Floristic patterns along a 43-km long transect in an Amazonian rain forest
HANNA TUOMISTO*, KALLE RUOKOLAINEN, MELCHOR AGUILAR†‡ and ABEL SARMIENTO†§
Department of Biology, University of Turku, FIN-20014 Turku, Finland, and
†
Universidad Nacional de la Amazonía Peruana, Iquitos, Peru
Summary
1
The floristic variation in Amazonian lowland forests is poorly understood, especiallyin the large areas of non-inundated (tierra firme) rain forest. Species composition may beeither unpredictable as abundances fluctuate in a random walk, more-or-less uniform,or it may correspond to environmental heterogeneity.
2
We tested the three hypotheses by studying the floristic variation of two phylogenet-ically distant plant groups along a continuous 43-km long line transect that crossedtierra firme rain forest in northern Peru.
3
The observed floristic patterns were compared to patterns in the spectral reflectancecharacteristics of the forest as recorded in a Landsat TM satellite image. The topographyof the transect was measured in the field, and surface soil samples were collected todocument edaphic conditions. The two plant groups, pteridophytes and the Melasto-mataceae, were assessed in 2-m wide and 500-m long sampling units.
4
Floristic similarity (Jaccard index) between sampling units ranged from 0.01 to 0.71(mean = 0.27), showing that some units were almost completely dissimilar while otherswere very alike.
5
Spatially constrained clustering produced very similar subdivisions of the transectwhen based separately on satellite image data, pteriophytes, and Melastomataceae, andthe subdivisions were also related to topography and soil characteristics. Mantel testsshowed that floristic similarity patterns of the two plant groups were highly correlatedwith each other and with similarities in reflectance patterns of the satellite image, andsomewhat less correlated with geographical distance.
6
Our results lend no support to the uniformity hypothesis, but they partially supportthe random walk model, and are consistent with the hypothesis that species segregateedaphically at the landscape scale within the uniform-looking forest.
Key-words
: Amazonia, beta diversity, chronological clustering., floristic composition,Mantel test, Melastomataceae, Peru, pteridophytes, random walk model, satelliteimagery, tropical rain forest
Journal of Ecology
(2003)
91
, 743–756
Introduction
In the past few decades, researchers have becomeincreasingly interested in documenting and under-
standing the spatial structure and species compositionof Amazonian lowland rain forests. There are a rangeof different views concerning the main factors thatcontrol plant species distributions in these forests, andthe kind of general distribution patterns that follow.The ecological and floristic differences between suchcontrasting habitats as inundated vs. non-inundatedforests, or forests on podzolized white sand soils vs. non-podzolized soils, are universally recognized. However,researchers differ fundamentally in their views on howspecies are distributed within these broad forest categories,especially within the non-inundated forests on non-podzolized soils, i.e. the ‘typical’ tierra firme rain forest.
*Correspondence: Hanna Tuomisto (e-mail: hanna.tuomisto@ utu.fi). ‡Present address: Consejo Transitorio de AdministracionRegional – Loreto, Av. A. Quiñones Km. 2, Iquitos, Peru§Present address: Instituto Nacional de Recursos Naturales,Carr. Federico Basadre Km. 4.200, Pucallpa, Peru
Much of the current discussion is centred around therelative importance of three alternative views: (i) plantspecies are competitively equal, and the local speciescomposition is in a state of ‘random walk’ as a resultof local immigration and extinction; (ii) the forest isessentially homogeneous, and a small proportion ofthe species are competitively superior and dominatethe forest over wide areas; and (iii) differences in soilswithin the forest are distinct enough to favour differentspecies at different locations, and thus create numerousfloristically differentiated forest patches.
If the plant species in a community come and go atrandom, as in the first view (Hubbell & Foster 1986;Condit 1996; Hubbell 1997, 2001; Chave
et al
. 2002;Condit
et al
. 2002), the variation in abundance of agiven species is expected to show strong spatialautocorrelation due to dispersal limitation, but notto be systematically correlated with the abundances(or even presence) of other species. As a result, a speciesthat is abundant at one site is likely to be both presentand abundant at nearby sites but not at faraway sites,and the overall floristic similarity among sites willdecrease monotonically with increasing inter-sitedistance. This decrease in floristic similarity is expectedto be approximately linearly related to the logarithmof geographical distance (Hubbell 2001; Condit
et al
.2002). The species are more or less equivalent ecolo-gically, so any one of them can become abundant, rare,or locally extinct by chance, and floristic patterns aretherefore not expected to correlate with patterns in localsite conditions. Furthermore, one would not expect tofind sharp floristic boundaries where several speciesappear and/or disappear simultaneously. Instead, spe-cies turnover would be gradual in space.
The second view maintains that species compositionand abundance patterns are relatively constant over wideareas, although only a few species may be shared betweensampling plots if plot size is small (Duivenvoorden1995; Pitman
et al
. 1999, 2001; Terborgh
et al
. 2002),and that, following disturbance, species compositionand abundance revert towards their prior state(Terborgh
et al
. 1996). Most species are expected to bewidespread, and the species that become abundant ata given site are not a random subset of all the speciespresent, but are likely to belong to a limited group ofspecies that possess biological characteristics thatenable them to compete successfully and dominate overlarge tracts of forest (Pitman
et al
. 1999, 2001). Thesedominant species are expected to be omnipresent inthe forest, at least at the landscape scale, so theirabundance patterns are not expected to show eitherspatial autocorrelation or correlation with patterns inlocal site conditions. Because the forest is consideredhomogeneous, no abrupt turnover zones in speciescomposition are expected.
The third view maintains that the environmentalvariation in western Amazonia is pronounced enoughto create floristically differentiated communities withinthe tierra firme forest. Spatial variation in species
composition is expected, both in response to landscape-scale soil differences and in relation to local factors suchas topography and associated soil catenas (Poulsen &Balslev 1991; Tuomisto
et al
. 1995; Ruokolainen
et al
.1997; Svenning 1999; see also Lieberman
et al
. 1985 andClark
et al
. 1995, 1998, 1999 for Central America, andSabatier
et al
. 1997 for Guiana). Species abundancesand floristic composition are expected to reflect spatialpatterns in the environmental conditions, so that if thereis spatial turnover in these patterns, corresponding andpredictable turnover is also expected in the vegetation(Tuomisto & Poulsen 1996, 2000). In a patchy environ-ment, spatial autocorrelation is expected to be a poorpredictor of similarity in species composition becauseenvironmentally, and hence floristically, dissimilar sitescan occur in close proximity (Poulsen & Tuomisto1996; Ruokolainen
et al
. 1997). Because each speciesis expected to be most abundant where the environ-mental conditions are most favourable for it, the samedominant species are expected at sites with similarenvironmental conditions, while different dominantsare expected at sites with differing environmentalconditions (Tuomisto
et al
. 1998).Similar discussion about the detailed variation in
species composition within broadly defined vegetationtypes has been conducted elsewhere in the tropics,especially in south-east Asia, where field results havealso been variously interpreted (Baillie
et al
. 1987).Poore (1968) concluded that, while rare species may behabitat specialists, the distribution of common speciesis determined by biotic interactions and chance. Othershave found evidence for mainly edaphically determineddistribution patterns (Austin
et al
. 1972; Ashton 1976;Baillie
et al
. 1987), or evidence of such patterns insome but not all of the forest types studied (Newbery& Proctor 1984).
-
To test the three hypotheses described above, it isnecessary that field inventories are both extensiveenough and detailed enough to reveal landscape-scalefloristic and edaphic patterns. Extensive tree samplinghas been carried out in western Amazonia (Gentry1988; Duivenvoorden 1995; Ruokolainen
et al
. 1997;Ruokolainen & Tuomisto 1998; Pitman
et al
. 1999,2001; Duque
et al
. 2002), but both sampling andspecies identification of trees are very laborious andtime-consuming, and the number of species involvedis very high, so it is difficult to obtain tree samples thatare both spatially and floristically representative enoughto give a detailed and reliable picture of landscape-scalespecies distribution patterns.
To be able to cover larger spatial extents in moredetail, we have concentrated our sampling effort on twoplant groups that are more easily observable and lessspecies-rich than canopy trees: pteridophytes (fernsand fern allies) and the Melastomataceae (which aremainly shrubs and small trees). In earlier studies, both
groups have been found to show roughly the samefloristic patterns as trees (Ruokolainen
et al
. 1997;Ruokolainen & Tuomisto 1998; Vormisto
et al
. 2000),and hence we call them here indicator groups.
Because our field sampling did not cover trees, wewanted to preclude the possibility that the indicatorgroups chosen conform unduly with each other becauseof phylogenetic relatedness or similarities in lifehistories. We therefore studied groups that are bothphylogenetically remote (pteridophytes vs. angiosperms)and have contrasting life histories and dispersal modes(pteridophytes have wind-dispersed spores and asessile self-supporting gametophyte generation, whileMelastomataceae are predominantly bee-pollinatedand bird-dispersed). Any congruence in species com-position patterns between such dissimilar groups islikely to reflect external factors that would also affectother plant groups.
Because Amazonian rain forests are spatiallyextensive and difficult to access, remotely sensedinformation has been used to help in recognizing plantcommunities and spatial patterns within them. Earlysuccessional forest, inundated forest, different kinds ofswamp and forest on podzolized white sand soil, whichare all clearly structurally different, can be readilyrecognized in satellite images (Kalliola
et al
. 1991,1998; Tuomisto
et al
. 1994; Foody & Hill 1996; Novo &Shimabukuro 1997; Tuomisto 1998; Hill 1999; Saatchi
et al
. 2000). Even non-inundated tierra firme forests onnon-podzolized soil, which mostly look homogene-ous in aerial photographs, show considerable spectralpatchiness in Landsat TM (Thematic Mapper) satelliteimages with variation at scales from hundreds ofmetres to kilometres (Tuomisto
et al
. 1995). It has beendebated whether or not these satellite image patternsindicate differences in the vegetation that are related tosoil differences (Condit 1996; Duivenvoorden & Lips1998), but recent results have indicated that this indeedis the case (Ruokolainen & Tuomisto 1998; Tuomisto &Ruokolainen 2001; Tuomisto
et al
. 2003a).We used a Landsat TM satellite image as a source of
spatially continuous information for landscape-scalevariation in the rain forest. The satellite measuresreflectance of the ground cover, which in this case ismainly determined by the canopy trees, lianas andepiphytes. On the basis of earlier studies (Tuomisto
et al
. 1995; Ruokolainen & Tuomisto 1998; Vormisto
et al
.2000; Tuomisto & Ruokolainen 2001; Tuomisto
et al
.2003a) we propose that canopy patterns follow under-lying edaphic conditions, and hence we use the reflect-ance values from the satellite image as a proxy forenvironmental variation. Soil samples were analysed toverify this relationship, but their spatial resolution wasnot sufficient to include them in the formal analyses.
Note that our indicator groups are understoreyplants, and therefore have hardly any direct influenceon the reflectance characteristics of the forest. Con-sequently, a correlation between understorey speciesdistribution patterns and canopy reflectance patterns
can only be found if the factors that determine thereflectance characteristics of the canopy are stronglycorrelated with those factors that determine thefloristic composition of the understorey. This putsthe hypothesis on environmental control of floristicpatterns under a stringent test, while favouring the accept-ance of the random walk hypothesis and the uniformityhypothesis.
We have recently tested the three hypotheses usingwidely spaced field sampling that ranged from southernPeru to Ecuador and Colombia (Tuomisto
et al
. 2003b).At such a wide spatial scale, the forests were clearly nothomogeneous, and the data indicated that both randomwalk with dispersal limitation, and environmentalfactors are needed to explain floristic patterns. Condit
et al
. (2002) found that the dispersal limitation modelwas sufficient to explain their field data at distancesbetween 0.2 km and 50 km. In the present study, ouraim is to concentrate on this landscape scale, and totest the three hypotheses using data from a continuous43-km long transect.
A single transect was used instead of discrete plotsbecause continuous sampling allows observations ofspatial change to be made and compared between datasets, and assessment of whether turnover is spatiallycontinuous or occurs more rapidly at certain points.A transect can also be georeferenced more readilythan separate plots, because it crosses tree-fall gapswhere GPS (Global Positioning System) coordinatescan readily be obtained. Our field survey extended over43 km of forest, and the analyses of both field data andsatellite image data were based on 500-m long samplingunits. This geographical scale is such that it is able todetect patches of the size recognized by Tuomisto
et al
.(1995), who assessed spectral patchiness along 30-kmlong transects that were drawn on satellite images butnot field-verified. The relatively coarse resolution furthermakes it unlikely that any correlation between patternsin satellite imagery and plant species composition iscaused by ordinary forest succession in tree fall gaps,because these are typically much less than 500 m across.
We also asked how many of the observed plantspecies are actually distributed in a way that correlateswith the reflectance patterns in satellite imagery. Thisquestion was answered by first classifying the samplingunits of the transect on the basis of information fromthe satellite image, and then testing, for each species,whether or not its distribution was biased towards anyof the recognized classes.
Materials and methods
Fieldwork was carried out in Amazonian Peru in theforest reserve of the Amazon Center for Environmental
Education and Research (ACEER) close to the con-fluence of the Sucusari and Napo rivers (Fig. 1). Theclimate in the area is tropical, humid and almostaseasonal. Mean monthly temperature in the nearby cityof Iquitos is 25–27
°
C throughout the year, and annualprecipitation is about 3100 mm. No month receives lessthan 180 mm of rain on average, but about half of theyears for which records exist experienced one or twomonths with less than 100 mm of rain (Marengo 1998).
The area is about 100–200 m above sea level, and thelandscape ranges from flat to hilly. The surface soil isformed of unconsolidated sediments of various originsand ages, including the mid-Miocene Pebas formationand more recent fluvial deposits. The geology of thearea has not been studied in detail, but accounts of thegeological history of the general region do exist (Hoorn1993; Räsänen
et al
. 1998).The vegetation in the study area consists mainly
of closed-canopy non-inundated forests, althoughseasonally or sporadically inundated zones are foundalong all rivers and major creeks. No special edaphicallydefined vegetation types (such as forests on white sandsoils) are known from the non-inundated area.
Soil and floristic studies were conducted in a con-tinuous 43.38-km long transect. The transect followedthe border of the ACEER reserve, which had been markedin the field a few months earlier by a crew of men fromnearby villages, some of whom also joined us on thisfield expedition. The transect was georeferenced usingGPS technology.
A visible mark was fixed every 50 m along the lengthof the entire transect. Sections of 100 m were used assampling subunits for the floristic inventory. Pterido-phytes and members of the family Melastomataceae(excluding Memecylaceae) were censused within anestimated 2.0 m to the left side of the transect. Presence-absence data were collected for each species. Collectingabundance data would have been too slow to bepracticable; as the length of our field expedition wasmainly limited by the quantity of provisions we were
able to carry, we had to compromise between localdetail and spatial extent.
For the purposes of the present paper, five consecut-ive subunits were fused to obtain 87 sampling unitswith an effective size of 0.1 ha (500
×
2 m). The last unitended at the shore of the Apayacu River, and wasshorter than the other units. The vegetation in the firstsampling unit (the one closest to the Sucusari River)showed, in parts, characteristics typical of secondaryforests. Because our purpose was not to study differ-ences between secondary and old-growth forests, butrather variation within old-growth forests, this unit wasexcluded from the analyses presented. The final samplesize was therefore 86 sampling units that coveredalmost 8.6 ha.
To facilitate pteridophyte observation and sampling,only individuals with at least one leaf longer than10 cm were considered, and epiphytic and climbingindividuals whose lowermost green leaves were higherthan 2 m above ground were ignored. Voucher col-lections for all species of both pteridophytes and theMelastomataceae are deposited in herbaria in Peru(AMAZ), Finland (TUR) and USA (KSP; herbariumacronyms follow Holmgren
et al
. 1990).The topographic profile of the transect was measured
using a clinometer (Suunto, Vantaa, Finland). Meas-urements were taken every 50 m, and in between if theslope of the terrain changed significantly. Surface soilsamples (the top 5 cm of the mineral soil) were collectedat roughly 2.5-km intervals, such that two samples weretaken from each location, one from the top of a hill andone from the bottom of the nearest valley. Each of thesoil samples consisted of five pooled subsamplescollected within an area of
c.
5
×
5 m. Physical andchemical analyses were carried out using standardprocedures (van Reeuwijk 1993). We report soil texture(percentage of coarse sand with particle size 0.25–2 mm) and the concentration of exchangeable bases(calcium, potassium, magnesium and sodium measuredin 1
NH
4
OAc at pH 7).A Landsat TM image (path 6, row 62, 1 November
1987) covering the study area was obtained from theLandsat Pathfinder HTF project of the University ofMaryland and NASA, USA. After the field work, thesatellite image was rectified using ground control pointsthat had either been obtained using a GPS in the field,or could be identified on a base map derived from LandsatMSS satellite images (IFG 1984). The transect was drawnon the rectified satellite image with the help of GPS co-ordinates and landmarks.
For each of the 500-m long sampling units, an areaextending 200 m to either side was delimited on thesatellite image. Such a large area was used for tworeasons. First, it reduced the effect of hilliness on theresults, as each unit was large enough to average out thedifferences in reflectance values between the sunlit andshaded sides of hills. Second, the error in GPS co-ordinates at the time of sampling may have exceeded100 m, so a 200-m buffer in pixel sampling was deemed
Fig. 1 Location of the study area in Peruvian Amazonia, andof the 43-km long transect near the confluence of the Napoand Sucusari rivers. The transect first runs 30 km to the east,and then 13 km to the north.
necessary to ensure that the transect was actuallycontained within the sampled area.
The original values for the pixels included withineach of the delimited areas were extracted for analysis.Most of the areas had 211–218 pixels, but four weremade smaller (135–180 pixels) to avoid including pixelswith clouds or cloud shadows. Only data from bands 1–5 and 7 were used, the thermal infrared band 6 beingexcluded. ER-Mapper 5.5 software (Earth ResourceMapping, Egham, UK) was used for all satellite imageanalyses.
Our numerical analyses aimed to reveal areas of majorchanges in pteridophyte and Melastomataceae speciescomposition, and to clarify the extent to which differ-ences in environmental factors (as inferred from pixelvalues in the satellite image) and geographical distancecan be used to predict differences in floristic composi-tion of the forest.
Almost all analyses are based on resemblancematrices, each of which consists of pairwise compar-isons among all 86 sampling units using one or moredescriptor variables. When a resemblance measure thathad originally been calculated as distance (
D
) needed tobe converted to similarity (
S
) or vice versa, the formula
S
= 1
−
D
was used.Floristic similarity matrices were calculated using
the Jaccard index [
S
=
a
/(
a
+
b
+
c
), where
a
is thenumber of species shared between the two samplingunits,
b
is the number of species only found in the firstunit, and
c
is the number of species only found in thesecond unit]. Three similarity matrices were calculated:one using pteridophytes, one using Melastomataceaeand one using both plant groups combined.
Differences in pixel values between sampling unitswere expressed in Euclidean distance using mean pixelvalues calculated separately for each band for eachunit. A total of 11 Euclidean distance matrices werecalculated using one or more of these satellite-derivedvariables. Seven matrices were based on a single variable:either the mean pixel values of one of the six bands, orthe green vegetation index [NDVI = (band 4
−
band 3)/(band 4 + band 3)]. For the remaining matrices, twodifferent combinations of satellite-derived variableswere used [all six bands (1–5 and 7), or band 2, band7 and NDVI]. The purpose of including just threevariables in the latter case was to build a model withreduced collinearity. The visible wavelength bands 1–3are highly intercorrelated, as are the infrared bands 4, 5and 7, so only one band from each group was used inthe reduced model. Bands 2 and 7 were chosen becausethey showed highest correlations with the floristic datain the Mantel test (see below). NDVI was chosen becauseit provides information from bands 3 and 4 but is lesscorrelated with bands 2 and 7 than either band alone.
Both combinations of satellite-derived variableswere first used to produce a distance matrix in which
each of the included variables was given equal weightby standardizing to zero mean and unit variance.However, there is no reason to believe that equal weightswould give an optimal relation to floristic variation,so for both variable combinations, a second distancematrix was constructed where each of the standardizedsatellite-derived variables were weighted individually.The weights were obtained from an equation of multipleregression on distance matrices that was obtained foreach variable combination. The independent distancematrices were in both cases based on the satellite-derived variables, and the dependent distance matrixwas the floristic distance matrix that included bothpteridophytes and Melastomataceae. For both modelswe report the standard partial regression coefficients(
B
) and the coefficients of multiple determination (
R
2
).The partial regression coefficients for each satellite-derived variable were used as weights in calculating thecorresponding Euclidean distance matrix. Backwardelimination was applied in the multiple regression ana-lysis that initially included only band 2, band 7 andNDVI to make sure that each of the variables in thefinal model would have a statistically significant (
P <
0.05 after Bonferroni correction) contribution to theamount of variance explained (see Legendre
et al
. 1994;Legendre & Legendre 1998).
Mantel tests of matrix correspondence were runto analyse the degree of predictability in the floristicpatterns of the sampling units. First, a Mantel test wasrun to quantify the correlation between the floristicdistances as measured separately with the two plantgroups (pteridophytes and the Melastomataceae).Then, Mantel tests were run to find out to what degreethe floristic distances correlated with distances in pixelvalues in the satellite image and with geographicaldistance. All possible pairwise tests involving one ofthe three floristic distance matrices and one of the nineunweighted satellite-derived distance matrices were run.The weighted satellite-based distance matrices were onlyused in one Mantel test, i.e. the one using the combinedpteridophyte and Melastomataceae distance matrixthat had been used in the multiple regression analysisthat provided the weights. All three floristic distancematrices were also used in two Mantel tests involvinggeographical distance: one test used original geographicaldistances, and the other used ln-transformed geographicaldistances.
Partial Mantel tests were run to find out how muchof the correlation between two distance matrices (suchas floristic and satellite-derived) remained after takinginto account the correlation with a third distance matrix(such as geographical).
For each Mantel test and partial Mantel test, wereport the standardized form of the Mantel statistic,which corresponds to a Pearson correlation coefficientcalculated between the two distance matrices inquestion. The statistical significance of each correlationwas determined by a Monte Carlo permutation testusing 999 permutations, which allows testing of the
< 0.001 level for eachindividual correlation. When interpreting the results,it is important to keep in mind that Mantel’s matrixcorrelation coefficient
r
M
is not comparable with thelinear Pearson’s correlation coefficient
r
P
, which is basedon the original variables rather than distances. Whenboth are calculated on the same univariate data,
r
M
obtains clearly lower values than
r
P
, although theygenerally show the same degree of statistical significance(Legendre 2000).
Cluster analyses were carried out to classify thesampling units on the basis of their floristic similarity(on the basis of pteridophytes, Melastomataceae, andthe two combined) and on the basis of their similarityin satellite image pixel values (using the matrix that gavethe highest Mantel correlation with the floristic matrix).Two different agglomerative clustering methods wereapplied, both of which use a proportional-link linkagealgorithm. The connectedness level was always set to0.5 (i.e. mid-way between single and complete linkage).
One of the clustering methods, chronological clus-tering, applies a unidimensional constraint and onlyallows the fusion of sampling units or groups of units ifthey are contiguous along the transect (Legendre
et al
.1985; Legendre & Legendre 1998). This method givesnon-hierarchical results, because the clusteringprocedure performs at each step a permutational test ofsignificance and decides according to the user-definedalpha significance level whether or not to fuse thetwo groups under consideration. When two groups aremaintained separate, the probability that they actuallybelong to the same random population is less than thevalue of alpha used. In the present study, three runswere made with each similarity matrix using alphasignificance levels 0.001, 0.01 and 0.05.
The second clustering method is hierarchical, andallows the fusion of sampling units irrespective of theirposition along the transect. This method was used toobtain a hierarchical classification of the clusters thathad been recognized in the chronological clustering ofthe satellite image data at alpha significance level 0.01.
To test whether the distribution patterns of theindividual plant species are related to the clustersobtained from the hierarchical classification of thesatellite image, we computed the indicator values ofDufrêne & Legendre (1997) for all species that wererecorded in five or more of the 86 sampling units. Theindicator value combines into a single index twoproperties of a plant species’ distribution: specificityand fidelity. Specificity of a species to a given cluster isthe proportion of its occurrences that are within thatcluster, and fidelity is the proportion of sampling unitsin that cluster that contain the species. The two valuesare multiplied by each other and by 100 to yield indexvalues in percentages. Indicator values were calculatedfor each of the clustering levels that contained betweenone and eight clusters, and the statistical significance ofeach indicator value was determined by a Monte Carlotest using 999 permutations.
Principal coordinates analysis on the floristicdistance matrix (pteridophytes and Melastomataceaecombined) was conducted to visualize the floristicrelationships between the 86 sampling units in anordination diagram.
The R-package was used to compute the resemblancematrices and to run the Mantel tests, cluster analysesand ordination analysis (version 3 was used for chron-ological clustering, version 4 for the other analyses).The multiple regression analyses were run with Permute!3.4 (alpha version). The indicator values were computedusing the program IndVal 2. All these programs are avail-able through the web site http://www.fas.umontreal.ca/BIOL/Legendre/indexEnglish.html.
Results
In general, two kinds of landscape were found withinthe transect: low lying areas with small hills, and a moreelevated area with higher hills. A part of the latter waslike a plateau with deeply dissected creek valleys. Soilanalyses showed that the soils in the low lying areascontained a smaller proportion of coarse sand butmore exchangeable bases than soils in the more hillyareas (Fig. 2a).
A total of 221 species were recorded in the 43-kmlong transect (130 pteridophytes and 91 Melastomata-ceae). The number of species within a 500-m longsampling unit ranged from 9 to 41 for pteridophytesand from 10 to 25 for Melastomataceae (Fig. 2b,c).
Floristic similarity (Jaccard index, calculated usingboth plant groups) between sampling units ranged from0.01 to 0.71, the mean being 0.27. The results obtainedwhen both plant groups were taken separately wereremarkably similar. Jaccard index values ranged from 0to 0.75 (mean 0.27) for pteridophytes and from 0 to0.80 (mean 0.26) for Melastomataceae. These valuesshow that some of the units were floristically completelydifferent while others were highly similar.
The actual similarity patterns for the two plant groupswere also very similar. The matrix correlation computedbetween pteridophytes and Melastomataceae was0.70 (Mantel test,
P
< 0.001). When the effect of geo-graphical distance on this correlation was partialledout, the correlation coefficient decreased only slightly(
r
M
= 0.68 with linear geographical distance;
r
M
= 0.65with ln-transformed geographical distance,
P
< 0.001in both cases). Partialling out the correlation withsatellite-derived distances had a more notable effecton the fern–Melastomataceae correlation (
r
M
= 0.62with bands 1–5 and 7,
r
M
= 0.60 with band 2, band 7and NDVI,
P
< 0.001 in both cases).The Mantel tests involving the distance matrices
based on satellite-derived variables (bands 1–5 and 7,
NDVI) gave rather similar correlation patterns withpteridophytes, Melastomataceae and the combinedfloristic data (Table 1). In all cases, the matrix correla-tions were statistically highly significant. The singlesatellite-derived variable that showed the highest
matrix correlation with all three floristic data sets wasband 7, whose correlations with the floristic distancematrices were about as high as those obtained wheninformation from all bands was combined butunweighted. Even higher correlations were obtained
Fig. 2 Spatial patterns along a 43-km long transect in Peruvian Amazonia. (a) Topographical profile of the transect, and somesoil characteristics measured from 36 soil samples. Species richness of (b) pteridophytes and (c) Melastomataceae in 500-m longunits of the transect. Chronological clustering of 500-m long units of the transect based on either (d) satellite image pixel valuesor species composition of (e) pteridophytes and the Melastomataceae combined, (f) pteridophytes only or (g) Melastomataceaeonly. Thin vertical lines emphasize the limits between transect sections obtained with α = 0.01 in (d). Hierarchical clustering ofthese sections is shown in Fig. 3.
Table 1 Matrix correlations as measured with the Mantel test between differences in floristic composition and pixel values of aLandsat TM satellite image in Peruvian Amazonia. Euclidean distance was used to measure differences in pixel values. TheJaccard index was used to measure floristic similarity, and this was converted to a distance measure before performing the Manteltest. Statistical significance of each correlation coefficient was calculated using 999 permutations; ***P < 0.001, **P < 0.01. Nosignificance values are reported for the Mantel tests involving weighted distance matrices, as the only purpose of running thesetests was to check if weighting improved the correlation coefficient
when the satellite-derived distance matrix included theunweighted band 2, band 7 and NDVI.
The multiple regression tests, where the distancematrices based on each of the satellite-derived variableswere used to explain the variation in the combinedfloristic distance matrix, produced the following models:
(−0.17 × band 1) + (0.42 × band 2) − (0.04 × band 3) + (0.12 × band 4) − (0.19 × band 5) + (0.49 × band 7) [R2 = 0.34, P < 0.001]
(0.29 × band 2) + (0.34 × band 7) + (0.14 × NDVI) [R2 = 0.33, P < 0.001]
In an attempt to optimize the Euclidean distancematrices, the partial regression coefficients from thesemodels were used to weight each of the satellite-derivedvariables before calculating the corresponding weightedEuclidean distances. It was indeed found in bothcases that higher correlations with the floristic distancematrix were obtained using the weighted than using theunweighted satellite-derived matrix (Table 1).
To test whether floristic distances between samplingunits can be predicted more accurately using satelliteimage reflectance patterns or geographical distances, aseries of Mantel tests and partial Mantel tests wererun. It is clear from the results in Table 2 that satelliteimages were superior in this respect. Geographicaldistances yielded Mantel correlations between 0.25 and0.39, while the satellite-based distance matrices yieldedcorrelations between 0.42 and 0.55. Furthermore, inthe partial Mantel tests the satellite-derived distancematrices had a larger decreasing effect on the correla-tions between geographical and floristic distancesthan the geographical distance matrices had on thecorrelations between satellite-derived and floristicdistances. The ln-transformed geographical distancesshowed clearly more explanatory power than the lineargeographical distances.
The satellite-derived matrix that yielded the highestMantel correlation coefficient in Table 1, i.e. theweighted matrix with band 2, band 7 and NDVI, waschosen to be used in the cluster analyses. To check therobustness of the results, chronological clusteringat α = 0.01 was performed using all four Euclideandistance matrices that combined several satellite-derivedvariables. The differences between the outcomes wereminor, so using one of the other matrices for the rest ofthe analyses would not have changed the overall results.Because differences in the Mantel test results weresmall when only pteridophytes, only Melastomataceae,or both plant groups were used, it was consideredunnecessary to repeat the multiple regression tests withpteridophytes only and Melastomataceae only.
The chronological clustering based on the satelliteimage data on the one hand and the three floristic datasets on the other were very similar (Fig. 2d–g). A breakpoint was invariably found at 23 km, whicheversimilarity index and whichever alpha significance levelwas used. Additional statistically strongly supported breakpoints were found at approximately 5 km, 8 km, at either15 km or 17 km, and between 36 km and 38 km. The exactpositions of the break points between 36 km and 38 kmvaried, indicating that the corresponding change wasgradual rather than abrupt, but still distinct enough tobe apparent in all data sets and at all alpha levels. Onlyone of the break points that were strongly supported byall three floristic data sets, namely that at about 21 km,was not apparent in the satellite image data.
The clustering hierarchy of the eight transect sectionsthat had been obtained from the chronological cluster-ing of the satellite image data (Fig. 2d, alpha significancelevel of 0.01) is shown in Fig. 3. The figure also shows foreach clustering level the strong indicator species, definedhere as those plant species whose indicator valueswere both 30% or higher and statistically significant
Table 2 Matrix correlations as measured with the Mantel test and partial Mantel test between differences in floristic composition,differences in pixel values of a Landsat TM satellite image, and geographical distance and its logarithm in Peruvian Amazonia.Euclidean distance was used to measure differences in pixel values. Jaccard index was used to measure floristic similarity, and thiswas converted to a distance measure before performing the Mantel test. Statistical significance of each correlation coefficient wascalculated using 999 permutations; ***P < 0.001
Fig. 3 Hierarchical clustering of a 43-km long transect in Peruvian Amazonia on the basis of satellite image pixel values. Therightmost branches in the clustering correspond to the eight transect sections in Fig. 2(d) that are marked with similar shading.The strong indicator plant species are shown for the different clustering levels; the species names are displayed next to the clusterfor which they are indicators. The leftmost occurrence of each species name is given in full, other occurrences are abbreviated.Indicator values are in parenthesis, and the highest indicator value for each species is in bold.
(P < 0.05). Out of the 84 pteridophyte species thatoccurred in at least 5 of the 500-m sampling units, 61(73%) had a statistically significant indicator valuefor at least one of the classification levels, and out of the58 Melastomataceae species that occurred in at least5 units, 49 (84%) did.
Different shadings are used in Figs 2(d) and 3 toindicate the four most spectrally distinct classes of sam-pling units. It was found that each of these four classeshad at least one strong indicator species for each plantgroup, and most had several. The highest number ofstrong indicator species (30 pteridophytes and 25 Mela-stomataceae) was found at the two-class level, whichcontrasts the plateau-like area at km 23–38 with the restof the transect (compare Figs 2 and 3). This separationis well defined in the chronological clustering based onboth satellite-derived data (Fig. 2d) and floristic com-position data (Fig. 2e–g). The four-class level, althoughnot emerging as consistently in the different chronolo-gical clusterings, has almost as many indicator species(27 pteridophytes and 27 Melastomataceae).
The distribution of the significant indicator valuesis different for the two plant groups. Pteridophytesclearly have more indicator species for the richer-soilareas with gentle topography than for poorer-soil areaswith hilly topography, whereas Melastomataceaeindicator species are more evenly distributed betweenthe two types of terrain. This pattern is consistent withthe observation that species richness of pteridophytesis clearly at its lowest in the hilly area, while speciesrichness of the Melastomataceae is more evenly dis-tributed along the transect (Fig. 2b,c).
A comparison between the floristic ordination of the86 sampling units (Fig. 4a) with the topographic pattern(Fig. 2a) indicates that the main floristic gradientcorresponds with the terrain gradient. Sampling unitsfrom terrain with small hills and fine grained, cation-rich soils are found towards the left of the ordination,and those from high hills with coarser grained, cation-poorer soils are found towards the right. Even though the
ordination is based only on floristic data, the spectralclasses (see Figs 2d and 3) form distinct groups. Spectralclass 4 is almost separated from the other three classesalong axes 1 and 2, and even more so when axis 3 is takeninto account. The other three classes segregate from eachother mainly along axis 3 (Fig. 4b). The agreement is evenmore remarkable because the first step in the satellite-based classification was spatially constrained, while thefloristic ordination did not involve any constraint.
Discussion
The random walk model predicts that species turnoverin space is gradual and that the species occurrencepatterns of independent plant groups are not correlatedeither with each other or with external factors otherthan geographical distance. These predictions are notsupported by our results. Species composition alongthe transect showed distinct turnover zones betweenmore homogeneous stretches and the turnover zoneswere very similar for pteridophytes, Melastomataceaeand the satellite image. Furthermore, the floristicpatterns were highly correlated between the two plantgroups and with satellite image patterns, even when theeffect of geographical distance was taken into account.
The third prediction of the random walk model, thatfloristic similarity among sites decreases linearly withthe logarithm of geographical distance, was partlysupported by our results. The highly significantMantel correlations between floristic and geographicaldistances are in agreement with this prediction, and thelogarithmic distance model gave a clearly better fit thanlinear distance. However, some sampling units fromdistant parts of the transect were floristically moresimilar to each other than to intervening units (Fig. 4),which is contrary to the prediction. Partial Mantel testssuggested that such discrepancies may be explainedby the effect of environmental patchiness: taking into
Fig. 4 Principal coordinates ordination of 500-m long units of a 43-km long transect in Peruvian Amazonia. The ordination isbased on the species composition of pteridophytes and Melastomataceae. Allocation to spectral classes is based on satellite imagepixel values, and the inset shows how sampling units belonging to each of the classes are referred to in the text (numbers), in thisordination diagram (symbols) and in Figs 2(d) and 3 (lines with shading).
account satellite image patterns reduced the correla-tion between geographical distances and floristicpatterns substantially, although the correlations involv-ing ln-transformed geographical distances remainedstatistically significant.
Condit (1996) has proposed that much of Amazoniamay be ecologically homogeneous enough for therandom walk model to operate at regional scales. In amore recent study, Condit et al. (2002) found that therandom walk model explains the tree species distri-butions in data from Amazonian Peru and Ecuadorfairly well at distances between 0.2 km to 50 km. Theobvious contradiction with our results may stem fromat least two different sources. First, Condit et al. (2002)did not test the floristic pattern they observed againstthe alternative hypothesis that edaphic or other envir-onmental factors may influence species distributions.Environmental conditions usually present high spatialautocorrelation, so the presence of spatial autocorre-lation alone is not sufficient to prove that environmentalfactors are not playing a role for species distributions(cf. Borcard et al. 1992). Second, it is possible that rainforest tree species distributions are not controlled bythe same factors as the distributions of pteridophytesand Melastomataceae. Duque et al. (2002) reportedthat the floristic composition of canopy trees showedweaker correlation with environmental factors thandid the floristic composition of understorey trees,which may indicate that large trees are less sensitive toenvironmental effects than smaller plants. However,the result may also be due to the small number of largetrees present in each study plot, which would increasethe effect of sampling error on the results.
Several studies have reported that patterns in thespecies composition of pteridophytes and Melasto-mataceae correlate − independently of geographicaldistance − with patterns in the species composition oftrees in Peruvian rain forests (Tuomisto et al. 1995;Ruokolainen et al. 1997; Ruokolainen & Tuomisto1998; Vormisto et al. 2000). Furthermore, in the presentstudy the floristic patterns of both pteridophytes andthe Melastomataceae were significantly correlatedwith the reflectance patterns in the satellite image, andthe indicator value analyses showed that distributionsof most species were highly linked to a reflectancebased classification of the transect. Because both plantgroups are mainly understorey plants that have littleeffect on the reflectance characteristics of the forest,such results are only possible if the forest canopy(consisting of trees, lianas and epiphytes) also showssimilar patterns. Whether the reflectance patterns arecaused by spatial pattern in the floristic composition ofthe forest canopy or in some structural or physiologicalproperties is not known at present. More detailedstudies on the degree of congruence among speciesdistributions of different plant life forms are needed toclarify these questions.
The random walk model has been developed forhomogeneous habitats (Hubbell 1997, 2001). The main
issue is to define which are the appropriate scales ingeographical, environmental, and temporal variationfor the model to operate. In fact, the conceptuallyrelated carousel model (van der Maarel & Sykes 1993)maintains that random turnover in species composi-tion is observed at small spatial scales over short timeintervals, but over longer time intervals the cumulativespecies list for each microsite converges towardsthe habitat-specific species pool as the species rotatebetween microsites. The carousel model thus empha-sizes recurrent patterns, and it has successfully (and,apparently, without major controversies) been appliedto herbaceous plant communities in temperate areas(Palmer & Rusch 2001 and references cited therein). Itmay well be that such a carousel would become evidentfrom the apparently random dynamics of tropical treecommunities, if a large enough number of trees wereobserved for a long enough time period.
The regional homogeneity model (Pitman et al. 1999,2001; Terborgh et al. 2002) predicts that speciescomposition is essentially the same over wide areas andshows no abrupt turnover zones, and that the same(dominant) species are found everywhere. Our resultsdo not support any of these predictions. The floristicsimilarity among our field samples ranged from 0.0 tomore than 0.7 (Jaccard index), distinct turnover zoneswere found with both pteridophytes and the Melasto-mataceae, and the indicator value analyses showed thatthe occurrence of most species of both plant groupswere significantly related to a satellite image basedclassification of the transect. Even though we did notrecord species abundance in the present study, it wasobvious in the field that the floristically distinct units ofthe transect were characterized by different dominantspecies. This has also been found in earlier studies(Tuomisto & Poulsen 1996; Tuomisto et al. 1998).
A possible source of difference between our resultsand those of Pitman et al. (1999, 2001) and Terborghet al. (2002) is the efficiency of sampling. With treesampling, a constant problem is the low number of indi-viduals recorded per species, which makes it difficultto judge whether differences among sampling unitsare due to real floristic differences or to low samplingintensity. The resultant loss of power in unravellingdistribution patterns of individual tree species hasbeen discussed, for example, by Clark et al. (1999). Theadvantage of using small plants for this kind of aninventory is that a high number of individuals isobserved per species, so species are unlikely to be absentfrom a particular sample just because too few indi-viduals are observed. In the present study, we recorded 221species, and densities that we have measured elsewherein the region suggest that these represent over 55 000individuals of pteridophytes and 13 000 individualsof Melastomataceae. This contrasts strongly with thecommon situation in tree surveys, where the number of
individuals recorded is usually much lower but thenumber of species is higher.
The conclusion that floristic differences exist withinthe transect is probably valid also for other plantgroups, including trees, as discussed above. However,our more specific results, such as the range of floristicsimilarity values or the proportion of species that shownon-random distribution patterns, are not directly trans-ferable to other plant groups. This is because there is atendency, at least among Amazonian trees and palms,that large statured species are more wide-spread bothgeographically and ecologically than small staturedspecies (Ruokolainen & Vormisto 2000; Ruokolainenet al. 2002). Therefore, canopy trees might present asmaller proportion of species with significant indicatorvalues for the satellite image patterns than did pterido-phytes and Melastomataceae. On the other hand, canopytrees have a direct effect on radiance values in thesatellite image, which may increase the proportion ofspecies with significant indicator values.
The edaphic patchiness model predicts that spatialvariation in floristic composition reflects spatialpatterns in the environmental conditions, that sharpspecies turnover occurs where the environmentalconditions change, that different (dominant) speciesare found at sites with different environmental condi-tions, and that the distribution patterns of independentplant groups are correlated. All these predictions wereconsistent with our results.
On the basis of visual analysis of satellite images,Tuomisto et al. (1995) estimated that the average patchsize in tierra firme forests of Peruvian Amazonia isslightly less than 5 km, and that a 30-km long randomtransect crosses, on average, four different kinds ofpatch. These results agree very well with our presentresults: on the basis of the pixel values in the satelliteimage, our 43-km long transect was divided into eightsections (at a significance level 0.01), which translatesinto an average patch size of slightly more than 5 km.Clustering of these, when interpreted together with anindicator species analysis, suggested the recognition offour different kinds of patch, even though the transectonly covered tierra firme forest and excluded flood-plains, swamps and other distinct vegetation types thatwere included in the earlier analysis.
Unlike Tuomisto et al. (1995), we now have groundtruth data to test the ecological relevance of the satelliteimage patterns. When our transect was subdividedusing floristic data of pteridophytes and the Melasto-mataceae, very similar patterns emerged both betweenthe two plant groups and when these were compared tothe patterns obtained from the satellite image data. Inaddition, a great majority of the plant species yieldedstatistically significant indicator values for the clustersbased on satellite image data, confirming that thesatellite image reflects ecological characteristics that
are important for the distribution patterns of plants.As different geological formations often show differ-
ent denudation patterns, the differences in topographyalong our transect suggest that different geologicalformations yield the soil parent material in differentparts of the transect. Soil samples confirmed that atleast soil texture and cation content vary accordingly.Even though the soil data along our transect are tooscarce to be analysed at the same degree of detail as thefloristic and satellite image data, they are consistentwith the hypothesis that the observed floristic patternsare caused by differences in soils. These results arein agreement with results obtained in EcuadorianAmazonia at the same spatial scale but using scattered500-m long transects with more detailed soil sampling(Tuomisto et al. 2002, 2003a).
The sampling unit size used in the satellite imageanalysis is rather large (500 × 400 m, or 20 ha), andtherefore each unit is in itself a mosaic of different localconditions, of which the variability in topography isobvious from Fig. 2(a). Generally it can be expectedthat the more heterogeneous the sampling unit, themore difficult it will be to unravel correlations betweenspecies distribution patterns and environmental condi-tions (e.g. Palmer & Dixon 1990). In spite of this, all theanalyses we conducted pointed to the result that speciesdistribution patterns are correlated with patterns insatellite imagery and terrain characteristics. In spite oflocal variability, the mean conditions therefore seemsufficiently different for species to segregate at thelandscape scale.
Conclusions
In an earlier study at a wider spatial scale, we found thatthe homogeneity hypothesis gained no support, andthat both the random walk model and the ecologicalspecialization model were needed in order to under-stand floristic patterns in western Amazonian tierrafirme forests (Tuomisto et al. 2003b). Our presentresults show that the same is true at the landscape scale.
When considering to what extent the results of thepresent study can be generalized, it is important toknow how representative the observed degree ofecological heterogeneity is. Two points indicate thatour present study area was not exceptionally hetero-geneous for western Amazonian tierra firme. First, theearlier satellite image analysis of Tuomisto et al. (1995)found that the area is one of the most homogeneousparts of Peruvian Amazonia. Second, the Mantelcorrelation coefficients we obtained between floristicdistances and satellite-based distances in the presentstudy were comparable to similar correlations obtainedin Amazonian Ecuador in an area that has earlier beenconsidered homogeneous (Tuomisto et al. 2003a).
The finding that non-obvious but distinct floristicand edaphic variation exists within tierra firme rainforests at the landscape scale has important practicalimplications. In ecological research, we should be
cautious about extrapolating field results from onesite to other sites under the assumption that tierra firmerain forests are all the same. In biodiversity conserva-tion, a high degree of habitat heterogeneity implies anincreased need for wide-scale information on speciesdistribution and endemism patterns to better assesswhere the different habitats are, which species theyharbour, and where conservation efforts should beconcentrated.
Acknowledgements
We thank ACEER and INRENA for the permissionsto work in the area, and Explorama Tours (especiallyPeter Jenson) for practical help in preparing the fieldwork. The crew of men from nearby villages providedinvaluable assistance during the making of the transect.The Landsat Pathfinder HTF project of the Universityof Maryland and NASA provided the satellite image.Soil analyses were done at the Agricultural ResearchCentre of Finland. The herbaria in Iquitos (AMAZ)and Turku (TUR) provided facilities, and Alan R. Smithand Robbin C. Moran helped with species identifica-tions. Daniel Borcard and Pierre Legendre providedconstructive comments on the manuscript and dataanalyses.
References
Ashton, P.S. (1976) Mixed Dipterocarp forest and its variationwith habitat in the Malayan lowlands: a re-evaluation atPasoh. Malaysian Forester, 39, 56–72.
Austin, M.P., Ashton, P.S. & Greig-Smith, P. (1972) Theapplication of quantitative methods to vegetation survey.III. A re-examination of rain forest data from Brunei.Journal of Ecology, 60, 305–324.
Baillie, I.C., Ashton, P.S., Court, M.N., Anderson, J.A.R.,Fitzpatrick, E.A. & Tinsley, J. (1987) Site characteristicsand the distribution of tree species in Mixed DipterocarpForest on Tertiary sediments in central Sarawak, Malaysia.Journal of Tropical Ecology, 3, 201–220.
Borcard, D., Legendre, P. & Drapeau, P. (1992) Partialling outthe spatial component of ecological variation. Ecology, 73,1045–1055.
Chave, J., Muller-Landau, H.C. & Levin, S.A. (2002) Compar-ing classical community models: theoretical consequencesfor patterns of diversity. American Naturalist, 159, 1–23.
Clark, D.B., Clark, D.A. & Read, J.M. (1998) Edaphicvariation and the mesoscale distribution of tree species in aneotropical rain forest. Journal of Ecology, 86, 101–112.
Clark, D.A., Clark, D.B., Sandoval, R.M. & Castro Vinicio, M.C.(1995) Edaphic and human effects on landscape-scaledistributions of tropical rain forest palms. Ecology, 76,2581–2594.
Clark, D.B., Palmer, M.W. & Clark, D.A. (1999) Edaphicfactors and the landscape-scale distributions of tropicalrain forest trees. Ecology, 80, 2662–2675.
Condit, R. (1996) Defining and mapping vegetation typesin mega-diverse tropical forests. Trends in Ecology andEvolution, 11, 4–5.
Condit, R., Pitman, N., Leigh, E.G. Jr, Chave, J., Terborgh, J.,Foster, R.B., Nuñez, V.P., Aguilar, S., Valencia, R., Villa, G.,Muller-Landau, H.C., Losos, E. & Hubbell, S.P. (2002) Beta-diversity in tropical forest trees. Science, 295, 666–669.
Dufrêne, M. & Legendre, P. (1997) Species assemblagesand indicator species: the need for a flexible asymmetricalapproach. Ecological Monographs, 67, 345–366.
Duivenvoorden, J.F. (1995) Tree species composition and rainforest-environment relationships in the middle Caquetáarea, Colombia, NW Amazonia. Vegetatio, 120, 91–113.
Duivenvoorden, J.F. & Lips, J.M. (1998) Mesoscale patternsof tree species diversity in Colombian Amazonia. ForestBiodiversity in North, Central and South America, and theCaribbean: Research and Monitoring (eds Dallmeier, F. &Comiskey, J.A.), pp. 535–549. Man and the Biosphere Series,Vol. 21. UNESCO, Paris and The Parthenon PublishingGroup, New York, USA.
Duque, A., Sanchez, M., Cavelier, J. & Duivenvoorden, J.F.(2002) Different floristic patterns of woody understoreyand canopy plants in Colombian Amazonia. Journal ofTropical Ecology, 18, 499–525.
Foody, G.M. & Hill, R.A. (1996) Classification of tropicalforest classes from Landsat TM data. International Journalof Remote Sensing, 17, 2353–2367.
Gentry, A.H. (1988) Changes in plant community diversityand floristic composition on environmental and geograph-ical gradients. Annals of the Missouri Botanical Garden, 75,1–34.
Hill, R.A. (1999) Image segmentation for humid tropicalforest classification in Landsat TM data. InternationalJournal of Remote Sensing, 20, 1039–1044.
Holmgren, P.K., Holmgren, N.H. & Barnett, L.C. (1990)Index Herbariorum. Part I: The Herbaria of the World. NewYork Botanical Garden, New York.
Hoorn, C. (1993) Geología del nororiente de la Amazoníaperuana: la Formación Pebas. Amazonía Peruana. VegetaciónHúmeda Tropical en el Llano Subandino (eds R. Kalliola,M. Puhakka and W. Danjoy), pp. 69–85. PAUT and ONERN,Jyväskylä, Finland.
Hubbell, S.P. (2001) The Unified Neutral Theory of Biodiversityand Biogeography. Monographs in Population Biology 32.Princeton University Press, USA.
Hubbell, S.P. (1997) A unified theory of biogeography andrelative species abundance and its application to tropicalrain forests and coral reefs. Coral Reefs, 16 (Suppl.), S9–S21.
Hubbell, S.P. & Foster, R.B. (1986) Biology, chance, andhistory and the structure of tropical rain forest treecommunities. Community Ecology (eds J. Diamond andT.J. Case), pp. 314–329. Harper & Row, New York, USA.
IFG. (1984) Mapa Planimétrico de Imágenes de Satélite1:250 000 (Perú). Institute for Applied Geosciences, NeuIsenburg, Germany.
Kalliola, R., Puhakka, M., Salo, J., Tuomisto, H. &Ruokolainen, K. (1991) The dynamics, distribution andclassification of swamp vegetation in Peruvian Amazonia.Annales Botanici Fennici, 28, 225–239.
Kalliola, R., Ruokolainen, K., Tuomisto, H., Linna, A. &Mäki, S. (1998) Mapa Geoecológico de la zona de Iquitos yvariación medioambiental. Geoecología y DesarrolloAmazónico: Estudio Integrado en la Zona de Iquitos, Perú(eds R. Kalliola and S. Flores Paitán), pp. 443–457.Annales Universitatis Turkuensis Ser A II 114. Universityof Turku, Finland.
Legendre, P. (2000) Comparison of premutation methods forthe partial correlation and partial Mantel tests. Journal ofStatistical Computation and Simulation, 67, 37–73.
Legendre, P., Dallot, S. & Legendre, L. (1985) Succession ofspecies within a community: chronological clustering,with applications to marine and freshwater zooplankton.American Naturalist, 125, 257–288.
Legendre, P., Lapointe, F. & Casgrain, P. (1994) Modelingbrain evolution from behavior: a permutational approach.Evolution, 48, 1487–1499.
Legendre, P. & Legendre, L. (1998) Numerical Ecology.Second English Edition. Developments in EnvironmentalModelling, 20, 1–853.
Lieberman, M., Lieberman, D., Hartshorn, G.S. & Peralta, R.(1985) Small-scale altitudinal variation in lowland wettropical forest vegetation. Journal of Ecology, 73, 505–516.
van der Maarel, E. & Sykes, M.T. (1993) Small-scale plantspecies turnover in a limestone grassland: the carouselmodel and some comments on the niche concept. Journal ofVegetation Science, 4, 179–188.
Marengo, J.A. (1998) Climatología de la zona de Iquitos,Perú. Geoecología y Desarrollo Amazónico: Estudio Integradoen la Zona de Iquitos, Perú (eds R. Kalliola and S. FloresPaitán), pp. 35–57. Annales Universitatis Turkuensis Ser AII 114. University of Turku, Finland.
Newbery, D.McC. & Proctor, J. (1984) Ecological studies infour contrasting lowland rain forests in Gunung Mulunational park, Sarawak. IV. Associations between treedistribution and soil factors. Journal of Ecology, 72, 475–493.
Novo, E.M. & Shimabukuro, Y.E. (1997) Identification andmapping of the Amazon habitats using a mixing model.International Journal of Remote Sensing, 18, 663–670.
Palmer, M.W. & Dixon, P.M. (1990) Small-scale environmentalheterogeneity and the analysis of species distributions alonggradients. Journal of Vegetation Science, 1, 57–65.
Palmer, M.W. & Rusch, G.M. (2001) How fast is the carousel?Direct indices of species mobility with examples from anOklahoma grassland. Journal of Vegetation Science, 12,305–318.
Pitman, N.C.A., Terborgh, J., Silman, M.R., Nuñez, V.P.(1999) Tree species distributions in an upper Amazonianforest. Ecology, 80, 2651–2661.
Pitman, N.C.A., Terborgh, J., Silman, M.R., Nuñez, V.P.,Neill, D.A., Cerón, C.E., Palacios, W.A. & Aulestia, M. (2001)Dominance and distribution of tree species in upperAmazonian terra firme forests. Ecology, 82, 2101–2117.
Poore, M.E.D. (1968) Studies in Malaysian rain forest. I.The forest on Triassic sediments in Jengka forest reserve.Journal of Ecology, 56, 143–196.
Poulsen, A.D. & Balslev, H. (1991) Abundance and cover ofground herbs in an Amazonian rain forest. Journal ofVegetation Science, 2, 315–322.
Poulsen, A.D. & Tuomisto, H. (1996) Small-scale to continentaldistribution patterns of neotropical pteridophytes: the roleof edaphic preferences. Pteridology in Perspective (edsJ.M. Camus, M. Gibby, and R.J. Johns), pp. 551–561. RoyalBotanical Gardens, Kew, London.
Räsänen, M., Linna, A., Irion, G., Rebata Hernani, L.,Vargas Huamas, R. & Wesselingh, F. (1998) Geología ygeoformas de la zona de Iquitos. Geoecología y DesarrolloAmazónico: Estudio Integrado en la Zona de Iquitos, Perú(eds R. Kalliola and S. Flores Paitán), pp. 59–137. AnnalesUniversitatis Turkuensis Ser A II 114. University of Turku,Finland.
van Reeuwijk, L.P. (1993) Procedures for Soil Analysis.Technical Paper 9, 4th edn. International Soil Referenceand Information Centre, Wageningen, the Netherlands.
Ruokolainen, K., Linna, A. & Tuomisto, H. (1997) Use ofMelastomataceae and pteridophytes for revealing phyto-geographic patterns in Amazonian rain forests. Journal ofTropical Ecology, 13, 243–256.
Ruokolainen, K. & Tuomisto, H. (1998) Vegetación naturalde la zona de Iquitos. Geoecología y Desarrollo Amazónico:Estudio Integrado en la Zona de Iquitos, Perú (eds R. Kalliolaand S. Flores Paitán), pp. 253–365. Annales UniversitatisTurkuensis Ser A II 114. University of Turku, Finland.
Ruokolainen, K., Tuomisto, H., Vormisto, J. & Pitman, N.(2002) Two biases in estimating range sizes of Amazonianplant species. Journal of Tropical Ecology, 18, 935–942.
Ruokolainen, K. & Vormisto, J. (2000) The most widespreadAmazonian palms tend to be tall and habitat generalists.Basic and Applied Ecology, 1, 97–108.
Saatchi, S.S., Nelson, B., Podest, E. & Holt, J. (2000) Mappingland cover types in the Amazon basin using 1 km JERS-1mosaic. International Journal of Remote Sensing, 21, 1201–1234.
Sabatier, D., Grimaldi, M., Prévost, M.-F., Guillaume, J.,Godron, M., Dosso, M. & Curmi, P. (1997) The influence ofsoil cover organization on the floristic and structuralheterogeneity of a Guianan rain forest. Plant Ecology, 131,81–108.
Svenning, J.-C. (1999) Microhabitat specialization in a species-rich palm community in Amazonian Ecuador. Journal ofEcology, 87, 55–65.
Terborgh, J., Foster, R.B. & Nuñez, V., P. (1996) Tropical treecommunities: a test of the nonequilibrium hypothesis.Ecology, 77, 561–567.
Terborgh, J., Pitman, N., Silman, M., Schichter, H. &Nuñez, V., P. (2002) Maintenance of tree diversity intropical forests. Seed Dispersal and Frugivory: Ecology,Evolution and Conservation (eds D.J., Levey, W.R. Silva andM. Galetti), pp. 1–17. CABI Publishing, Wallingford, UK.
Tuomisto, H. (1998) What satellite imagery and large-scalefield studies can tell about biodiversity patterns in Amazo-nian forests. Annals of the Missouri Botanical Garden, 85,48–62.
Tuomisto, H., Linna, A. & Kalliola, R. (1994) Use of digitallyprocessed satellite images in studies of tropical rain forestvegetation. International Journal of Remote Sensing, 15,1595–1610.
Tuomisto, H. & Poulsen, A.D. (1996) Influence of edaphicspecialization on pteridophyte distribution in neotropicalrain forests. Journal of Biogeography, 23, 283–293.
Tuomisto, H. & Poulsen, A.D. (2000) Pteridophyte diversityand species composition in four Amazonian rain forests.Journal of Vegetation Science, 11, 383–396.
Tuomisto, H., Poulsen, A.D. & Moran, R.C. (1998) Edaphicdistribution of some species of the fern genus Adiantum inWestern Amazonia. Biotropica, 30, 392–399.
Tuomisto, H., Poulsen, A.D., Ruokolainen, K., Moran, R.C.,Quintana, C., Celi, J. & Cañas, G. (2003a) Linking floristicpatterns with soil heterogeneity and satellite imagery inEcuadorian Amazonia. Ecological Applications, 13, 352–371.
Tuomisto, H. & Ruokolainen, K. (2001) Variación de losbosques naturales en las áreas piloto a lo largo detransectos y en imágenes de satélite. Evaluación de RecursosVegetales no Maderables en la Amazonía Noroccidental(eds J. Duivenvoorden, H. Balslev, J. Cavelier, C. Grandez,H. Tuomisto and R. Valencia), pp. 63–96. IBED, Universiteitvan Amsterdam, The Netherlands.
Tuomisto, H., Ruokolainen, K., Kalliola, R., Linna, A.,Danjoy, W. & Rodriguez, Z. (1995) Dissecting Amazonianbiodiversity. Science, 269, 63–66.
Tuomisto, H., Ruokolainen, K., Poulsen, A.D., Moran, R.C.,Quintana, C., Cañas, G. & Celi, J. (2002) Distribution anddiversity of pteridophytes and Melastomataceae alongedaphic gradients in Yasuni National Park, EcuadorianAmazonia. Biotropica, 34, 516–533.
Tuomisto, H., Ruokolainen, K. & Yli-Halla, M. (2003b)Dispersal, environment, and the floristic variation ofwestern Amazonian forests. Science, 299, 241–244.
Vormisto, J., Phillips, O.L., Ruokolainen, K., Tuomisto, H. &Vásquez, R. (2000) A comparison of fine-scale distributionpatterns of four plant groups in an Amazonian rainforest.Ecography, 23, 349–359.
Received 11 December 2002 revision accepted 13 May 2003
