floristic segregation by soil fertility/geology, the remaining floristic clusters segregate along the catena as previously described in Kahn & Castro (1985) and as partially suggested by Ribeiro et al. (1999). These authors suggested a floristic segregation between valleys, slopes and plateaus. However, we find the separation between valleys and plateaus to be relatively fuzzy, making it difficult to set a clear-cut line between forest types along the catena. Presence of transition zones (ecotones) between such topographic classes has been suggested to bring about differences in herb species composition at the Ducke Reserve (Drucker et al. 2008). The floristic ecotone is likely linked with a hydrological ecotone, a transition zone between valley and uplands marked by fluctuations of the water table level (Rennó et al. 2008; Nobre et al. 2011).

The environmental interpretation of forest types based on valley and upland classes appears to be more complex than previously thought. At the BDFFP and Ducke Reserves, floristic separation between cluster one (valleys) and two (uplands) was linked to soil sand content, as in previous studies (Costa et al. 2005; Zuquim et al. 2008). However at Uatumã, floristic separation of upland and valleys (clusters two and three) is also associated with slope, and weakly associated with sand. Also, a significant floristic separation is observed within valleys (clusters one and three), related to differences in cation content. Valley forests at Uatumã are slightly richer in cations than valleys in the other sites, possibly due to subtle physical and chemical differences between the underlying Palaeozoic and Cretaceous sediments. Even within poor soil sites, geological and land-form features may be controlling floristic differences between similar topographical habitats.

Spatial floristic patterns are partially related to spatial structure of environment and neighbourhood effects

The sampling procedure adopted in our study precludes inference of detailed spatial patterns; spatial distances <1 km and between 20–50 km are not represented in the data. Nevertheless, some spatial patterns can be seen. Dispersal limitation has usually been identified as an important process that contributes to floristic differences between sites in Amazonian forests, modelled by the decay of floristic similarity with increasing geographical distance (Vormisto et al. 2004a; Bohlman et al. 2008). However, floristic differences between distant sites can be associated with site-specific environmental conditions and do not necessarily result from geographic separation per se. The spatial distribution of clusters in our analyses reflects a spatially structured environment, making it difficult to distinguish the effects of dispersal limitation (endogenous spatial processes) from the effects of spatial structure of the environment (exogenous processes). Although there are analytical tools for partitioning floristic variation into pure environment effects, pure spatial effects and interactions, such procedures have been criticized (Smith & Lundholm 2010): the interpretation of environmental control depends not only on the environmental variation but also on the spatial structure of the environment.

Model-based clustering with geodesic distances is a useful approach for vegetation classification

The MC and the parsimony-based approaches have shown promising results for classification of vegetation data sets (Dale et al. 2001, 2010), but have seldom been used. Although MC does not perform well using the ASWE evaluator, geometric evaluation criteria have some drawbacks. ASWE turn out to be able to identify only sharply defined floristic clusters or outlier groups while being insensitive to variation in the non-outlier groups. The segregation of isolated clusters does not mean that a valid cluster solution has been identified (Aho et al. 2008). Also, ASWE produced few clusters as an optimal solution, contrary to the outcome provided with ISE. The BIC was more related to non-geometric validation (ISAE). Internal non-geometric criteria should be preferred over internal geometric criteria in floristic classification because phytosociologists are often interested in relating floristic clusters with indicator species (Dale 1995).

Unlike UPGMA using a non-geometric evaluator, MC coupled with BIC will not produce clusters containing just one or a few sites (Fraley & Raftery 1998). Classification methods that allow small clusters are not desirable because there are difficult to interpret ecologically and statistically (Tichý et al. 2010). Indeed, the parsimony-based approach avoids high complexity in the final clustering solution, but still provides an informative classification regarding species associations and environmental conditions. Future studies using stability-based criteria to set final cluster solutions (Tichý et al. 2011) and other species-based approaches (Aho et al. 2008) should provide a more rigorous test of model-based clustering.

The MC searches for geometric structures with Gaussian distributions. Parameter estimates for multivariate mixture models are methodologically constrained by high-dimensional data (Fraley & Raftery 1998). Geodesic distances, therefore, can be useful in transforming non-geometric shapes present in the original dissimilarity space into relatively geometric cluster shapes (Schmidtlein et al. 2010) and in revealing essential underlying features of multivariate ecological data using few dimensions (Mahecha et al. 2007). Geodesic distances also eliminate a subtle curvilinear distortion in our NMDS ordination. Minchin (1987) suggested the use of local non-metric multidimensional scaling (LNMDS) coupled with Bray–Curtis distance as a robust technique to deal with arch distortion. However, arch distortions have been observed even with LNMDS (Økland & Eilertsen 1993). Curvilinear distortion is often associated with high turnover of species (beta-diversity; Minchin 1987). High beta-diversity is observed in fine-grained sampling of a broad environmental space. Fine-grained sampling of the environment is necessary to capture subtle environmental specialization of tropical plants (Drucker et al. 2008). Therefore, we recommend the use of geodesic distances to extract subtle and reliable ecological patterns from a multivariate data set, especially when employing fine-grained sampling of a broad environmental gradient.

Conclusions and perspectives for floristic classification in Amazonia

Geological and broad landform features control edaphic and floristic patterns in central Amazonia. Given this tight control, simple topography-based vegetation types, such as upland and valleys, fail to represent regional floristic variation due to site-specific soil conditions. Hence, lithology and broad landform features appear to be reliable proxies for floristically defined forest types in central Amazonia. Landform features can be derived from digital elevation models, satellite images and geological maps (Rossetti et al. 2010; Higgins et al. 2011) and, therefore, can be used to map floristic patterns across more extensive and inaccessible geographic regions in Amazonia. The use of a subset of taxa, such as monocot herbs or other taxa (Higgins & Ruokolainen 2004), seems to provide a cost-efficient validation of such maps. MC and geodesic distance can make the analytical procedure of classification more reliable. In the near future, improved vegetation maps could become a tangible and useful framework for management and conservation planning of Amazonian forests.

Acknowledgements

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

We thank INPA (Instituto Nacional de Pesquisas da Amazônia), PPBio (Programa de Pesquisas em Biodiversidade), CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico through the Pilot Program for Tropical Forests – PPG7), the BDFFP project, Instituto Internacional de Educação do Brasil (BECA program) and ICMBio – ReBio Uatumã for funding and support. Drs. Helen Kennedy and Paul Maas helped with species identification. Gabriela Zuquim, Joelson Nogueira and Fábio Espinelli assisted during fieldwork and in the collection of some of the environmental data. The staff of INPA's Soil and Plant Lab conducted most of the soil analyses.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information
  • Aho, K., Roberts, D.W. & Weaver, T. 2008. Using geometric and non-geometric internal evaluators to compare eight vegetation classification methods. Journal of Vegetation Science 19: 549562.
  • Bohlman, S.A., Laurance, W.F., Laurance, S.G., Nascimento, H.E.M., Fearnside, P.M. & Andrade, A. 2008. Importance of soils, topography and geographic distance in structuring central Amazonian tree communities. Journal of Vegetation Science 19: 863874.
  • Chauvel, A., Lucas, Y. & Boulet, R. 1987. On the genesis of the soil mantle of the region of Manaus, Central Amazonia, Brazil. Experientia 43: 234241.
  • Cohen, J. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20: 3746.
  • Costa, F.R.C. & Magnusson, E.M. 2010. The need for large-scale, integrated studies of biodiversity - the experience of the Program for Biodiversity Research in Brazilian Amazonia. Brazilian Journal of Nature Conservation 8: 312.
  • Costa, F.R.C., Magnusson, W.E. & Luizao, R.C. 2005. Mesoscale distribution patterns of Amazonian understorey herbs in relation to topography, soil and watersheds. Journal of Ecology 93: 863878.
  • Costa, F.R.C., Guillaumet, J.-L., Lima, A.P. & Pereira, O.S. 2009. Gradients within gradients: the mesoscale distribution patterns of palms in a central Amazonian forest. Journal of Vegetation Science 20: 6978.
  • Costa, F.R.C., Espinelli, F.P. & Figueiredo, F.O.G. 2011. Guide to the Zingiberales of PPBio sites in Brazilian Western Amazonia. Áattema, Manaus, BR.
  • Dale, M.B. 1995. Evaluating classification strategies. Journal of Vegetation Science 6: 437440.
  • Dale, M.B., Salmina, L. & Mucina, L. 2001. Minimum message length clustering: an explication and some applications to vegetation data. Community Ecology 2: 231247.
  • Dale, M.B., Allison, L. & Dale, P.E.R. 2010. Model selection using minimal message length: an example using pollen data. Community Ecology 11: 187201.
  • De Cáceres, M., Legendre, P. & Moretti, M. 2010. Improving indicator species analysis by combining groups of sites. Oikos 119: 16741684.
  • De'ath, G. 1999. Extended dissimilarity: a method of robust estimation of ecological distances from high beta diversity data. Plant Ecology 144: 191199.
  • Dempster, A.P., Laird, N.M. & Rubin, D.B. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society. Series B (Methodological) 39: 138.
  • Drucker, D.P., Costa, F.R.C. & Magnusson, W.E. 2008. How wide is the riparian zone of small streams in tropical forests? A test with terrestrial herbs. Journal of Tropical Ecology 24: 6574.
  • Dufrene, M. & Legendre, P. 1997. Species assemblages and indicator species: the need for a flexible asymmetrical approach. Ecological Monographs 67: 345366.
  • Emilio, T., Nelson, B.W., Schietti, J., Desmoulière, S.J.-M., Espírito Santo, H.M.V. & Costa, F.R.C. 2010. Assessing the relationship between forest types and canopy tree beta diversity in Amazonia. Ecography 33: 738747.
  • Fittkau, E.J., Junk, W., Klinge, H. & Sioli, H. 1975. Substrate and vegetation in the Amazon region. In: Dierschke, H. (ed.) Substrate and vegetation in the Amazon Region, pp. 7393. J. Cramer, Vaduz, LI.
  • Fraley, C. & Raftery, A. 1998. How many clusters? Which clustering method? Answer via model-based cluster analysis. The Computer Journal 41: 578588.
  • Fraley, C. & Raftery, A. 2002. Model based clustering, discriminant analysis, and density estimation. Journal of The American Statistical Association 97: 611631.
  • Franke, W.A. 2009. The durability of rocks – developing a test of rock resistance to chemical weathering. American Journal of Science 309: 711730.
  • Higgins, M.A. & Ruokolainen, K. 2004. Rapid tropical forest inventory: a comparison of techniques based on inventory data from western Amazonia. Conservation Biology 18: 799811.
  • Higgins, M., Ruokolainen, K., Tuomisto, H., Llerena, N., Cardenas, G., Phillips, O.L., Vásquez, R. & Räsänen, M. 2011. Geological control of floristic composition in Amazonian forests. Journal of Biogeography 38: 21362149.
  • Higgins, M.A., Asner, G.P., Perez, E., Elespuru, N., Tuomisto, H., Ruokolainen, K. & Alonso, A. 2012. Use of Landsat and SRTM data to detect broad-scale biodiversity patterns in Northwestern Amazonia. Remote Sensing 4: 24012418.
  • Hoorn, C., Wesseling, F.P., ter Steege, H., Bermudez, M.A., Mora, A., Sevink, J., Sanmartin, I., Sanchez-Meseguer, A., Anderson, C.L., Figueiredo, J.P., Jaramillo, C., Riff, D., Negri, F.R., Hooghiemstra, H., Lundberg, J., Stadler, T., Sarkinen, T. & Antonelli, A. 2010. Amazonia through time: Andean uplift, climate change, landscape evolution and biodiversity. Science 330: 927931.
  • Ibge. 2004. Mapa de Vegetação do Brasil, 3rd ed. Fundação Instituto Brasileiro de Geografia e Estatística, Rio de Janeiro, BR.
  • Ibge-Embrapa. 2001. Mapa de Solos do Brasil. Instituto Brasileiro de Geografia e Estatística, Ministério da Agricultura, Rio de Janeiro, BR.
  • Irion, G. 1978. Soil infertility in the Amazonian rain forest. Naturwissenschaften 65: 515519.
  • Irion, G. 1984. Clay minerals of Amazonian soils. In: Sioli, H. (ed.) The Amazon: limnology and landscape ecology of a mighty tropical river and its basin, pp. 537579. Dr. W. Junk, Dordrecht, NL.
  • Jackson, P.C., Cavelier, J., Goldstein, G., Meinzer, F.C. & Holbrook, N.M. 1995. Partitioning of water resources among plants of a lowland tropical forest. Oecologia 101: 197203.
  • Kahn, F. & Castro, A. 1985. The palm community in a forest of central Amazonia, Brazil. Biotropica 17: 210216.
  • Kass, R. & Raftery, A.E. 1995. Bayes factor. Journal of the American Statistical Association 90: 773795.
  • Kreft, H. & Jetz, W. 2010. A framework for delineating biogeographical regions based on species distributions. Journal of Biogeography 37: 20292053.
  • Kroonenberg, S.B. & De Roever, W.F. 2010. Geological evolution of Amazonian Craton. In: Hoorn, C. & Wesselingh, F. (eds.) Amazonia: landscape and species evolution. A look into the past, pp. 928. Wiley-Blackwell, Chichester, UK.
  • Lambers, H., Chapin, F.S. & Pons, T.L. 2008. Plant physiological ecology, 2nd ed. Springer, New York, NY, US.
  • Landeiro, V.L., Bini, L.M., Costa, F.R.C., Franklin, E., Nogueira, A., de Souza, J.L.P., Moraes, J. & Magnusson, W.E. 2012. How far can we go in simplifying biomonitoring assessments? An integrated analysis of taxonomic surrogacy, taxonomic sufficiency and numerical resolution in a megadiverse region. Ecological Indicators 23: 366373.
  • Luizão, R.C.C., Luizão, F.J., Paiva, R.Q., Monteiro, T.F., Sousa, L.S. & Kruijt, B. 2004. Variation of carbon and nitrogen cycling processes along a topographic gradient in a central Amazonian forest. Global Change Biology 10: 592600.
  • Mahecha, M.D., Martínez, A., Lischeid, G. & Beck, E. 2007. Nonlinear dimensionality reduction: alternative ordination approaches for extracting and visualizing biodiversity patterns in tropical montane forest vegetation data. Ecological Informatics 2: 138149.
  • Margules, C.R., Pressey, R.L. & Williams, P.H. 2002. Representing biodiversity: data and procedures for identifying priority areas for conservation. Journal of Biosciences 27: 309326.
  • McCune, B. & Grace, J.B. 2002. Analysis of ecological communities. MjM Software, Gleneden Beach, OR, US.
  • Minchin, P.R. 1987. An evaluation of the relative robustness of techniques for ecological ordination. Vegetatio 69: 89107.
  • Nobre, A.D., Cuartas, L.A., Hodnett, M., Rennó, C.D., Rodrigues, G., Silveira, A., Waterloo, M. & Saleska, S. 2011. Height above the nearest drainage – a hydrologically relevant new terrain model. Journal of Hydrology 404: 1329.
  • Økland, R.H. & Eilertsen, O. 1993. Vegetation–environment relationships of boreal coniferous forests in the Solhomfjell area, Gjerstad, S Norway. Sommerfeltia 16: 1254.
  • Ozinga, W.A., Schaminée, R., Bekker, M., Bonn, S., Poschlod, P., Tackenberg, O., Bakker, J. & Groenendael, J.M.V. 2005. Predictability of plant species composition from environmental conditions is constrained by dispersal limitation. Oikos 3: 555561.
  • Pansonato, M., Costa, F.R.C., Castilho, C. & Zuqim, G. 2013. Spatial or amplitude of predictors as determinants of the relative importance of environmental factors to plant community structure. Biotropica. 45: 299307.
  • Phillips, O.L., Vargas, P.N., Lorenzo, A., Cruz, A.P., Chuspe, M., Sánchez, W.G., Yli-halla, M. & Rose, S.A.M. 2003. Habitat association among Amazonian tree species: a landscape-scale approach. Journal of Ecology 91: 757775.
  • Projeto Radambrasil. 1978. Projeto RADAMBRASIL – Levantamento de recursos naturais, Vol 1–34. Depertamento de Produção Mineral, Rio de Janeiro, BR.
  • Quesada, C.A., Lloyd, J., Schwarz, M., Patiño, S., Baker, T.R., Czimczik, C., Fyllas, N.M., Martinelli, L., Nardoto, G.B., Schmerler, J., Santos, A.J.B., Hodnett, M.G., Herrera, R., Luizão, F.J., Arneth, A., Lloyd, G., Dezzeo, N., Hilke, I., Kuhlmann, I., Raessler, M., Brand, W.A., Geilmann, H., Moraes Filho, J.O., Carvalho, F.P., Araujo Filho, R.N., Chaves, J.E., Cruz Junior, O.F., Pimentel, T.P. & Paiva, R. 2010. Variations in chemical and physical properties of Amazon forest soils in relation to their genesis. Biogeosciences 7: 15151541.
  • Rennó, C.D., Nobre, A.D., Cuartas, L.A., Soares, J.V., Hodnett, M.G., Tomasella, J. & Waterloo, M.J. 2008. Environments in Amazonia. Remote Sensing & Environment 112: 34693481.
  • Ribeiro, J.E.L.S., Hopkins, M.J.G., Vicentini, A., Sothers, C.A., Costa, M.A.S., Brito, J.M., Souza, M.A.D., Martins, L.H.P., Lohmann, L.G., Assunção, P.A.C.L., Pereira, E.C., Silva, C.F., Mesquita, M.R. & Procópio, L. 1999. Flora da Reserva Ducke: Guia de Identificação das Plantas Vasculares de uma Floresta de Terra-firme na Amazônia Central. National Institute of Amazonian Research, Manaus, BR.
  • Ross, J.L.S. 2005. Os fundamentos da geografia da natureza. In: Ross, J.L.S. (ed.) Geografia do Brasil, pp. 1366. Edusp, São Paulo, BR.
  • Rossetti, D.F., Almeida, S., Amaral, D.D., Lima, C.M. & Pessenda, L.C.R. 2010. Coexistence of forest and savanna in an Amazonian area from a geological perspective. Journal of Vegetation Science 21: 120132.
  • Rousseeuw, P.J. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computation and Applied Mathematics 20: 5365.
  • Salovaara, K.J., Cárdenas, G.G. & Tuomisto, H. 2004. Forest classification in an Amazonian rainforest landscape using pteridophytes as indicator species. Ecography 6: 689700.
  • Santos, J.O.S. 1984. A parte setentrional do Cráton Amazônico (Escudo das Guianas) e a Bacia Amazônica. In: Shobbenhaus, C., Campos, D.A., Derze, G.R. & Asmus, H.E. (eds.) Geologia do Brasil, pp. 5791. Depertamento de Produção Mineral, Brasília, BR.
  • Schmidtlein, S., Tichý, L., Feilhauer, H. & Faude, U. 2010. A brute-force approach to vegetation classification. Journal of Vegetation Science 21: 11621171.
  • Smith, T.W. & Lundholm, J.T. 2010. Variation partitioning as a toll to distinguish between niche and neutral processes. Ecography 33: 648655.
  • Sombroek, W. 2000. Amazon landforms and soils in relations to biological diversity. Acta Amazonica 30: 81100.
  • Sombroek, W. 2001. Spatial and temporal patterns of Amazon rainfall: consequences for the planning of agricultural occupation and the protection of primary forests. Ambio 30: 388396.
  • Stropp, J., ter Steege, H., Malhi, Y., ATDN & RAINFOR 2009. Disentangling regional and local tree diversity in the Amazon. Ecography 32: 4654.
  • Tenenbaum, J.B., de Silva, V. & Langford, J.C. 2000. A global geometric framework for nonlinear dimensionality reduction. Science 290: 23192323.
  • Tichý, L., Chytrý, M., Hájek, M., Talbot, S.S. & Botta-Dukát, Z. 2010. OptimClass: using species-to-cluster fidelity to determine the optimal partition in classification of ecological communities. Journal of Vegetation Science 21: 287299.
  • Tichý, L., Chytrý, M. & Šmarda, P. 2011. Evaluating the stability of the classification of community data. Ecography 34: 807813.
  • Tuomisto, H., Poulsen, A.D., Ruokolainen, K., Morna, R.C., Quintana, C., Celi, J. & Cãnas, G. 2003a. Linking floristic patterns with soil heterogeneity and satellite imagery in Ecuadorian Amazonia. Ecological Applications 13: 352371.
  • Tuomisto, H., Ruokolainen, K., Aguilar, M. & Sarmiento, A. 2003b. Floristic patterns along a 43-km long transect in an Amazonian rain forest. Journal of Ecology 91: 743756.
  • Veloso, H.P., Rangel-Filho, A.L.R. & Lima, J.C.A. 1991. Classificação da vegetação brasileira, adaptada a um sistema universal. Instituto Brasileiro de Geografia e Estatística, Rio de Janeiro, BR.
  • Vormisto, J., Svenning, J.-C., Hall, P. & Balslev, H. 2004a. Diversity and dominance in palm (Arecaceae) communities in terra firme forests in the western Amazon basin. Journal of Ecology 92: 577588.
  • Vormisto, J., Tuomisto, H. & Oksanen, J. 2004b. Palm distribution patterns in Amazonian rainforests: what is the role of topographic variation? Journal of Vegetation Science 15: 485494.
  • Williamson, M.H. 1978. The ordination of incidence data. Journal of Ecology 66: 911920.
  • Zobel, M. 1997. The relative role of species pool in determining plant species richness: an alternative explanation of species coexistence? Trends in Ecology & Evolution 12: 266269.
  • Zuquim, G., Costa, F.R.C., Prado, J. & Braga-Neto, R. 2008. Distribution of pteridophyte communities along environmental gradients in Central Amazonia, Brazil. Biodiversity and Conservation 18: 151166.
  • Zuquim, G., Tuomisto, H., Costa, R.C., Prado, J., Magnusson, W.E., Pimentel, T., Braga-Neto, R. & Figueiredo, F.O.G. 2012. Broad scale distribution of ferns and lycophytes along environmental gradients in central and northern Amazonia, Brazil. Biotropica 44: 752762.

Supporting Information

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information
FilenameFormatSizeDescription
jvs12078-sup-0001-AppendixS1.pdfapplication/PDF79KAppendix S1. List of the 45 species of Zingiberales recorded in 123 inventory plots in central Amazonia, their respective number of individuals and multi-group indicator species analysis.
jvs12078-sup-0002-AppendixS2.pdfapplication/PDF227KAppendix S2. Results from the cumulative explained variance for the k nearest-neighbour solution and the number of clusters in best MC solution as a function of the number of dimensions in NMDS-G.
jvs12078-sup-0003-AppendixS3.pdfapplication/PDF167KAppendix S3. Model-based classification of 123 inventory plots displayed in all 2D projections of four NMDS-G dimensions.
jvs12078-sup-0004-AppendixS1.txtplain text document3K 
jvs12078-sup-0005-Suppinfo.docWord document403K 

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.