2.1 Study area

The Amazon forest, as circumscribed here (see outline in Figure 1), includes most of the Amazon and Orinoco river basins (excluding the Andean headwaters of some rivers and the mid-Orinoco Llanos), and the North Atlantic coastal river basins between the states of Delta Amacuro, in Venezuela, and Maranhão, in Brazil. With regards to elevation, a maximum altitude of 1100 m was established on the Andean flanks to exclude the complex and extensive vegetation and environment gradients associated with the massive mountain chain. No altitudinal limit was established, however, for the highlands of the Guiana Shield, which are entirely embedded in the Amazon Province. This is a controversial issue in the sense that some authors consider the Guiana Highlands as a separate biogeographic province (e.g., Cabrera & Willink, 1980; Cardoso et al., 2017), supported by the high number of endemic species, many of which are restricted to particular tepuis or highlands (Berry & Riina, 2005). We based our decision on the following facts: (a) unlike the Andes, which make up a natural limit, the Guiana Highlands are encircled and pervaded by Amazonian lowlands; (b) the highest altitudes reached by the Guiana Highlands (2500–3000 m) are modest compared to those of the Andes; and (c) tepuis and highlands also share a considerable number of species with lowland Amazonian vegetation types (Steyermark et al., 1995–2005).

2.2 Nomenclature

The white-sand vegetation complex is particularly thorny when it comes to nomenclature. The main reason for this is the remarkable variation in physiognomy, which ranges from grass/shrublands to forests with slender-trunked trees and more open canopies compared to those of adjacent terra firme forests growing on more clayey soils (Adeney et al., 2016). Throughout the Amazon, various local terms are also used to designate both the whole white-sand complex and its physiognomic expressions, for example, bana, caatinga amazônica, campina, campinarana, varillal, and chamizal (Demarchi et al., 2018; Fine et al., 2010; García-Villacorta et al., 2016; Stropp et al., 2011). Following Daly et al. (2016), we here adopt campinarana, because of its official use in Brazil (IBGE, 2012) and because it embraces the whole array of physiognomies growing on podzolized sands liable to ground water saturation, but conveniently excludes white-sand floodplain forests, which we distinguish in this paper.

Nomenclature for vegetation occurring on seasonal floodplains is less complex. We use the prevailing nomenclature for two main seasonally flooded vegetation types in the Amazon, distinguishing igapó and várzea forests, depending on the types of rivers along which they occur (see Junk et al., 2011; Kubitzki, 1987; Prance, 1979). Várzea forests are found along rivers carrying copious quantities of sediments (and nutrients), mostly brought from the Andes, with variation in the amount of clay resulting in waters that are many shades of brown. Confusingly, these rivers are often called white-water rivers (ríos de agua blanca, rios de água branca). In contrast, igapó forests are found along rivers with small amounts of suspended mineral particles, which are called black or clear-water rivers (ríos de agua negra o clara, rios de água negra ou clara). These rivers drain basins where white sands or other highly leached soils prevail (e.g., flowing from the Brazilian and Guiana Shields) and can carry vast loads of humic acid colloids resulting from the arrested litter decay in these hypo-dystrophic soils. A similar process takes place in black-water oxbow lakes severed from white-water rivers as well as in narrower upstream floodplains throughout the basin. The dichotomy of várzea and igapó falls short when it comes to rivers with “mixed” waters, and rivers with temporal and spatial variations in suspended particles, of which the Casiquiare Channel in Venezuela is an example. In both igapós and várzeas, vegetation structure varies from tall forests to floodplains with more open formations, depending on local flooding dynamics and related processes of either erosion or sedimentation (Kalliola et al., 1992; Luize et al., 2018; Salo et al., 1986; Worbes et al., 1992). The timing and duration of flooding in these forests can be variable, from once every few decades in rivers close to the Andes (e.g., on the Manu River in Peru, pers. comm. John Terborgh) to multiple months annually for the iconic várzeas and igapós along major rivers such as the Amazon and the Rio Negro.

2.3 Dataset

We extracted the dataset from the NeoTropTree (NTT) database (http://www.neotroptree.info/), which consists of tree species checklists compiled for geo-referenced sites, from southern Florida (U.S.A.) and Mexico to Patagonia in Argentina and Chile. Trees here are defined as freely standing woody plants >3 m in height, including tree ferns. NTT currently holds 7485 sites/checklists, 20,562 tree species, and 1,206,314 occurrence records. A site/checklist in NTT is defined by a single vegetation type, following the classification system proposed by Oliveira-Filho (2015), contained in a circular area with a 10 km diameter. Where two or more vegetation types co-occur in the area, there can be multiple geographically overlapping sites in the NTT database.

The data were originally compiled from an extensive survey of published and unpublished (e.g., PhD theses) literature, particularly those on tree community surveys and floristic inventories. Additional occurrence records obtained from both major herbaria and taxonomic monographs have been added to the checklists when they were collected within the 10-km diameter of the original NTT site, and within the same vegetation type. NTT does not include sites with an indication of high anthropogenic disturbance nor those with low species richness, because this is often due to low sampling/collecting efforts, which results in poor descriptive power. Thus, secondary forests, which might be considered a distinct vegetation type, are not included in our study. Lowest species richness in the Amazon dataset ranged from 20 species in savanna s.s. and campinarana to 100 in terra firme forest, while plot size (in tree community surveys derived from plot data) ranged from 1 to 5 ha.

All species and their occurrence records were checked for taxonomic circumscriptions and geographical distributions as accepted by the teams of specialists responsible for the online projects Flora do Brasil, Catalogue of the Vascular Plants of Ecuador, Peru Checklist, Bolivia Catalogue (available at http://floradobrasil.jbrj.gov.br/, http://www.tropicos.org/Project/CE/, http://www.tropicos.org/Project/PEC, and http://www.tropicos.org/Project/BC/, respectively) and published floras (Bernal et al., 2016; Boggan et al., 1997; Cardoso et al., 2017; Steyermark et al., 1995–2005). We eventually eliminated records for 111 species due to synonymy (59), invalid or dubious names (7), incorrect growth habit (15), and incorrect distribution (30).

The final dataset contained presence/absence data for 8224 tree species across 1584 sites, with a total of 364,965 presences. Sites derived exclusively from herbarium data represented 41% of the full matrix (654 sites). The dataset also included 24 environmental variables (30 arc-sec resolution) for all its sites, derived from multiple sources. Procedures and protocols concerning variables’ sources and extraction are thoroughly detailed at http://www.neotroptree.info/.

We adopted the vegetation descriptors provided by NTT and based on Oliveira-Filho (2015) to classify the sites into nine vegetation types: 776 terra firme forests, 171 campinaranas, 291 várzeas, 176 igapós, 55 rock outcrops, 36 tepuis, 29 coastal mosaics, 28 savannas sensu stricto (hereafter savanna s.s.), and 22 montane forests (Figure 1; Table S1). All sites classified as tepuis and montane forests occur above 1100 m of altitude (see Study Area), with tepuis differing from montane forests in their rocky soils and dwarfish physiognomy. The map in Figure 1 was designed using the packages maptools (Bivand & Lewin-Koh, 2017) and raster (Hijmans, 2016) in R Statistical Environment (R Core Team, 2018).

The NTT database also includes environmental variables for all its sites, derived from multiple sources (at a 30 arc-second resolution). Altitude at the NTT site center was used as an integrative environmental variable. Variables representing average climate (mean annual precipitation and temperature) as well as climate extremes (e.g., precipitation in driest month) and seasonality (e.g., precipitation seasonality) were obtained from WorldClim 1.4 data layers (Hijmans et al., 2005). Frost frequency (days) and cloud interception (mm) were obtained from interpolating known values as response variables (data obtained from 135 and 57 Brazilian Meteorological Stations measuring frost frequency and cloud interception, respectively) with elevation, latitude, and the WorldClim layers as predicting variables. Soil coarseness (% sand) and soil fertility (% base saturation) and surface rockiness (% exposed rock) were obtained from the Harmonized World Soil Database v1.2 (available at http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/harmonized-world-soil-database-v12/en/) and ranked afterward by mid-class percentage. The use of classes was adopted because high local soil heterogeneity can make raw figures unrepresentative. Soil Water Storage capacity (%) was obtained from the International Soil Moisture Network (available at www.ipf.tuwien.ac.at/insitu/).

2.4 Analyses of community composition

We first explored the patterns of floristic differentiation among previously defined vegetation types by performing non-metric multidimensional scaling (NMDS; McCune & Grace, 2002), and tested its overall significance by applying an analysis of similarities (ANOSIM; Clarke, 1993). Beforehand, we excluded 832 singletons (species found at a single site), as they commonly increase the noise in ordination analyses without contributing information (Lepš & Šmilauer, 2003), and then computed pairwise compositional distances between all sites using Simpson distance as the dissimilarity metric (Simpson, 1960), which describes community turnover without the influence of richness gradients (Baselga, 2010).

We used the vegetation types confirmed in the ordination analysis to produce sets of diagnostic species based on their coefficient of fidelity (phi; Tichý & Chytrý, 2006). An advantage of this coefficient is that they can take negative values, which expresses the fact that a species tends to “avoid” a particular habitat and its environmental conditions (De Cáceres & Legendre, 2009; De Cáceres et al., 2008). In this study, diagnostic species represent those statistically associated with one or more vegetation types so that their presence in species lists may be a strong indicator of the vegetation types themselves. Significance of phi was obtained via Monte Carlo permutations (999). Species’ phi for each vegetation type are provided as Supplementary Information (see Table S2).

We then used the major axes of compositional variation summarized by the ordination analysis to test whether the observed patterns of floristic differentiation in the Amazon are underpinned by extreme environmental conditions segregating terra firme forests from environmentally marginal vegetation types. First, we selected a subset of significant environmental variables for each of the major NMDS axes through an AIC-based forward selection method for generalized linear models and then performed an additional and progressive elimination of collinear variables based on their variance inflation factor (VIF), informed by their ecological relevance, until maintaining only those with VIF < 4 (Quinn & Keough, 2002). We tested the significance of the selected environmental variables by applying ANOVA permutation tests (999 permutations). We explored the results visually by fitting the values of the most important environmental variables in ordination space (NMDS). The variable selection, VIF, NMDS and phi analyses were conducted using the vegan (Oksanen et al., 2016), recluster (Dapporto et al., 2015), usdm (Naimi et al., 2014) and indicspecies (De Cáceres & Legendre, 2009) packages in R Statistical Environment (2018).

Finally, we used a chord diagram to assess the patterns of compositional overlap among Amazonian vegetation types. The chord diagram was designed using the D³ Java Environment (Bostock et al., 2011; custom codes available at https://bl.ocks.org/nbremer).

4.1 Continental-scale patterns

The composition of the tree flora across the Amazon region shows variation congruent with traditional vegetation classifications. The most species rich and geographically widespread vegetation type is terra firme forest, while marginal vegetation types, such as campinarana, savanna s.s., igapó or várzea, diverge in species composition along distinct environmental gradients. These marginal vegetation types house many tree species not found in terra firme forest, yet large numbers of them are shared among the different marginal vegetation types themselves, for example between savanna s.s. and campinarana.

The marginal vegetation types are placed at extreme values of the significant environmental gradients, potentially indicating eco-physiological stress, and our results highlighted that substrate, not climate, is the most important environmental driver controlling the major axes of composition in Amazonian tree communities. Different from other forest biomes in South America, where variation in temperature and water availability are clearly the most important factors controlling continental-scale patterns of tree community composition (e.g., in seasonally dry tropical forests (Neves et al., 2015), or in the Mata Atlântica (Neves et al., 2017)), climatic conditions are relatively unimportant in Amazonia (but see discussion for montane forests and tepuis).

Moreover, because these edaphic gradients are consistently important in segregating Amazonian vegetation types from local to continental scales, our results run counter to previous findings which have suggested that tree community composition in the Amazon is primarily driven by Amazon-wide gradients in environmental conditions (e.g., precipitation seasonality, soil fertility; Silva-Souza & Souza, 2020; ter Steege et al., 2006). Previous Amazon-wide studies analyzed tree species composition data without separating or considering the different Amazonian vegetation types. If composition is summarized within geographic grid cells (e.g., Silva-Souza & Souza, 2020; ter Steege et al., 2006), then a given grid cell may take on the compositional identity of the dominant vegetation type in the grid cell, and if there are geographic gradients in the prevalence of vegetation types, the geographic grid cell approach may lead to geographically driven results, which mask vegetation heterogeneity within grid cells. Our approach ensured that every sample unit (i.e., site or community) represents only a single vegetation type, which is likely why we find a clearer signal for vegetation type than for geography in our results. That these vegetation types are floristically coherent across the Amazon basin also suggests that dispersal among areas of the same vegetation type is not particularly limited by geographic distance, in agreement with a recent study of several Amazonian tree genera (Dexter et al., 2017).

Below we delve into the main floristic patterns observed in our results to discuss the compositional identity and environmental distinctiveness of Amazonian vegetation types. Because the dataset used in this study does not include sites with a high indication of anthropogenic disturbance, we stress that analyses including community inventories (e.g., floristic checklists, plot data) from recently degraded areas, such as early-stage secondary forests, may reveal additional vegetation types.

4.2 Forest types

Flooded forests share a similar environmental condition driving their compositional distinction from terra firme forests: seasonal flooding, potentially combined with soil waterlogging during the low water season. This, however, does not lead to homogeneous stands of flooded forests throughout the Amazon, and one of their main variations was captured here as follows: the floristic, edaphic and distributional differentiation of igapós and várzeas. Várzea forests are more evenly distributed across major river basins in Amazonia, while most igapó forests are concentrated in the Rio Negro and upper Orinoco River Basins where the substrate is of highly leached and impoverished white sands. Nonetheless, igapó forests are also found in other Amazonian regions under similar edaphic conditions (Montero et al., 2014; Wittman et al., 2010). In addition, both types of flooded forests have species restricted to them in our dataset (160 species restricted to várzeas and 168 to igapós), though the largest proportion of their species composition is either shared between them or with terra firme forests (Figure 3; Scudeller, 2018).

The tree flora of montane forests in Amazonia is compositionally coherent with the main floristic patterns described for Neotropical montane flora in general, such as the presence of some genera that are rare to absent in the lowland flora, including Bonnetia, Brunellia, Drimys, Hedyosmum, Ilex, Laplacea, Meriania, Podocarpus, Symplocos and Weinmannia (Webster, 1995). The overall lack of these taxa in lowland Amazonia is likely driven by temperature, an important environmental factor driving floristic differentiation between montane and terra firme forests in our dataset. Nonetheless, variation in temperature across the range and location of elevations sampled in our study is not large, thus explaining the high proportion of tree species shared between montane forests and other Amazonian habitats (88%), and supporting the claim that these forests should be treated as Amazonian (contrasting with views in Cabrera & Willink, 1980; Cardoso et al., 2017). Igapó forests, for instance, share a lower proportion of tree species with other Amazonian habitats (82%), yet igapós are consistently treated as Amazonian (Wittmann et al., 2010). The fact is that many lowland terra firme species do extend their distribution toward high altitudes (1100–3000 m), such as Annona symphyocarpa, Coussapoa crassivenosa, Cyathea bipinnatifida, Cyathea macrosora, Elaeoluma nuda, Hieronyma oblonga, Miconia dodecandra, Miconia pseudocapsularis, Miconia punctata, Mollinedia ovata, Nectandra reticulata and Quiina cruegeriana; to cite a few diagnostic species of both montane and terra firme forests (Table S2).

The scarcity of mineral nutrients in white-sand environments is probably the leading environmental distinction of campinaranas, and plant species there are known to have acquired morphological, physiological and mutualistic traits to maximize both nutrient capture and retention (Adeney et al., 2016). This specialized flora explains much of the differentiation of campinaranas from other vegetation types, which is evident in the high proportion of restricted species (42%; Figure 3). This is almost twice the proportion of endemics in western Amazonian campinaranas found by Garcia-Villacorta et al. (2016; 23%), but this is probably explained by the fact that those authors worked with the whole spectrum of growth habits, and not only trees, considered all available herbarium voucher data (not just those collected near NTT sites as done here) and concentrated their efforts only in western Amazonia.

4.3 Open formations

The coastal sand deposits along the Atlantic shores, covered by a mosaic of mangroves and sandy beaches, represent another Amazonian vegetation type associated with white-sand substrates. However, soils in these coastal mosaics are more fertile than in campinaranas, and they are mostly found in the eastern Amazon, where precipitation seasonality is relatively more pronounced. Nonetheless, these white-sand, seasonally dry coastal environments are not too extensive (Cremers & Hoff, 2003; González, 2011; Silva et al., 2010), nor do they seem to be restrictive floristically, as 94% of species in coastal mosaics are also found in other vegetation types. Accordingly, they have one of the lowest proportions of restricted tree species in Amazonian habitats—only 15 tree species are restricted to coastal sand deposits in this analysis, nine of which are typical of mangroves.

Water deficit intervals, mediated by climate, substrate or both, drive tree community differentiation in two other environments: savannas s.s. and rock outcrops. Most savannas s.s. are found where the dry season is longest in the Amazon, and where fire outbreaks may occur in the dry season. Interestingly, many of these savannas are hyper-seasonal (sensu Sarmiento, 1983), in that they also face some form of water excess in the rainy season, mostly due to soil waterlogging either caused by poor drainage or floods, as in the Bolivian Llanos de Moxos and in the Brazilian estuarine island of Marajó. Nevertheless, there are also non-hyper-seasonal savannas, particularly on hills with shallow soils in the Brazilian state of Pará, where the flora shares a great number of species with that of the Cerrado savannas in Central Brazil (Devecchi et al., 2020). In fact, the tree flora of most Amazonian savannas does show some floristic affinity with the savannas of either the Brazilian Cerrados or the Venezuelan Llanos (Buzatti et al., 2018; Ratter et al., 2006; see also Devecchi et al. 2020 for comparisons between all plant life-forms).

Rock outcrops are another common feature in the Amazon that may experience local water shortage, even in ever-wet areas, because rainwater is promptly drained from the substrate. Rock outcrops are particularly common across inselbergs on both the Guiana and Brazilian Crystalline Shields, where they host tree species not found elsewhere in the Amazon (Gröger, 2000; Raghoenandan, 2000). The xeric nature of rock outcrops is confirmed by the disjunct occurrences of a considerable number of species that are also typical of seasonally dry deciduous forests outside of the Amazon, such as Aspidosperma cuspa, Brasiliopuntia brasiliensis, Bursera simaruba, Cereus hexagonus, Guapira cuspidata, Senegalia riparia and Vachellia farnesiana. Apart from rock outcrops, another substrate related to periods of water deficit in the Amazon is the hardened surface of some mudflats, such as those found on the Guyanese coastal plains and on the meanders of the Lower Amazon and mid-Marañon rivers. This condition probably explains the occurrence of a typical savanna vegetation and flora in these regions, even under year-round wet climates.

The tepuis are formed of the steep slopes and plateaus of the massive Paleozoic sandstones atop the Guiana Shield. Environmental factors along altitudinal gradients are hardly ever easily summarized by one variable and include factors such as increasing cloud interception (an important factor in our models) and declining temperatures toward higher altitudes. Additional sources of environmental heterogeneity include slope, aspect, and surface rockiness, and there is usually a local combination of stressful factors at play. In general, scrubs and savannas prevail on the shallow soils and bare rocks at the summits of tepuis (Huber, 2005), where cloud interception is the major source of water, and are replaced by montane forests in colder environments with deeper soils.

It is important to bear in mind that, different from forest types, the open formations in our analyses do not represent a tree-dominated habitat. Therefore, our comparisons are based on the few tree species that occur in these plant communities, while herbs, forbs and shrubs are not included. If data for the whole plant community were available, we would expect even higher dissimilarities between these open habitats and other tree-dominated habitats, with increased number of endemic species in the former.