IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
Seasonal Variation in Spectral Signatures of Five
Genera of Rainforest Trees
Monica Papeş, Raul Tupayachi, Paola Martínez, A. Townsend Peterson, Gregory P. Asner, and
George V. N. Powell
Abstract—Recent technological and methodological advances
in the field of imaging spectroscopy (or hyperspectral imaging)
make possible new approaches to studying regional ecosystem
processes and structure. We use Earth Observing-1 Hyperion
satellite hyperspectral imagery to test our ability to identify tree
species in a lowland Peruvian Amazon forest, and to investigate
seasonal variation in species detections related to phenology.
We obtained four images from 2006–2008, and used them to
spectrally differentiate crowns of 42 individual trees of 5 taxa
using linear discriminant analysis. Temporal variation of tree
spectra was assessed using three methods, based on 1) position
of spectra in a two-dimensional canonical variable space, 2) a
broadband, multispectral dataset derived from sets of narrow
bands identified as informative to spectrally separate taxa, and 3)
narrow band vegetation indices (photochemical reflectance index
and anthocyanin reflectance index) sensitive to leaf pigments. We
obtained high classification success with a reduced set of trees (28
individuals) whose crowns were well represented on Hyperion 30
m resolution pixels. Temporal variability of spectra was confirmed
by each of the three methods employed. Understanding seasonality
of spectral characteristics of tropical tree crowns has implications
in spectral based multi-seasonal species mapping and studying
Index Terms—EO-1 Hyperion, hyperspectral, narrow bands, re-
mote sensing, seasonality, tropical tree.
complex issues such as climate change, carbon budgets, and
biodiversity conservation. Detailed field-based measurements
are time-consuming, limited in spatial and temporal coverage,
localized spatially, and logistically and financially difficult to
in high diversity tropical ecosystems.
These challenges are being addressed via remote sensing
(from air or space) of natural surfaces based on the spectral
NDERSTANDING the functional properties of ecosys-
tems at large spatial scales is necessary for addressing
and September 11, 2012; accepted September 27, 2012. Date of current version
May 13, 2013.
M. Papeş is with the Department of Zoology, Oklahoma State University,
Stillwater, OK 74078 USA (corresponding author, e-mail: firstname.lastname@example.org).
Ecology, Carnegie Institution for Science, Stanford, CA 94305 USA.
Center, The University of Kansas, Lawrence, KS 66045 USA.
G.V. N.Powellis with Conservation Science Program,WorldWildlife Fund,
Washington, DC 20037 USA.
Color versions of one or more of the figures in this paper are available online
Digital Object Identifier 10.1109/JSTARS.2012.2228468
characteristics of their absorption, reflectance, and scattering
properties of solar energy . Vegetation studies in the
400–2500 nm region of the electromagnetic spectrum of re-
flected light have shown that the visible range (VIS, 400–700
nm) is characterized by strong pigment absorption (chloro-
phyll, carotenoids, and anthocyanin); the near-infrared range
(NIR, 700–1300 nm) is dominated by scattering, the major
biochemical contributor being water, and to a smaller degree
lignin and cellulose; and the shortwave infrared range (SWIR,
1300–2500 nm) is characterized by low reflectance and high
absorption of light by water and vegetation components such
as starch, cellulose, lignin, and nitrogen –.
This fundamental knowledge of vegetation spectral proper-
ties forms the basis of imaging spectroscopy (or hyperspectral
imaging), which measures reflected light in hundreds of narrow,
contiguous spectral bands . These information-rich datasets
have been used to study tropical ecosystem function [forest
canopy chemistry, , ] and structure [tree species richness
and identification, , –]. Findings emerging from
previous remote sensing studies suggest that: 1) hyperspectral
imagery can be used to quantitatively assess biochemical prop-
erties of plants and to estimate ecosystem species composition;
2) translation of measurements of spectra from individual
leaves directly to the canopy scale is difficult, due to atmo-
spheric effects and tree crown structural elements (e.g., bark,
variable crown cover, epiphytes, etc.) influencing signals;
and 3) within-species variation exists. Although individual
variation of species’ spectra exists, within-species variation is
of a lower magnitude than between-species variation for many
tropical trees –. However, much less attention has
been given to temporal variation of canopy spectra that may
influence studies of ecosystem processes and biodiversity in
contrasting ways: such variation could be informative of tree
physiology or phenology, yet on the other hand, it could hinder
repeatability of tree identification and mapping between time
steps of observation.
Given the limited availability of airborne or spaceborne
hyperspectral imagery, temporal variation of species has
been studied mainly at the leaf level using on-site reflectance
measurements , . At the canopy level, narrow band
vegetation indices calculated from time-series imaging spec-
troscopy have been used to compare biochemical and physio-
logical changes of invasive and native tree species in Hawaii
. Other studies have focused on time-series analysis of
broad-band vegetation indices [e.g., Normalized Difference
Vegetation Index, NDVI; ] derived from easily accessible
multispectral, satellite imagery like the Moderate Resolution
Imaging Spectroradiometer [MODIS, ]. Given the coarse
1939-1404/$31.00 © 2012 IEEE
340IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
LIST OF TAXA WITH NUMBER OF INDIVIDUALS AND CLASSIFICATION ERRORS (FRACTION OF N) USING FOUR HYPERSPECTRAL DATASETS ( —JULY, DRY
SEASON; —NOVEMBER, BEGINNING OF WET SEASON; —DECEMBER, WET SEASON; AND
BASED ON CROWN SIZES (SEE METHODS)
—MAY, BEGINNING OF DRY SEASON) AND TWO TREE DATASETS,
spatial (250–1000 m) and spectral (10–100
these data (19 channels for 400–2500 nm range), however,
only a few vegetation indices can be derived, which are mostly
limited to describing broad changes in forest canopy coverage,
canopy water content, and photosynthetic activity –.
These studies are important in that they describe regional sea-
sonal phenological patterns, albeit irrespective of tree species.
In this study, we use Hyperion hyperspectral satellite im-
agery  to examine spectra of five common tropical tree
taxa, and to analyze their variation at four points in time from
2006–2008. More specifically, our aim was two-fold: (i) to
investigate tree spectral separation across time using linear
discriminant analysis and (ii) to quantify temporal variation
of spectra using the multidimensional narrow band space, a
simulated broadband space, and two narrow-band vegetation
indices. The location of the research site in hyperdiverse
southeastern Peru  makes for a challenging study, both in
terms of sample sizes of trees and availability of cloud-free
imagery. However, our aim is to improve understanding of
tree spectral seasonality at the crown level, and this analysis
represents a first step. We use only image-derived tree canopy
spectra to study temporal-spectral variation, avoiding the issues
of up-scaling leaf spectra to the canopy level; however, we
must take into account issues regarding quality of hyperspectral
satellite imagery, such as signal-to-noise ratio, striping, and
differences in sun angle, and atmospheric water absorption that
can mask vegetation absorption features.
) resolution of
II. MATERIALS AND METHODS
A. Study Site and Tree Samples
Large crowned, emergent trees belonging to five taxa
(Table I) identified as food resources for keystone animal
species such as macaws and peccaries (as part of a related,
ongoing study of animal foraging ecology) were mapped
in tropical lowland evergreen forests along the Los Amigos
River, a tributary of the Madre de Díos River, in southeastern
Peru (Fig. 1). The area is part of a conservation concession
managed by Asociación para la Conservación de la Cuenca
Amazónica (ACCA). Rainfall at this site is seasonal, being
lowest in June–September . Because of lack of precision in
species-level identifications, two tree taxa were examined at the
level of genus (Table I, Hymenaea sp. and Parkia sp.). Of the
five taxa selected for this study, four belong to Fabaceae, the
(Bertholletia excelsa) belongs to Lecythidaceae family, con-
taining 10 genera of mostly emergent trees . The Fabaceae
tree species in our study are grouped into Caesalpinioideae
which contains about 40 genera of trees (here represented by
Apuleia leiocarpa and Hymenaea sp.—a genus with 13 known
species) and Mimosoideae which contains about 20 genera of
large canopy, wet forest trees (here represented by Cedrelinga
cateniformis and Parkia sp.—a genus with about 40 known
species, including Old World species). Locations of trees were
recorded with GPS units, and checked on a panchromatic
PAPEŞ et al.: SEASONAL VARIATION IN SPECTRAL SIGNATURES OF FIVE GENERA OF RAINFOREST TREES 341
Fig. 1. The location of the study area in southeastern Peru (12.544S, 0.0611W) and the area where trees were sampled along a trail system (displayed with
a Quickbird image) are shown. The four Hyperion datasets analyzed are inserted as composite images (R:650 nm, G:847 nm, and B:548 nm bands). The four
July 2006; November 2007; December 2006; andMay 2008.
CHARACTERISTICS OF THE FOUR HYPERION IMAGES ACQUIRED 2006–2008
QuickBird satellite image with a spatial resolution of 61 cm,
acquired on 24 June 2006 (inset, Fig. 1). Because not all crowns
30 m in diameter, smaller crowns were represented by
mixed signal hyperspectral imagery pixels. We subsampled the
initial, “full” set of 42 trees to a “screened” subset of 28 trees
(Table I), all with crowns covering
pixel in the Hyperion imagery (see below). In this analysis, the
full dataset was used only in preliminary analyses of species’
spectral separation to check for effects of mixed pixel signals.
40% of the 3030 m
B. Hyperspectral Satellite Imagery
imagery is available, NASA’s Earth Observing 1 (EO-1) Hype-
rion, which provides data at 30 m spatial resolution. The EO-1
spacecraft was launched in 2000 and it is scheduled to con-
on-board fuel has been expended and the satellite is precessing
, . Due to cloud cover, out of a total of 30 attempts,
we obtained four images of the study area (Fig. 1), acquired on
20 July 2006 (dry season), 19 November 2007 (beginning of
wet season), 29 December 2006 (wet season), and 21 May 2008
(beginning of dry season). We will refer to these four different
“snapshots” in time as
lected at near nadir (view angle
Original radiance datasets (not georeferenced) were atmo-
spherically corrected to apparent surface reflectance using the
ACORN-6 atmospheric correction model (ImSpec, Palmdale,
CA). To reduce atmospheric correction modeling errors around
, and. All images were col-
; Table II).
342IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
Fig. 2. Summary of selection of Hyperion narrow bands, with bands removed from analyses in black and bands retained in gray (panel 1: original Hyperion
dataset). Panel 2 shows 126 narrow bands remaining after removal of uncalibrated bands. Panel 3 shows the channels used in separate linear discriminant analyses
for each season
. Selection of narrow bands when datasets are pooled is shown in panel 4. Panel 5 presents the combination of channels used to
simulate 9 multispectral bands.
water absorption channels (940 and 1140 nm), a cubic spline
curve was fit to these two spectral regions in each pixel. Inter-
mittent pixels with lower values create striping in some Hype-
rion bands . This effect was corrected using a destriping al-
on a band-by-band basis . The sensor collects data across
the electromagnetic spectrum in the 400–2500 nm range, in
242 separate bands.However,because 42 channels were uncali-
brated and others had low signal-to-noise ratios, we retained for
analysis only 126 narrow bands from each of the four temporal
datasets (Fig. 2, middle horizontal panel).
Finally, we georeferenced Hyperion images to the QuickBird
image with Geomatica 10.0 (PCI Geomatics Enterprises, Inc,
an “inverse rubber sheeting” procedure : Hyperion images
were fit to a vector file containing contours of oxbow lakes de-
rived from the high spatial resolution image (QuickBird). The
final locational root mean square error (RMSE) was
for all four images. The crown spectral measurements for all 42
trees were extracted from these datasets in ENVI 4.5 (Exelis Vi-
sual Information Solutions, Boulder, CO).
C. Statistical Analyses
1) Tree Crown Spectral Separation: The four datasets corre-
sponding to different times of the year and different years were
inant function analysis using PROC STEPWISE in SAS (SAS
Institute Inc., Cary, NC, USA) to identify the subset of the 126
bands that produced thebest discriminatory model,asmeasured
by Wilk’s lambda . This approach has been used previously
in remote sensing to classify tree species , , offering a
convenient means by which to reduce the high dimensionality
of hyperspectral imagery and correlations between channels.
This subset of bands was then used in a linear discriminant
analysis that seeks the highest ratio of between-groups sums
of squares to within-group sums of squares  to allow clas-
sification of individual trees as to species and genus. To as-
sess the degree to which the model predicts group memberships
PAPEŞ et al.: SEASONAL VARIATION IN SPECTRAL SIGNATURES OF FIVE GENERA OF RAINFOREST TREES343
Fig. 3. Individual trees (lighter symbols) and taxon means (darker, filled symbols) in canonical variable spaces were generated independently for the
datasets, demonstrating the seasonal change per taxa: (A) A. leiocarpa (x), B. excelsa
, and C. cateniformis ; and (B) Hymenaea sp. and Parkia sp.
better than random, the Wilk’s lambda, Pillai’s trace, and Roy’s
tracted from each of the four images
spectra of the screened datasets across the four data sets to test
whether discriminatory power is time-independent. We tested
classification results via cross-validations, omitting each tree
discriminant analysis on the missing crown. Canonical trans-
formation of the original variables (narrow spectral bands) al-
lowed for visualization and comparison of spectral separation
of species through time (see below).
2) Temporal Variation of Species’ Spectra: Before investi-
gating significance of tree spectral variability between time pe-
riods, we took a random sample of forest pixels across the entire
image to assess general seasonal variation in spectra. We tested
the ability of linear discriminant analysis to classify pixels as to
the correct time of image collection. This preliminary analysis
was intended to provide information regarding general, non-
species specific, vegetation spectral variation across
Temporal variation of spectral characteristics of the five taxa
studied was assessed using three different approaches: 1) based
on each taxon’s location in canonical spectral-dimensional
. We also pooled
space, 2) based on aggregating spectral subsets obtained via
stepwise discriminant function analysis and testing for re-
flectance differences by species and time, and 3) based on two
narrow band vegetation indices, the photochemical reflectance
index, PRI , and the anthocyanin reflectance index, ARI
. Variation in location of taxa in canonical space was
assessed through analysis of changes in measured spectral
distances among taxa means, and an analysis of geometric
differences in positions of the means.
For the first analysis, we calculated Euclidean distances
among mean spectral signatures for all taxa in a two-dimen-
sional canonical space (first and second derived canonical
variables, Fig. 3) as derived in the independent analyses de-
scribed above. Canonical axes were scaled to the same values
among the four independent image datasets. We thus generated
four distance matrices, one for each of the
datasets. In each matrix, all possible pairs (10) among the five
taxa were represented by Euclidian distance values calculated
using the taxon mean values of first and second derived canon-
ical variables. To test for changes in the spatial arrangement of
species’ centroids across the four seasonal datasets, we used a
Mantel test , a randomization technique that compares sets
of distance matrices for similar structure. The null hypothesis
tested in this pairwise approach is that no relationship exists
,, , and
344IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
between the matrices. In the original implementation of this
test , products of corresponding elements of the matrices
are summed and the observed Z statistic compared to values
obtained when the rows of one matrix are shuffled at random.
A more widely used implementation computes Pearson’s cor-
relation coefficients between corresponding elements of the
matrices . We calculated pairwise comparisons of matrices
random permutations in each test. These tests were carried
out in the statistical program R (R Development Core Team
2008), using “vegan”, a package of R functions for vegetation
We also tested the significance of relative positional change
of taxon means in two-dimensional canonical space among
the four seasonal datasets via a method adapted from the field
of geometric morphometrics . Taxonomists have devel-
oped techniques for studying individual shape variation and
delineation of morphological characters, by digitizing such
characters and their specific landmarks. These methods seek
statistically significant morphological differences between
individuals, and groups them based on these differences. We
used the positions of the means of the taxa in the four datasets
in canonical space (Fig. 3) as analogs of the landmarks to
track positional changes of taxa in the time domain across
. The variation in configuration of the landmarks in
two-dimensional canonical space among the four data sets
was summarized by calculating relative warps with respect
to an average shape of the location of five taxa means ob-
tained after fitting least squares of one configuration of taxa
means to another on complex regression functions . This
analysis is, in fact, a principal component analysis of the
covariance matrix of the partial warp scores, which are com-
puted when the shape of one configuration of landmarks is
projected on the average shape. We used two freeware pack-
ages available from the Stony Brook Morphometrics group
(http://life.bio.sunysb.edu/morph/index.html): tpsTDig to digi-
tize “landmarks” (in our study, taxon means in two-dimensional
canonical space), and tpsRelW to calculate relative warps of
datasets. In other words, we tested whether relative locations of
taxon means in canonical space change significantly between
datasets: significant change would imply that variation exists
in the geometric arrangement of species spectra in canonical
space across time, thus spectra vary from one season to another.
Second, a multivariate analysis of variance in spectral
space was carried out across the four datasets. The challenge
in this case was to reconcile the four band subsets that had
been selected independently for each seasonal analysis (see
above) to optimize taxon classification. We averaged selected
narrow bands into nine broad bands that represent the main
regions of the electromagnetic spectrum and are effectively
multispectral in nature (Fig. 2, right lower vertical panel):
three in the VIS range, one in the NIR, and five in the SWIR
region of the spectrum. The spectral range of the SWIR broad
bands was defined such that each of the four subsets previously
selected for seasonal analyses
least one narrow band. The NIR broad band spanned over 26
narrow bands, however each subset was represented by 5–11
narrow bands ( —eight narrow bands;
, and , and used 1000
was represented by at
—eleven). This multispectral dataset lost the specificity and
information richness of the original narrow band datasets, but
allowed for temporal analysis of spectral variation within taxa.
As such, the dataset generated for this analysis contained the
average reflectance in the nine multispectral bands for the 28
trees, grouped by time period. We used a repeated measures
Multivariate Analysis of Variance [MANOVA; ] to test
the null hypotheses 1) that variation in spectral reflectance
is not taxon-specific, and 2) that time period does not affect
changes in the reflectance. Significance was tested using PROC
GLM in SAS (SAS Institute Inc., Cary, NC, USA) which
calculates four multivariate tests: Wilks’ lambda, Pillai’s trace,
Hotelling-Lawley trace, and Roy’s greatest root.
Finally, to focus on a subset of possible phenological changes
in tree canopy, we calculated the average photochemical re-
flectance index (PRI) and anthocyanin reflectance index (ARI)
for each taxon in each of the four datasets
measures the reflectance in the green spectrum: the reference
tion region. The physiologically active signal is expressed in a
narrow spectral feature at 531 nm, due to the environmentally
or growth stage induced changes in the xanthophyll cycle, but
is located within the broad overlapping chlorophyll/carotenoid
PRI is used as an indicator of photosynthetic efficiency ,
. Values outside of the green vegetation range ( 0.2 to 0.2)
are considered indicative of factors that affect photosynthetic
activity . A review of PRI applications found that this ratio
is a reliable estimator of photosynthetic efficiency at leaf and
canopy levels . Another signal of early leaf development,
senescence, as well as response to stress is represented by the
anthocyanins, epidermal layer screening pigments which show
an absorption peak around 550 nm , . The chlorophyll
absorption at this wavelength is close to that at 700 nm, thus
the reflectance values at these two wavelengths are used in ARI
calculation to account for the confounding effect of chlorophyll
ARI values range between 0.001 and 0.1.
. The PRI
. Green vegetation
A. Tree Crown Spectral Separation
When the screened dataset was used, the stepwise linear
discriminant analysis selected 25–27 narrow bands with the
most discriminatory power (Fig. 2, panel 3, bands in dark gray)
for each temporal dataset; the full dataset (Fig. 2, panel 4)
produced very different results, with multivariate statistical
tests marginally significant and large errors in cross-validation
tests (Table I). In contrast, discriminant analyses based on the
screened dataset yielded statistically significant models (all
for Wilk’s lambda,
for Roy’s greatest root) and the cross-vali-
dation tests returned zero or low classification errors ( 25%,
Table I). The least successful classification observed was for
Hymenaea sp., the taxon for which sample sizes were lowest
(only 4 individuals in the screened dataset). Taxon classes were
well-defined in canonical space (Fig. 3).
for Pillai’s trace,
PAPEŞ et al.: SEASONAL VARIATION IN SPECTRAL SIGNATURES OF FIVE GENERA OF RAINFOREST TREES345
Fig. 4. Temporal spectral variation observed in canonical variable space: (A) representation of
through time, not by taxon (symbols as in Fig. 3); (B) a random sample of Hyperion pixels falling in forested areas (see Fig. 1), also grouped by time.
datasets and all (28) individual trees, showing separation
The screened dataset spectra extracted from the four tem-
poral images were also analyzed in a combined dataset. The
stepwise procedure identified 30 spectral bands (Fig. 2,
pooled), but these data produced large classification errors in
terms of taxa in the linear discriminant analysis (Table I, bottom
section); rather, this analysis identified group membership by
season time period (Fig. 4A). It appears that temporal spectral
variation is much greater than taxonomic variation. Within time
periods, however, spectral separation of the tree taxa was clear
and highly significant, as in our previous analyses .
B. Temporal Variation of Tree Crown Spectra
Linear discriminant analysis of the random sample of forest
pixels was highly successful in classifying pixels as to tem-
poral dataset (Fig. 4B). The largest error based on cross-vali-
dation tests was observed for
of the linear discriminant model (Wilk’s lambda, Pillai’s trace,
and Roy’s greatest root) were statistically significant (all
(0.0022); the multivariate tests
). This result can be interpreted as evidence for signifi-
cant overall temporal signal in the Hyperion measurements of
vegetation spectral properties.
We next tested for similarity in the geometric positions of
species’ spectral means in canonical space among seasons.
Estimates based on Euclidean distances among means with
Mantel test comparisons yielded no significant relationship
between any pair of distance matrices (
), which indicates that no significant (nonrandom)
similarity could be detected in positions of taxon means among
time periods. Simple Mantel tests have been shown to be
robust, even with small or skewed samples, and to avoid type
I error (rejecting the null hypothesis when it is actually true)
consistently through permutation tests .
The relative warps analysis adapted from geometric morpho-
metrics provided further support for the hypothesis of temporal
re-assortment of variation of spectral characteristics of these
tree taxa. We found no overlap between positions of taxon
means in two-dimensional canonical space over
346 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
Fig. 5. Position of taxa through time
RW2 represent the first two axes of the warp space. The separation of
in relative warp space generated with respect to an average canonical configuration of five taxa means; RW1 and
indicates that the canonical configuration of taxa (Fig. 3) changes over time.
the plots generated with the first two axes of the warp space.
Finally, the MANOVA of the multispectral dataset (nine
broad bands) confirmed the significance of interactions be-
tween time and variation in crown spectra
did not find a significant effect of taxon on observed spectral
. This result is not surprising, given that
crown spectra were averaged over numerous (73) narrow bands
(Fig. 2, right lower vertical panel), with little overlap with the
narrow bands identified as informative for taxon classification
when data were analyzed separately. The reduced overlap is
the result of applying a lax selection criterion that retained any
band present in one of the four linear discriminant analysis
datasets. Our expectation for the potential of repeated measures
MANOVA to find significant interaction between taxa and
spectral variation was low due to lack of specificity of the broad
band dataset. Nevertheless, even this rather coarse spectral
dataset again emphasized temporal variation in tree crown
C. Narrow Band Ratios
PRI average values by taxon and across the four datasets
0.11 to 0.05 (Fig. 6). All five taxa exhibited a
peak in average PRI in
which corresponds to the wet season,
suggesting an increase in photosynthetic function. In contrast,
the peak in average ARI values ( 0.04 to 0.04) occurred in
is an important issue because it is conceivable that present and
distributions and ecosystem structure and function . It is
thus important to understand the details of these measurements
that may influence ways and the scopes for which these tech-
nologies are employed. One of these details is the focus of the
We investigated spectral variability of tree crowns using
spaceborne hyperspectral data acquired at four different times
over a two-year period. Classification of individual trees to
taxon showed low or no error within each of the temporal
datasets, indicating that spectral separation is possible, as had
been the conclusion of our previous, simpler analyses .
The selection of narrow bands via linear discriminant analysis
, especially in the VIS region: in December
(wet season) most of the bands were selected from the green
portion of the spectrum. The reduction of the available samples
for analysis to 28 trees inevitably affects the statistical power
of our findings, limiting the extent to which they can be gener-
alized, but it also underlines the difficulty of generating large
datasets in very diverse tropical forests. Pooling data across
time in a single time- and taxon analysis masked the taxon
signal, probably owing to greater seasonal effects and possible
season by taxon interactions implied by results shown in Fig. 5.
In fact, simple visualizations of tree spectra (Fig. 7) call
attention to within-species variation among the four temporal
datasets. Some of the decreases in reflectance that are observed
are related to water absorption, particularly around 970, 1140,
and 1900 nm. As such, it can be difficult to separate the relative
contributions of atmospheric water vapor versus vegetation
characteristics in producing these drops in reflectance. These
absorption features are most distinct in the wet (December)
and dry (July) seasons. Nevertheless, since the narrow bands
used in the statistical tests did not overlap with these particular
regions of the spectrum, we consider that the statistically sig-
nificant variation observed among measurements of individual
PAPEŞ et al.: SEASONAL VARIATION IN SPECTRAL SIGNATURES OF FIVE GENERA OF RAINFOREST TREES347
Fig. 6. Temporal change in the average photochemical reflectance index (PRI, upper panel) and anthocyanin reflectance index (ARI, lower panel) values by
species. Peak values occur in the wet season for the PRI but in the dry season for the ARI.
tree crown spectra results from physiological and/or structural
changes that occur at different times of the year. However,
the differences in the solar illumination of the four Hyperion
images (Table II) may confound the canopy seasonal variation
signals . Additional, observational evidence in support
of the spectral variation hypothesis comes from phenological
monitoring in the field: most of the trees analyzed here showed
differences in the crown for growth stages (flowering, fruiting,
new leaves) between dry and wet seasons. The flowering season
occurred during the wet months (November and December)
40% of crowns had flower buds in A. leiocarpa and
Hymenaea sp. trees, as well as most (four out of six) of B.
excelsa individuals, whereas only one C. cateniformis and no
Parkia sp. trees had flower buds. The narrow band vegetation
indices suggested that the four taxa had higher photosynthetic
efficiency (but lower assimilation) under cloudy skies 
in December (PRI peak, Fig. 6) and increased anthocyanin
absorption in July (ARI peak, Fig. 6) that could be related to
waterstressin the dry seasonorwith phenological changes(leaf
development and senescence). The tree crown observations for
July do not corroborate the “new leaf” premise (leaf senescence
was not assessed). However, over half of the individuals of two
taxa (A. leiocarpa and B. excelsa) had mature fruit in July, and
we expect that this change in crown phenology could influence
Our crown-level level spaceborne imaging spectroscopy
approach to study spectral variation differs from previous
studies addressing, to a certain degree, similar questions. In
general, previous work focused on quantifying intraspecific
variation in leaf spectra, usually involving measurements of
leaf spectra with field or laboratory spectrometers , ,
, . Those studies provided information necessary to
evaluate the possible utility of spectra for species identification
and mapping. Another field of study that has seen considerable
attention is that of temporal variation in vegetation indices.
Hyperspectral datasets can be used to calculate narrow band
vegetation indices that are sufficiently specific to identify
changes in pigment concentration, photosynthetic activity, leaf
area, or crown cover , .
The analysis of multiple time frames of hyperspectral im-
agery provided support for the use of tree spectra to separate
taxa and showed that spectra are variable across seasons. Re-
lating spectral variation to individual phenological stages was
not possible, however successful tracking of seasonal spectral
348 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013
Fig. 7. Spectra of tree crowns of the 5 tree taxa studied (screened dataset shown only) at the 4 times examined
removed; colors represent different individuals of the same taxon. Apparent reflectance values were scaled from floating point into integer values using a 10,000
, with uncalibrated Hyperion bands
changes at broader, ecosystem level for assessing forest canopy
The main scope of this study was to investigate spectral vari-
ation in a set of common tropical tree taxa identified as im-
portant food sources for keystone animal species. Our limited
sample sizes precluded us from analyzing possible correlations
between specific tree phenological stages (leafing, fruiting, etc.)
and crown spectral variability, a link we feel to be worth ex-
ploring in more depth in future studies. Such information is
necessary for building a future, regional-scale understanding of
habitat requirements and phenology influences on seasonal dis-
tribution of animal species that will ultimately aid biodiversity
The authors are grateful to World Wildlife Fund program in
Peru for financial and logistic support. The authors thank J. Ja-
cobson for help with the atmospheric correction of EO-1 Hype-
rion images, and F. Ballantyne,K.Jensen,and S.Martinuzzi for
their comments on an earlier version of this manuscript. The au-
thors also greatly appreciate the suggestions provided by three
anonymous reviewers and the associate editor to improve the
clarity and focus of the manuscript.
 S. L. Ustin, D. A. Roberts, J. A. Gamon, G. P. Asner, and R. O. Green,
“Using imaging spectroscopy to study ecosystem processes and prop-
erties,” BioScience, vol. 54, pp. 523–534, 2004.
 P. J. Curran, “Remote sensing of foliar chemistry,” Remote Sens. Env-
iron., vol. 30, pp. 271–278, 1989.
 G. P. Asner, “Biophysical and biochemical sources of variability in
canopy reflectance,” Remote Sens. Environ., vol. 64, pp. 234–253, Jun.
 G. A. Blackburn, “Remote sensing of forest pigments using airborne
imaging spectrometer and LIDAR imagery,” Remote Sens. Environ.,
vol. 82, pp. 311–321, 2002.
trometry for Earth remote sensing,” Science, vol. 228, pp. 1147–1153,
 G. P. Asner and P. M. Vitousek, “Remote analysis of biological inva-
USA, vol. 102, pp. 4383–4386, Mar. 22, 2005.
 M. E. Martin and J. D. Aber, “High spectral resolution remote sensing
Applications, vol. 7, pp. 431–443, 1997.
 G. P. Asner, R. F. Hughes, P. M. Vitousek, D. E. Knapp, T. Kennedy-
Bowdoin, and J. Boardman et al., “Invasive plants transform the three-
dimensional structure of rain forests,” Proc. National Academy of Sci-
ences USA, vol. 105, pp. 4519–4523, 2008.
 A. Held, C. Ticehurst, L. Lymburner, and N. Williams, “High resolu-
 K. M. Carlson, G. P. Asner, R. F. Hughes, R. Ostertag, and R. E.
Martin, “Hyperspectral remote sensing of canopy biodiversity in
Hawaiian lowland rainforests,” Ecosystems, vol. 10, pp. 536–549, Jun.
 M. L. Clark, D. A. Roberts, and D. B. Clark, “Hyperspectral discrim-
ination of tropical rain forest tree species at leaf to crown scales,” Re-
mote Sens. Environ., vol. 96, pp. 375–398, 2005.
 J. Zhang, B. Rivard, A. Sánchez-Azofeifa, and C. Castro-Esau, “Intra-
and inter-class spectral variability of tropical tree species at La Selva,
Costa Rica: Implications for species identification using HYDICE im-
agery,” Remote Sens. Environ., vol. 105, pp. 129–141, 2006.
PAPEŞ et al.: SEASONAL VARIATION IN SPECTRAL SIGNATURES OF FIVE GENERA OF RAINFOREST TREES349
 M. Kalacska, G. A. Sánchez-Azofeifa, B. Rivard, T. Caelli, H.
P. White, and J. C. Calvo-Alvarado, “Ecological fingerprinting of
ecosystem succession: Estimating secondary tropical dry forest
structure and diversity using imaging spectroscopy,” Remote Sens.
Environ., vol. 108, pp. 82–96, 2007.
 R. Pouteau and B. Stoll, “SVM Selective Fusion (SELF) for multi-
source classification of structurally complex tropical rainforest,” IEEE
J. Selected Topics in Applied Earth Observations and Remote Sensing,
vol. 5, pp. 1203–1212, 2012.
 M. Papeş, R. Tupayachi, P. Martínez, A. Peterson, and G. Powell,
“Using hyperspectral satellite imagery for regional inventories: A test
with tropical emergent trees in the Amazon Basin,” J. Vegetation Sci-
ence, vol. 21, pp. 342–354, 2010.
 G. P. Asner and R. E. Martin, “Airborne spectranomics: Mapping
canopy chemical and taxonomic diversity in tropical forests,” Fron-
tiers in Ecology and the Environment, vol. 7, pp. 269–276, Jun. 2009.
 J. B. Feret and G. P. Asner, “Spectroscopic classification of tropical
forest species using radiative transfer modeling,” Remote Sens. Env-
iron., vol. 115, pp. 2415–2422, Sep. 15, 2011.
 L. Naidoo, M. A. Cho, R. Mathieu, and G. Asner, “Classification of
savanna tree species, in the Greater Kruger National Park region, by
integrating hyperspectral and LiDAR data in a Random Forest data
mining environment,” ISPRS J. Photogramm. Remote Sens., vol. 69,
pp. 167–179, 2012.
 G. P. Asner, R. E. Martin, R. Tupayachi, R. Emerson, P. Martínez, and
F. Sinca et al., “Taxonomy and remote sensing of leaf mass per area
(LMA) in humid tropical forests,” Ecological Applications, vol. 21,
pp. 85–98, Jan. 2011.
M. Quesada, “Variability in leaf optical properties of Mesoamerican
trees and the potential for species classification,” Am. J. Botany, vol.
93, pp. 517–530, Apr. 2006.
 M. A. Cochrane, “Using vegetation reflectance variability for species
level classification of hyperspectral data,” Int. J. Remote Sens., vol. 21,
pp. 2075–2087, 2000.
 G. P. Asner, R. E. Martin, K. M. Carlson, U. Rascher, and P. M.
Vitousek, “Vegetation-climate interactions among native and invasive
species in Hawaiian rainforest,” Ecosystems, vol. 9, pp. 1106–1117,
 A. Huete, C. Justice, and H. Liu, “Development of vegetation and
soil indexes for MODIS-EOS,” Remote Sens. Environ., vol. 49, pp.
 C. O. Justice, E. Vermote, J. R. G. Townshend, R. Defries, D. P. Roy,
and D. K. Hall et al., “The Moderate Resolution Imaging Spectrora-
diometer (MODIS): Land remote sensing for global change research,”
IEEE Trans. Geosci. Remote Sens., vol. 36, pp. 1228–1249, 1998.
 X. M. Xiao, S. Hagen, Q. Y. Zhang, M. Keller, and B. Moore, “De-
tecting leaf phenology of seasonally moist tropical forests in South
vol. 103, pp. 465–473, Aug. 30, 2006.
 A. Huete, Y. Kim, P. Ratana, K. Didan, Y. E. Shimabukoro, and T.
Miura, “Assessment of phenologic variability in Amazon tropical rain-
perspectral Remote Sensing of Tropical and Subtropical Forests, M.
Kalacska and G. A. Sánchez-Azofeifa, Eds.
Press, 2008, pp. 233–259.
 R. Houborg, H. Soegaard, and E. Boegh, “Combining vegetation index
and model inversion methods for the extraction of key vegetation bio-
physical parameters using Terra and Aqua MODIS reflectance data,”
Remote Sens. Environ., vol. 106, pp. 39–58, Jan. 15, 2007.
and F. Gao et al., “Monitoring vegetation phenology using MODIS,”
Remote Sens. Environ., vol. 84, pp. 471–475, Mar. 2003.
 S. G. Ungar, J. S. Pearlman, J. A. Mendenhall, and D. Reuter,
“Overview of the Earth Observing One (EO-1) mission,” IEEE Trans.
Geosci. Remote Sens., vol. 41, pp. 1149–1159, 2003.
 R. B. Foster, “The floristic composition of the Rio Manu floodplain
forest,” in Four Neotropical Rainforests, A. H. Gentry, Ed.
Haven, CT: Yale Univ. Press, 1993, pp. 99–379.
 L. J. Osher and S. W. Buol, “Relationship of soil properties to parent
material and landscape position in eastern Madre de Dios, Peru,” Geo-
derma, vol. 83, pp. 143–166, 1998.
Boca Raton, FL: CRC
 A. H. Gentry, A Field Guide to the Families and Genera of Woody
Plants of Northwest South America (Colombia, Ecuador, Peru) With
Supplementary Notes on Herbaceous Taxa.
Chicago Press, 1996.
 S. G. Ungar, E. M. Middleton, L. Ong, and P. E. Campbell, “EO-1 Hy-
perion onboard performance over eight years: Hyperion calibration,”
presented at the 6th EARSeL SIG IS, Tel-Aviv, Israel, 2009.
 E. M. Middleton, S. G. Ungar, D. Mandl, L. Ong, S. Frye, and P. E.
Campbell et al., “The Earth Observing One (EO-1) satellite mission:
Over a decade in space,” IEEE J. Selected Topics in Applied Earth
Observations and Remote Sensing, 2013, EO-1 Special Issue.
 D. G. Goodenough, A. Dyk, K. O. Niemann, J. S. Pearlman, C. Hao,
and T. Han et al., “Processing Hyperion and ALI for forest classifi-
cation,” IEEE Trans. Geosci. Remote Sens., vol. 41, pp. 1321–1331,
 G. P. Asner and K. B. Heidebrecht, “Imaging spectroscopy for deserti-
fication studies: Comparing AVIRIS and EO-1 Hyperion in Argentina
drylands,” IEEE Trans. Geosci. Remote Sens., vol. 41, pp. 1283–1296,
 A. Dyk, D. G. Goodenough, A. S. Bhogal, J. Pearlman, and J. Love,
“Geometric correction and validation of Hyperion and ALI data for
EVEOSD,” in Proc. IEEE Int. Geoscience and Remote Sensing Symp.,
Toronto, ON, Canada, 2002, pp. 579–583.
 C. Huberty, Applied Discriminant Analysis.
“Species classification of tropical tree leaf reflectance and dependence
on selection of spectral bands,” in Hyperspectral Remote Sensing of
Tropical and Subtropical Forests, M. Kalacska and G. A. Sánchez-
Azofeifa, Eds. Boca Raton, FL: CRC Press, 2008, pp. 141–159.
Techniques. Thousand Oaks, CA: Sage Publications, 2007.
 J. A. Gamon, L. Serrano, and J. S. Surfus, “The photochemical re-
flectance index: An optical indicator of photosynthetic radiation use
efficiency across species, functional types, and nutrient levels,” Oe-
cologia, vol. 112, pp. 492–501, 1997.
 A. A. Gitelson, M. N. Merzlyak, and O. B. Chivkunova, “Opticalprop-
erties and nondestructive estimation of anthocyanin content in plant
leaves,” Photochem. Photobiol., vol. 74, pp. 38–45, 2001/07/01 2001.
 N. Mantel, “Detection of disease clustering and a generalized regres-
sion approach,” Cancer Research, vol. 27, pp. 209–220, 1967.
 P. Legendre, “Comparison of permutation methods for the partial cor-
relation and partial Mantel tests,” J. Statistical Computation and Sim-
ulation, vol. 67, pp. 37–73, 2000.
 J. Oksanen, R. Kindt, P. Legendre, B. O’Hara, G. L. Simpson, and M.
Henry et al., Vegan: Community Ecology Package. R Package Version
 D. C. Adams, F. J. Rohlf, and D. E. Slice, “Geometric morphometrics:
Ten years of progress following the ‘revolution’,” Italian J. Zoology,
vol. 71, pp. 5–16, 2004.
 F. J. Rohlf, “Relative warp analysis and an example of its application
to mosquito wings,” in Contributions to Morphometrics, L. F. Marcus,
E. Bello, and A. García-Valdecasas, Eds.
Nacional de Ciencias Naturales, Consejo Superior de Investigaciones
 J. W. L. Cole and J. E. Grizzle, “Applications of multivariate analysis
of variance to repeated measurements experiments,” Biometrics, vol.
22, pp. 810–828, 1966.
N. C. Coops et al., “Linking foliage spectral responses to canopy-level
ecosystem photosynthetic light-use efficiency at a Douglas-fir forest in
 D. A. Sims and J. A. Gamon, “Relationships between leaf pigment
content and spectral reflectance across a wide range of species, leaf
structures and developmental stages,” Remote Sens. Environ., vol. 81,
pp. 337–354, 2002.
 M. F. Garbulsky, J. Peñuelas, J. Gamon, Y. Inoue, and I. Filella, “The
photochemical reflectance index (PRI) and the remote sensing of leaf,
canopy and ecosystem radiation use efficiencies: A review and meta-
analysis,” Remote Sens. Environ., vol. 115, pp. 281–297, 2011.
stress responses,” Photochem. Photobiol., vol. 70, pp. 1–9, 1999.
Chicago, IL: Univ.
New York: Wiley-Inter-
Madrid, Spain: Museo
350IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013 Download full-text
 J. Q. Chambers, G. P. Asner, D. G. Morton, L. O. Anderson, S. S.
Saatchi, and F. D. B. Espírito-Santo et al., “Region a ecosystem struc-
ture and function: Ecological insights from remote sensing of tropical
forests,” Trends in Ecology & Evolution, vol. 22, pp. 414–423, 2007.
 L. S. Galvão, J. R. dos Santos, D. A. Roberts, F. M. Breunig, M.
Toomey, and Y. M. de Moura, “On intra-annual EVI variability in
the dry season of tropical forest: A case study with MODIS and
hyperspectral data,” Remote Sens. Environ., vol. 115, pp. 2350–2359,
 C. Castro-Esau and M. Kalacska, “Tropical dry forest phenology and
discriminationoftropicaltree species using hyperspectraldata,” inHy-
perspectral Remote Sensing of Tropical and Subtropical Forests, M.
Kalacska and G. A. Sánchez-Azofeifa, Eds.
 M. Kalacska, G. A. Sánchez-Azofeifa, J. Calvo-Alvarado, B. Rivard,
Index and spectral vegetation indices in three Mesoamerican tropical
dry forests,” Biotropica, vol. 37, pp. 486–496, 2005.
 P. B. Reich, “Key canopy traits drive forest productivity,” Proc. Royal
Soc. B, vol. 279, pp. 2128–2134, Jan. 25, 2012, 2012.
 C.E.Doughty, G.P.Asner, and R.E.Martin,“Predictingtropicalplant
physiology from leaf and canopy spectroscopy,” Oecologia, vol. 165,
pp. 289–299, Feb. 2011.
Boca Raton, FL: CRC
Monica Papeş received the B.S. degree in biology from Universitatea de Vest,
Biology from University of Kansas, in 2003 and 2009, respectively.
She is currently Assistant Professor in the Department of Zoology at Okla-
homa State University. Her research focuses on species’ geographic distribu-
tions, and in particular on remote sensing applications to mapping biodiversity
and seasonal variation of species’ distributions.
Raul Tupayachi received the B.S. degree in biology with a minor in plant
ecology from Universidad San Antonio Abad del Cusco, Peru.
Currently, he is a Staff Member in the Department of Global Ecology,
Carnegie Institution for Science, Stanford, CA. His research focuses on plant
taxonomy, diversity, and distribution.
Paola Martínez received the B.S. degree in biology from Universidad San
Agustin of Arequipa, Peru.
Currently, she is a Staff Member in the Department of Global Ecology,
Carnegie Institution for Science, Stanford, CA. Her research interests are
taxonomy and diversity of trees of the Amazon Basin.
A. Townsend Peterson received the B.S. degree in zoology from Miami Uni-
versity in 1985, and the M.S. and Ph.D. in evolutionary biology from the Uni-
versity of Chicago in 1987 and 1990, respectively.
He is currently University Distinguished Professor in the Biodiversity Insti-
tute of the University of Kansas, and works on diverse topics related to distri-
butional ecology and evolutionary biology of organisms.
Gregory P. Asner received the B.S. degree in civil and environmental engi-
neering, the M.A. degree in biogeography, and the Ph.D. degree in environ-
mental, organismic, and population biology from the University of Colorado,
Boulder, in 1991, 1995, and 1997, respectively.
Currently, he is a Faculty Member in the Department of Global Ecology,
Carnegie Institution for Science, Stanford, CA. He also holds a faculty posi-
tion in the Department of Environmental Earth System Science, Stanford Uni-
versity, Stanford. His scientific research centers on how human activities alter
the composition and functioning of ecosystems at regional scales. He combines
field work, airborne and satellite mapping, and computer simulation modeling
to understand the response of ecosystems to land use and climate change.
George V. N. Powell received the Ph.D. degree from the University of Cali-
fornia, Davis in 1977.
He is a SeniorConservation Scientistwiththe ConservationScience Program
as a function of seasonal migrations, conservation landscape design and priority
setting, and conservation applications of GIS.