Introduction

Despite recent progress in the analysis and modeling of species diversity (Elith et al. 2011; Pavoine and Bonsall 2010; Rosindell et al. 2011), spatial patterns in the tree community composition of hyper-diverse rainforests are still poorly understood. Spatio-temporal processes that are supposed to explain them are partly theorized in niche (species habitat preferences) and neutral (population dispersal dynamics) models (Leibold et al. 2004). To disentangle these mechanisms, empirical ecological studies often infer ecological processes from an analysis of spatial patterns (McIntire and Fajardo 2009). One can, for instance, relate the spatial variation in species diversity to that in environmental determinants (e.g. soil, climate), while some spatial descriptors (e.g. geographical position, distance) can be used as surrogates for uncovering the scales of action of unmeasured or unmeasurable ecological processes (Dray et al. 2012). Among the environmental determinants, geodiversity (or geomorphodiversity) has often been mentioned (Thomas 2012; Tricart 1965) because the variations in background material directly influence soil properties and thus habitat suitability (Gray 2004; Parks and Mulligan 2010). It is also hypothesized that geomorphodiversity arise from paleo-environmental transformations that have potentially influenced migration routes by creating barriers and corridors (Santucci 2005). Geomorphodiversity, which results from interactions between many environmental processes over geological ages (climate change, erosion dynamics, tectonic movements, etc.), can thus shed light on biogeographical processes (Thomas 2011).

Several studies support the notion that modern Amazonian forest communities are marked by the imprint of major global environmental changes that occurred in pre-Pleistocene times (Ribas et al. 2012; Silva and Oren 1996). Paleo-climatic stability (inferred from latitude) and bedrock age are useful in explaining current alpha-diversity, whereas current climatic conditions are not (Stropp et al. 2009). The role of history suggests that new variables, which could account for past environmental filters, may help explain regional patterns of forest diversity. Landscape features shaped by diastrophic geological activities, have already been suggested as possible indicators of Amazonian beta-diversity (Hammond 2005; Sombroek 2000). However the explanatory power of this “geomorphic control” (as defined by Hammond in Hammond 2005) has never been tested explicitly.

French Guiana in Northeast Amazonia has fairly weak geological and altitudinal gradients but shows substantial geomorphodiversity resulting from differences in the susceptibility to erosion of its granitic substrates (Filleron et al. 2004). This has led the weathering mantle to diversify into contrasted landforms and soil covers (Sabatier et al. 1997), and these are thought to have influenced the long-term evolution of the forest. French Guiana may thus be viewed as an ideal case study to test the extent to which geomorphodiversity can explain tree beta-diversity (i.e. variation in floristic composition). Should geomorphodiversity prove to be a strong predictor of beta-diversity, multi-scale geomorphological analyses could be an efficient tool for forest management planning.

The specific aims of our study are thus: (i) to identify spatial patterns in the floristic composition of French Guiana forest at local and regional scales, and (ii) to identify the relative and absolute contributions of geomorphological and other geographical factors in explaining these spatial patterns (i.e. facets, landforms, landscapes). We use a holistic geomorphological stratification (i.e. at facet, landform, landscape scales) and an extensive forest inventory to assess multi-scale environmental variations across French Guiana. We analyze the data using multivariate partitioning methods based on Simpson’s metric of species diversity. Finally, we discuss the implication of our findings in terms of ecological insights that could help forest managers and conservation practitioners to prioritize their efforts both in French Guiana and throughout Amazonia.

Materials and methods

Study area

As part of the Guiana shield in Northern Amazonia, French Guiana (4°13′N, 52°59′W) covers about 85,000 km2 and has a mean altitude of about 140 m above sea level with few mountainous peaks exceeding 800 m. Its climate is equatorial with a mean annual temperature of about 26 °C and rainfall ranging from 4,000 mm/year in the northeast to 2,000 mm/year in the south and west. The number of consecutive months with less than 100 mm precipitation (dry season) varies from two in the north to three in the south, with marked inter-annual variations. Fully 95 % of French Guiana is located on a 2.2–1.9 G year plutonic and volcanic basement. This substrate corresponds to the oldest and most homogeneous part of the Guiana Shield (Delor et al. 2003) and mainly supports ferralsols and acrisols (see Table 1). The few sedimentary rocks found in the country correspond to Quaternary deposits in the coastal lowlands, rare Precambrian sandstone, and conglomerates and quartzite formations in a narrow northern belt, supporting more varied soil types (see Table 1). Savannas and mangroves are found only in the coastal sedimentary plain, and rainforest covers more than 90 % of the territory. Timber harvesting and agriculture are restricted to sub-coastal areas currently covering less than 500,000 ha, close to the largest towns and main roads.

Table 1 Description of the 10 geomorphological landscape types defined from landforms distribution and classified into five larger categories (plains and valleys, multi-concave relief, multi-convex relief, plateaus, mountains)
Full size table

Collection of forest inventory data

Plot sampling design

We selected 33 sites to represent the country’s main geomorphological features and geological substrates, and to cover the forest extending outside the most disturbed coastal areas (see Fig. 1). Given the logistic constraints stemming from difficulties with accessibility, we made special efforts to sample the less documented southern part of French Guiana that is a National Park and is only reachable by river and helicopter after special authorization. To sample local environmental variability, we established two to four transects at each site, about 2.5–3 km long (total = 111 transects) in different directions. Each 20-m wide transect was divided into 100-m segments such that our 3,132 basic sampling units were 0.2-ha plots, nested within transects themselves nested within sites. The field campaign was conducted from 2006 to 2013. This nested sampling design had the advantage of allowing us to control scale-dependence effects by accounting for the appropriate spatial instrumental variables (sampling units) in analyses and by selecting the appropriate exchangeable units in Monte Carlo randomization tests of statistical significance (see "Data analysis" section).

Fig. 1
figure 1

Study area and sampling design a location of the different sites (circles) on geomorphological landscapes map (from Guitet et al. 2013—see table 1 for details); b example of sampling design within a site (transects in black); c elevation map from SRTM d geomorphological landforms (from Guitet et al. 2013); e expert-based biogeographic map (from Paget 1999); f vegetation types based on remote-sensing approach (from Gond et al. 2011)

Full size image

Topographic description and soil sampling

All the plots were georeferenced using a Garmin 76CSx GPS receiver (Garmin Ltd., Southampton, UK) and delineated in the field using a Vertex laser clinometer (Haglöf Sweden AB, Langsele, Sweden). Each plot was classified by field operators based on a nine category toposequence position index (from riverside to hilltop). This classification was used to supplement the geomorphological descriptors (see Table 1). We also measured plot slope, length and elevation to correct surface areas and crosscheck the DEM-derived variables. Soil samples were also collected from 490 plots selected to represent the different topographical positions along each transect. These samples were analyzed for their structure and chemical content (unpublished data) to assist with ecological interpretations, but were not directly included in the study analyses.

Tree sampling

All trees (including palms) greater than 20 cm in diameter at breast height (DBH at 1.3 m or above the buttresses) were identified by an experienced team of 13 “tree-plotters” who employed a standardized vernacular nomenclature refined to correspond either to botanical species, genera or families (Guitet et al. 2014). A total of 123,906 trees were sampled. Of these, 29 % were identified to species, 64 % to genus and 6 % to family level. This nomenclature proved to be 83 % accurate at the family level and 74 % accurate at the level of the most precise botanical equivalent of our vernacular names (i.e. the species, genus or family level) depending on the precision reached by the tree-plotters (Guitet et al. 2014). Only 810 trees were unidentified and these were grouped into a single category. In the following text, we use the generic term of taxa to refer to the most precise botanical equivalents of the vernacular names. The sampled trees belonged to 221 taxa of 51 families. As the Leguminosae family accounted for 23.8 % of all the trees, it was split into Mimosoideae, Papilionoideae and Caesalpinioideae sub-families (The Legume Phylogeny Working Group 2013).

Geomorphological stratification

We used the 30-m spatial resolution digital elevation model (DEM) provided by NASA’s Shuttle Radar Topography Mission (Farr et al. 2007) to produce a regional multi-scale characterization of geomorphodiversity. We first selected four local parameters derived from the DEM: elevation considered as a good proxy for altitude above sea level, mean slope (over eight neighboring pixels), catchment area (log-transformed) considered as a basic wetness index, and Height Above the Nearest Drainage [HAND; Renno et al. (2008)] as a proxy for local soil water conditions.

We also included geomorphological landform and landscape classifications resulting from a previous object-oriented analysis of medium-scale variations in micro-relief (Guitet et al. 2013). This previous analysis defined 224,000 landform units (<1 km2) that we further classified into 13 different types based on elevation, vertical profile, horizontal shape, wetness, slope, and area parameters (Fig. SI1). As the landforms showed a spatially structured distribution at the regional scale (Fig. 1d), a segmentation algorithm using dominance, diversity and aggregation indices was used to delineate 82 homogeneous geomorphological landscapes (100–5,000 km2), which we subsequently classified into ten different landscape types (Table 1; Fig. 1c, Fig. SI2). This multi-scale analysis allowed us to capture landforms and landscape diversity without defining the scale of perception a priori. The geomorphological landform and landscape classifications were further cross-validated based on topographical data collected at the study sites during field work.

Data analysis

Our general approach was based on multivariate multi-scale analyses (Dray et al. 2012) of the forest inventory and environmental data described above. These analyses can be summarized as four steps. First, we identified the most significant patterns of variation in floristic composition (beta-diversity patterns) using Non-Symmetric Correspondence Analysis (NSCA, Gimaret Carpentier et al. 1998) of both the family-by-plot and taxa-by-plot abundance tables, then attempted to explain these patterns by introducing the seven geomorphological variables (Table 2) as explanatory variables in constraint NSCA. Secondly, we characterized and tested the spatial patterns displayed along the main constrained and unconstrained NSCA axes using a variogram-based approach (Couteron and Ollier 2005). Thirdly, we used Moran’s Eigenvector Maps (MEM; Dray et al. 2012) to explicitly introduce spatial components into the constrained model and to partition the determinants of beta-diversity between environmental, instrumental (i.e. sampling) and spatial effects (Borcard et al. 1992). We then built the most parsimonious ANOVA-like model that accounted for the most significant environmental and spatial effects, and designed Monte Carlo randomizations to test these effects while taking to account of scale-dependent effects. Finally, this model was compared to alternative models routinely used in French Guiana for operational planning and based on either more conventional environmental variables (geology, rainfall, etc.) or expert-based, remote-sensing forest classifications (see Table 2).

Table 2 Variables tested to explain diversity patterns and spatial patterns in floristic composition of French Guiana forest at multiple scales
Full size table

All statistical analyses were performed using R software version 3.0 (R Development Core Team 2009) and the add-on packages “ade4” (Dray and Dufour 2007), “vegan” (Oksanen et al. 2007) and personal routines available at http://pelissier.free.fr/Diversity.html.

Step 1: Analysis of floristic composition using constraint and unconstraint ordinations

To detect beta-diversity patterns (i.e. patterns of variation in floristic composition), we first analyzed our forest inventory data using NSCA, an analytical approach that relies on Simpson’s metric of species diversity (Gimaret Carpentier et al. 1998) and is thus insensitive to the low frequencies of rare species (Pelissier and Couteron 2007). NSCA therefore places emphasis on the most common species that are also the most accurately identified by forest-plotters. It thus shows variations in main floristic background instead of highlighting the most peculiar situations, as does for instance classical Correspondence Analysis (Couteron et al. 2003; Rejou-Mechain et al. 2008). In fact, Simpson’s diversity computed from forest inventory data shows very good correlations with precise botanical data from the same sites (Guitet et al. 2014).

The effect of environmental variables on beta-diversity patterns was then investigated by NSCA on Instrumental Variables (NSCAIV) and partial NSCAIV (pNSCAIV). These analyses amount to applying to NSCA the classical principles of canonical and partial canonical analyses sensu Legendre and Legendre (1998). NSCAIV and pNSCAIV can be used to detect beta-diversity patterns that are (or are not) explained by environmental variables. Comparing the variance of the floristic tables obtained by NSCA with that obtained by NSCAIV (or pNSCAIV), partitions the diversity into fractions that are explained (or unexplained) by the explanatory variables (Pelissier et al. 2003). The correlation of the explanatory variables and their contribution to the main axes also helps to quantify their relative influence on floristic variations.

We used as explanatory environmental variables the seven geomorphological variables derived from the DEM and our geomorphological stratification (Table 2), introduced one by one or all together.

Step 2: Detection and characterization of spatial patterns using variograms

Floristic spatial structures (i.e. spatial patterns) were characterized and tested against spatial randomness using variogram-like functions that partition beta-diversity between pairs of plots with respect to their inter-plot geographical distances. The method was applied both to the complete abundance tables with beta-diversity measured in Simpson’s metric or to the fraction of (Simpson’s) beta-diversity displayed along the NSCA axes in the framework of the Multi-Scale Ordination (MSO; Couteron and Ollier 2005). MSO applies to constrained (NSCAIV) and residual (pNSCAIV) ordination axes so that comparing the variogram-like functions derived from NSCA, NSCAIV and pNSCAIV can indicate whether the explanatory variables account for the main spatial patterns observed in the data. Monte Carlo randomization procedures were used to test the statistical significance of the deviations of the variogram-like functions, with the null hypothesis being the absence of any spatial structure. These procedures were used to determine whether the floristic composition of two plots located a given distance apart was on average more or less similar (i.e. had higher or lower Simpson’s beta-diversity values) than expected for spatially independent plots.

Step 3: Partitioning the environmental and spatial effects in a global model

Environmental and spatial effects were disentangled by introduced MEM as explanatory variables in NSCAIV and pNSCAIV (Dray et al. 2012). MEMs are orthogonal eigenvectors derived from a neighborhood matrix of the sampled locations that yield spatial correlation templates ranked from large- to fine-scale spatial structures (i.e. from highest to lowest Moran’s I values, independently of the plots’ floristic composition). We applied a forward selection procedure (Dray et al. 2012) in order to retain in the spatial matrix only the most significant MEMs (i.e. with p < 0.05), and used these MEMs as explanatory variables, in NSCAIV and pNSCAIV, to highlight spatially structured beta-diversity patterns. With reference to Borcard’s et al. (1992) variance partitioning procedure, we then used the above spatial matrix and the environmental matrix to estimate:

  1. (i)

    the fraction of beta-diversity explained by the environmental variables once the effect of the spatial variables had been removed, i.e. the spatially unstructured environmental component or “pure environmental effect” (fraction a in Borcard et al. 1992);

  2. (ii)

    the fraction explained by the spatial variables once the effect of the environmental variables had been removed, i.e. the “pure spatial effect” (fraction c in Borcard et al. 1992);

  3. (iii)

    the fraction of the environmental effect that is spatially structured, called hereinafter the “mixed effect” (fraction d in Borcard et al. 1992).

This framework allowed us to select the environmental variables accounting for the largest proportion of the “pure environmental effect” (variable E) and the largest proportion of the “mixed effect” (variable M). Introducing sampling compartments (transects and sites) as instrumental variables in the same framework allowed us to determine the most appropriate sampling level (S) accounting for the “pure spatial effect”.

In order to test to what extent geomorphological variables can explain beta-diversity patterns while avoiding spatial dependence effects and pseudo-replication bias, we used restricted randomization procedures based on ANOVA-like pseudo-F ratios (Pelissier and Couteron 2007). We tested the effect of the environmental variable that showed the greatest “mixed effect” (M) by randomizing the appropriated exchangeable sampling compartments (S) between the categories of M. Likewise, we tested the effect of the environmental variable that showed the greatest “pure environmental effect” (E) by randomizing the elementary sampling units (i.e. the plots) between the categories of E restricted within the sampling compartments (S). This meant that variable M, which approximated the “mixed effect”, was tested while minimizing pseudo-replication bias due to distance-dependence from the pure spatial effect, and variable E, which approximated the “pure environmental effect”, was tested while taking the largest part of spatial structures into account (see for instance Anderson and Braak 2003).

Step 4: Comparison with conventional approaches

In order to compare the efficiency of our geomorphology-based approach with that of more conventional approaches we ran the same analyses with environmental explanatory variables such as climate and geology that are generally considered as the main factors influencing tropical beta-diversity (Higgins et al. 2011) and are often used to defined ecoregions (Bailey 2004). Data on dry season length and annual rainfall were provided by TRMM (Huffman et al. 2007) and geology data were extracted from the map in Delor et al. (2003). We also introduced two layers commonly used as references for practical forest management and conservation planning in French Guiana: Paget’s (1999) classification into five latitudinal biogeographical domains based on an expert-based approach combining climate, geology and expert phytogeographical knowledge (Fig. 1e), and the five forest types identified by Gond et al. (2011) from remote-sensing data derived from SPOT-vegetation satellite images, and which showed a broad-scale north–south organization partly consistent with the main rainfall gradient and the biogeographical domains (Fig. 1f).

Results

Floristic analyses revealed marked broad-scale beta-diversity patterns

The NSCA at the family level showed three prominent axes accounting for 55 % of total between-plot diversity and exhibiting significant spatial patterns (Fig. 2): (i) along axis 1 (24 % of total between-plot diversity), a decreasing abundance of Burseraceae and Mimosoideae anti-correlated with an increasing abundance of Lecythidaceae and Caesalpinioideae, resulting in a regular regional gradient from southeast to northwest French Guiana; (ii) along axis 2 (17 % of total between-plot diversity), two anti-correlated gradients in abundances of Lecythidaceae and Caesalpinioideae showing a marked spatial pattern for inter-plot distances of less than 150 km resulting in sub-regional patches dominated by Lecythidaceae mainly in the north-east and patches dominated by Caesalpinioideae in the east, south-west and north-west; (iii) along axis 3 (14 % of total between-plot diversity), a gradient of increasing abundance of both Chrysobalanaceae and Sapotaceae showing a marked local auto-correlation below 10 km resulting in sparse clusters dominated by these two families.

Fig. 2
figure 2

Central panel projection of plot (points) and family scores (circles) on the 3 main axes of the NSCA with the family-by-plot abundance table—circle size is proportional to abundance and grey levels indicate family position on axis 3—only the families with the greatest contributions are indicated. Left column MSO variograms of axes 1–3 with dotted lines indicating 95 % confidence intervals of the null hypothesis (significant points in white and non-significant in black). Right column mean per site of plot scores along axes 1–3, projected on maps

Full size image

The same NSCA at the taxa level (i.e. on the 221 most precise botanical equivalents) showed two prominent axes accounting for 26 % of total between-plot diversity (Fig. 3). This analysis appeared to be very consistent with the family-level NSCA and supported the observations made at the family level: (i) axis 1 (15 % of total between-plot diversity) was dominated on its positive side by Protium spp. (the most abundant Burseraceae genus) and Inga spp. (the most abundant Mimosoideae genus), while the negative side was dominated by Eperua falcata (the most abundant Caesalpinioideae species) along with the two most abundant Lecythidaceae genera (Lecythis spp. and Eschweilera spp.), and corresponded to the marked north–south regional gradient also detected at the family level; (ii) axis 2 (11 % of total between-plot diversity) was correlated on the positive side with Licania spp. (the most abundant Chrysobalanaceae genus), Pouteria spp. (the most abundant Sapotaceae genus) and Dicorynia guianensis, (the second most common Caesalpinioideae species), whereas Inga spp. and the group including E. falcata, Lecythis spp. and Eschweilera spp. were located on the negative side. This axis showed significant spatial dependence over short distances (ca. 50 km) that featured two northern clusters (corresponding to high abundances of Licania spp., Pouteria spp. and D. guianensis) separated by a central strip.

Fig. 3
figure 3

Central panel projection of plot (grey points) and taxa (black circles) scores on the 2 main axes of the NSCA on the taxa-by-plot abundance table—circle size is proportional to taxa abundance. Left panel variograms of MSO on axes 1 and 2 with dotted lines indicating 95 % confidence intervals of the null hypothesis (significant points in white and non-significant in black). Right panel mean per site of plot scores along axes 1 and 2, projected on maps

Full size image

Beta-diversity and spatial patterns are clearly related to geomorphological variables

When the seven geomorphological variables were used to approximate the floristic table by NSCAIV they explained 19 and 20 % of the family and taxa beta-diversity patterns, respectively. At the family level (Fig. 4), the first NSCAIV axis (35 % of NSCAIV inertia) was negatively correlated with the first NSCA axis and was mainly determined by opposition between multi-concave landscapes (D) and multi-convex landscapes (B, I, J). The second NSCAIV axis (17 % of NSCAIV inertia), which was closely correlated with the second NSCA axis, was mainly formed by opposition between mountainous landscapes (H) and plateaus with inselbergs (E, F). The other geomorphological variables appeared to be less important except height above nearest drainage (HAND) and also elevation and landform type 8 (very massive hills) which, as expected, were correlated with mountainous landscapes (H). Landform 9 (half-orange) and slope were also associated with multi-convex landscapes (B, J) on axis 1. Landform 12 (wet hillocks on low base-level) and valley landscapes (C) were negatively correlated with axis 2.

Fig. 4
figure 4

Upper panel correlation circle for the first 2 axes of NSCAIV performed on the family abundance table with respect to geomorphological explanatory variables (letters in brackets refer to landscape codes while numbers refer to landform codes explained in Table 1—topographical positions are noted in bold and other variables are underlined—only the most contributing variables are indicated). Lower panel projection of plots on NSCAIV factorial plan 1–2 clustered by landscape types (letters and ellipses) that can be grouped into five main categories (dotted lines)

Full size image

Consistent results were obtained at the taxa level. A strong correlation was observed between NSCAIV and NSCA axes, and a prominent contribution was made by landscape types (Fig. 5). Mountains (H) and multi-concave landscapes (D) were opposed to valleys (C) and multi-convex landscapes (B, I, J) on the first axis (52 % of NSCAIV inertia); and type E plateaus were opposed to multi-concave landscapes (D) and valleys (C) on axis 2 (18 % of NSCAIV inertia). However, we noted that other variables made a greater contribution at the taxon than the family level, especially topographical position on axis 2 that shaped a gradient from the upper topographical positions on the positive side (flat hill-top, upper-slope and ridge) to the lower topographical positions on the negative side (terrace, large talweg and narrow talweg). As expected, these lower topographical positions were also correlated with the smallest landform types 12, 13 (wet hillocks), with type 14 (large flattened and wet relief) and with the wetness index.

Fig. 5
figure 5

Upper panel correlation circle for the first 2 axes of NSCAIV performed on the taxa abundance table with respect to geomorphological explanatory variables (letters in brackets refer to landscape codes while numbers refer to landform codes explained in Table 1—topographical positions are noted in bold and other variables are underlined—only the most contributing variables are indicated). Lower panel projection of plots on NSCAIV factorial plan 1–2 grouped by topographical position (lower panel) showing a variation in composition from lower to upper positions

Full size image

In order to represent the spatially structured components of the floristic gradients, we selected 124 and 113 significant MEMs (p < 0.05) for the family and taxa abundances tables, respectively, and introduced these into the corresponding NSCAIV. Though slightly different in the two cases, the selected MEMs represented only very broad-scale structures (i.e. high Moran’s values). The MEMs explained 36 and 33 % of family and taxa between-plot diversity, respectively, indicating that the beta-diversity patterns were strongly structured on a broad scale.

We then used the geomorphological variables to approximate the fitted and residual tables resulting from the NSCAIV constrained by the MEMs. This led us to partition between-plot diversity into a pure environmental component, a pure spatial component and a mixed component corresponding to the spatially structured environmental effect (Fig. 6). The pure environmental component was weak at both family and taxa levels (1.3 and 2.6 %, respectively) whereas the mixed component and pure spatial component corresponded to larger and fairly equivalent parts of between-plot diversity (respectively 17.5 and 17 % for “mixed effect” and 18.4 and 16 % for “pure space effect”). Regarding the geomorphological variables, geomorphological landscapes explained the largest fraction of the spatially structured between-plot diversity (“mixed effects”) at both family (13 %) and taxa (12 %) levels, while the other geomorphological variables explained no more than 1–6 %. Topography was the geomorphological factor that showed the highest “pure environmental effect”, accounting for 1.4 and 0.5 % of between-plot diversity at the family and taxa levels, respectively.

Fig. 6
figure 6

Variance partitioning in relation to different instrumental variables (given on the vertical axis) and spatial variables (defined by the MEMs). Percentages indicate the ratio between explained variance obtained with NSCAIV and total diversity obtained with original NSCA. Explanatory factors are geomorphological variables used as environmental variables one by one and all together in the first part, combination of different forest descriptors as explanatory variables in the second part, and sampling compartments as instrumental variables in the last part

Full size image

The same partitioning procedure based on conventional environmental variables (geology and climate) explained less than 10 % of between-plot diversity, while biogeographical domains and remotely-sensed vegetation types customarily used by managers to approximate forest types led to even lower explanatory power (3–8 %) at both family and taxa levels (Fig. 6).

Variance partitioning using the sampling compartments as explanatory variables showed that the site effect accounted for more than two-thirds of the spatial effects, explaining 27 and 26 % of between-plot diversity at family and taxa levels, respectively, while the transect effect accounted for almost all the spatial effects, i.e. 33 and 30.5 % of between-plot diversity at the family and taxa levels, respectively (Fig. 6).

Geomorphological landscapes are sufficient to account for the main broad-scale variations in floristic composition

Based on the results presented above, we tested to what extent the geomorphological landscapes effect could explain the main taxa abundances, using transects as exchangeable sampling units in our permutation tests (considering their efficiency to include almost all the spatial dependence effects). This showed that geomorphological landscapes significantly accounted for the abundance variations of 80 taxa and 20 families (p < 0.005), corresponding to 65 and 77 % of the trees (see Supplementary Information Table SI3). In contrast, the topography effect, locally tested by permuting the plots between topographical categories within each transect, accounted for the abundance variations of 21 taxa and 7 families (p < 0.005), corresponding to 37 and 43 % of the trees.

The geomorphological landscape effect also explained the main spatial structures seen when introduced in MSO and applied to the total family-by-transect and taxa-by-transect tables, though slight autocorrelation was still seen below 50 km (Fig. 7).The same analyses using geology and climate as predictors showed significant scale-dependence effects below 100 km and above 150 km (not shown).

Fig. 7
figure 7

Multi-Scale Ordination (MSO) performed on the family-by-transect table (on the left) and the taxa-by-transect table (on the right). Full points correspond to the MSO on total tables. Empty points correspond to the MSO after removal of the geomorphological landscape effect. Dotted lines represent the 95 % bilateral intervals computed from 99 permutations between the sampling units

Full size image

Discussion

Regional beta-diversity patterns nested in continental gradients

Our multivariate analyses demonstrated that the rainforest of French Guiana is subject to complex, significant floristic variations, with regional patterns that include a marked latitudinal gradient. Variance partitioning showed that these broad-scale patterns are not closely related to geology or climate, as is commonly assumed by previous hypotheses, but are more comprehensively explained by geomorphological landscapes that express regional geomorphodiversity. It was no surprise that these patterns on a regional scale fit within the continental patterns previously described across Amazonia.

The Guiana shield differs from the Amazonia basin by its high level of endemism (Da Silva et al. 2005; Lopez-Osorio and Miranda-Esquivel 2010) and an unusual species composition (ter Steege et al. 2003). This particularity stems from a marked continental-scale family gradient of increasing Leguminosae abundance from southwest to northeast Amazonia (ter Steege et al. 2006) and secondary gradients of increasing Burseraceae at both genus (Emilio et al. 2010) and family (ter Steege et al. 2006) levels from northwest to southeast. On our regional scale, the most prominent pattern was also a marked latitudinal northwest to southeast gradient closely related to an increasing abundance of Burseraceae and a decreasing abundance of Lecythidaceae. This gradient also partly correlated with variations in the abundances of the principal Leguminosae sub-families corresponding to a shift between Caesalpinioideae and Mimosoideae from north to south. The same pattern has previously been detected in nearby Guyana with far more Lecythidaceae and E. falcata (the most abundant Caesalpinioideae) in the central and northern parts of the country (ter Steege 1998). These matches between regional and more continental gradients suggest that the regional broad-scale patterns detected at both family and taxa levels are partly nested within a larger continental framework and could be explained by similar large-scale structuring processes.

Rainfall, dry season intensity and geology were the main factors proposed in recent studies to explain these forest diversity patterns in Amazonia (Albernaz et al. 2012; Emilio et al. 2010; Stropp et al. 2009). These factors do indeed vary greatly on the continental scale (Sombroek 2000): from 1,200 mm/y to 6,400 mm/y for rainfall, from 0 to 6 months for dry season length and, for geology, from old crystalline bedrock from the Proterozoic (>2500 Ma) to recent sediments originating from Andean orogeny during the Cenozoic (<66 Ma) resulting in very different soil fertilities. They are consequently proposed as surrogates in the selection of reserve areas and for sampling Amazonian species diversity (Schulman et al. 2007). However, on our scale, these factors are more homogeneous and in our study explained less than 10 % of the variations in floristic composition.

In fact, even though French Guiana has a significant rainfall gradient, the fairly substantial total rainfall it receives (far above the 1,500 mm/y threshold that marks the transition between evergreen and semi-deciduous forest), and the short dry season (1–3 months with temporally well distributed rains even in its driest area), combine to reduce the impact of climate on vegetation in this region. Moreover, a series of ice-ages and episodic ENSO events may also have drastically modified spatio-temporal rainfall patterns over the past millennium in this region (Hammond 2005) such that present meteorological data (over a fairly short period of a few decades) may not reflect the average conditions that tree communities have experienced during their assembly. For all these reasons, we assume that the weakness of the climatic effect is not due to a lack of accuracy or precision of the tested data but rather to a lack of representativeness of the short period available for the climatic data in comparison with the ecosystem’s life-time.

Because of its influence on soil, geology often has a significant impact on tropical forest vegetation, particularly when geological substrates with very dissimilar ages or properties are compared (Fayolle et al. 2012; Phillips et al. 2003). Beyond the direct effect of soil chemical properties on species composition, geology is sometimes recognized as influential on long-term forest dynamics: forests are more dynamic on rich than on poor soils, such that speciation rates and evolutionary processes may differ (Stropp et al. 2009). The conventional distinction between nutrient-rich floodplains on quaternary sediment (“varzea”) and less productive drained uplands on older terrains (“terra firme”) is a common illustration of this effect on forest composition in the Amazonian basin (Clark et al. 1999; Sollins 1998). A rather similar effect has been put forward in Guyana to explain variations in composition between north and central Guyana on Berbice sedimentary formation and southern forests on crystalline substrate (ter Steege 1998). However, this effect has little relevance in French Guiana where (i) Precambrian crystalline rocks are very dominant over sedimentary rock, (ii) the very old, deep soil cover seems to have masked potential differences between soils developed from the alteration of various types of granites that are nevertheless chemically different, and (iii) soil weathering sequences are organized more locally along short catenas that vary according to geomorphological stages (Sabatier et al. 1997). As a result, even though similar patterns are detected in French Guiana, suggesting that similar ecological mechanisms affect forest composition on continental and regional scales, the same surrogates cannot be used to represent them. In particular, geology and climate are not reliable factors when considering conservation issues on this regional scale, as demonstrated herein.

Using geomorphological landscapes to approximate forest types

Our results demonstrate that a stratification based on regional geomorphodiversity, expressed by geomorphological landscapes, significantly explains beta-diversity patterns in French Guiana. We also show that a forest stratification based on this surrogate more efficiently captures variations in composition than expert- or remote sensing-based approaches as it explains a larger proportion of floristic variations and especially the distribution of many important taxa.

Previous botanical studies identified two forest types in French Guiana, a Caesalpinioideae-dominated forest assumed to be in the north, and a Burseraceae-dominated forest assumed to cover the south (Sabatier and Prévost 1990). The new insight provided herein by geomorphological landscapes expands on this distinction for taxa corresponding to more than 65 % of all individuals, including well-distributed taxa (e.g. E. falcata, Licania spp., Lauraceae—see Table SI4) and some rare and/or threatened species (e.g. Vouacapoua americana—see Table SI4). Based on these results we were able to define at least five main forest types defined by major landscape categories (i.e. plains & valleys, plateaus, mountains, multi-concave relief, multiconvex relief—see Fig. 4) coupled with dominant families (Tab. SI3): (i) forests dominated by Lecythidaceae covering the coastal plains and large valleys; (ii) Lecythidaceae mixed with Caesalpinioideae, and especially E. falcata, in multi-convex (hilly) landscapes that are common in the north; (iii) forests on plateaus, more common in central Guiana, associating Caesalpinioideae and especially D. guianensis that can be locally very abundant, with Burseraceae; (iv) Burseraceae becoming significantly more dominant on multi-concave landscapes (locally called “peneplains”) that are mainly situated in southern French Guiana, also including very abundant Vochysiaceae, Simaroubaceae and Mimosoideae; (v) Mimosoideae reaching their maximum abundance in the mountains that were seen to harbor greater local richness (Guitet et al. 2014) than other landscapes with far more diversified families and taxa including many Lauraceae. More subtle forest sub-types can also be documented based on the 10 landscape features identified in Fig. 1.

Our results, which underline the usefulness of geomorphological features as surrogates of forest composition, are consistent with those of a recent study that used monocotyledons as subset taxa in central Amazonia and showed broad landform features as drivers of floristic patterns (Figueiredo et al. 2014). Our results also support the hypothesis that long-term forest dynamics are under “geomorphographic control”, as previously proposed for the Guiana shield (Hammond 2005), but which had not been formally verified before our study. Ecological conditions and geomorphological landscapes are of course closely related, such that relations between geomorphodiversity and biodiversity can be partly interpreted as a habitat filter (e.g. generally higher rainfall and lower temperatures with altitude in mountain landscapes, more intense waterlogging in plains and valleys, etc.). Moreover, soil type seems to be partly linked to landscape features, as observed with the soil samples collected during the same field work (see Table 1) and as previously demonstrated across Amazonia (Quesada et al. 2010). However, except in sporadic cases (such as small white-sand patches or duricrust), the major soil types encountered in French Guiana (i.e. ferralsols and acrisols) show few chemical or structural differences (Quesada et al. 2009). Consequently, we assume that beyond this direct but weak niche effect, geomorphological landscapes principally include the imprint of past environmental filters and historical biogeographical processes that underlie differences in forest composition. In fact, the process of landforms formation is driven by tectonic, climatic and marine events that mark the geomorphological landscapes (Thomas 2011). These long-lasting events can also modify floral and faunal dynamics (Haffer 2008; Stehli and Webb 2002) and influence survival and migration processes, which seem to be relatively slow in tropical rainforests (Malcolm et al. 2006). We therefore assume that geomorphological landscapes are relevant proxies for forest long-term history. Geomorphological landscapes may therefore be viewed as an integrative marker of an ecological trajectory leading to major divergences in habitats and vegetation.

When geomorphology reveals divergent forest dynamics

Several factors point to regional tectonic background processes and recent climate changes being linked to observed geomorphological landscapes and their specific forest composition. For instance, the accumulation of sediments carried by Amazonian waters along the Guianan coasts has resulted in slight subsidence of the continental shelf (Warne et al. 2002) in front of the Guiana Shield. A consecutive 40-m upheaval over the last 300,000 years has also been detected in inland north-western French Guiana (Palvadeau 1998) due to crustal deformations. The sinking of the river network subsequent to this uplifting probably enhanced superficial erosion, which would explain the dominance of multiconvex landscapes and rejuvenated thin soils in this part of French Guiana (Beaudet and Coque 1994; Boulet et al. 1979). In fact, the original plateaus and their deep weathered ferralsols, which are still dominant in the stable south-eastern part of the territory, are gradually being replaced by a more dissected topography, with thinner soils and more superficial drainage that is shaping northwestern multi-convex landscapes. This long-evolving process may in turn have modified tree population dynamics and the composition equilibrium, especially for Eperua falacta which colonizes hillslopes in this northern landscape whereas it preferably occupies downhill and hydromorphic soils in plateau landscapes (Sabatier et al. 1997). Recent molecular phylogeographic analyses of E. falcata corroborate a still ongoing colonization dynamic accompanied by genetic differentiation (Audigeos et al. 2013). Thus, the multi-convex landscapes of French Guiana comprise forest with rapid dynamics resulting from a recent environmental imbalance, whereas plateau landscapes, on the contrary, indicate sustained stability which is conducive to slower evolution.

The case of mountainous landscapes (mainly between 300 and 500 m) also illustrates the strong relationship between long-term forest dynamics and geomorphology. Indeed, Guianan mountains show richer forest composition than lower elevations and therefore support the “mid-altitude bulge” effect which has also been detected in other tropical forests (Eisenlohr et al. 2013; Lomolino 2001). This effect is sometimes explained by the optimal soil temperature and intermediate fertility found at mid-altitude (Sanchez et al. 2013), which are assumed to reduce both competition and environmental constraints on vegetation. However, these zones could also have benefited from a “refugia” effect (Haffer 2008) that may explain the generally greater richness of the mountainous forests in both the north and south of the country (Guitet et al. 2014). In fact, a large body of evidence from palynology, sedimentology and anthracology suggests that regional forests were fragmented during Quaternary dry periods, corresponding to the series of ice ages and precession cycles (Duputie et al. 2009; Haffer 2008; van der Hammen and Absy 1994). But the deep, ancient and well-drained ferralsols that cover these high reliefs prove that they did not experience these recent dry phases because long periods of humidity are essential to ferralsols that need deep weathering with slow superficial erosion (Ferry et al. 2003; Sombroek 2000). In any case, this particular geomorphological landscape appears to be an efficient indicator of a specific ecological trajectory that leads to high alpha-diversity.

Systematic geomorphological approach: an efficient tool for conservation planning

Several recent studies conducted on a continental scale have concluded that pre-quaternary tectonic and climate modifications have had major effects on landscape and biodiversity changes in Amazonia (Cheng et al. 2013; Hoorn et al. 2010). Modern forests have been markedly influenced by climate fluctuations, fault reactivations and changes in the dynamics of water flows in several Amazonian localities (Rossetti et al. 2012). Therefore, an expansion of the geomorphological landscape mapping presented here for French Guiana could be useful for modeling beta-diversity patterns across Amazonia.

Several studies have underlined that geodiversity should be taken into consideration when drawing up global-scale biodiversity maps or identifying biodiversity hotspots as guides to conservation strategies (Parks and Mulligan 2010; Sombroek 2000; Thomas 2012). However, the practical application of this idea generally reduces the approach to local DEM variations or focuses on geological variability as the sole geodiversity approach (e.g. Schulman et al. 2007). In fact, geomorphodiversity is rarely evaluated on a large scale in tropical forests. But, thanks to recent remote-sensing products, particularly RADAR and LiDAR, and dedicated software developed for GIS, it is now easier to create objective and accurate geomorphological segmentations at all scales (e.g. Camargo et al. 2012). We therefore suggest that geomorphodiversity should be used more systematically for conservation planning, in addition to conventional environmental factors or multi-spectral remote-sensing images. The approach we describe in this study could be reproduced in other hyper-diverse tropical forests where environmental gradients are fairly smooth (e.g. Brazilian shield, west and central Africa or southern India) but also in more contrasted ecological contexts where geomorphological landscape features could provide richer surrogates of past environmental filters and long-term species migrations than the conventional, unsophisticated descriptors commonly used (i.e. latitude, longitude or age of geological substrate). Moreover, beyond their direct effects on flora, geomorphological features also appear to influence faunal communities as demonstrated by concurrent vertebrate inventories conducted along the same forest transects described in this study (Richard-Hansen et al. submitted). As a consequence, new geomorphological maps provided by satellite imagery could be very widely used in tropical regions to direct regional planning and forest conservation. Such approaches might more efficiently link geomorphodiversity and biodiversity, and also provide new insights to guide future surveys.