| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
What's this? |
Ecology |
2Earth Observation Systems Laboratory, Department of Earth and Atmospheric Sciences, University of Alberta, Edmonton, Alberta, Canada, T6G 2E3; 3Smithsonian Tropical Research Institute, Apartado 2072, Balboa, Ancon, Republic of Panama; 4Centro de Investigaciones en Ecosistemas, Universidad Nacional Autonoma de Mexico, Morelia, Michoacan, Mexico
Received for publication August 10, 2005. Accepted for publication January 31, 2006.
ABSTRACT
Leaf traits and physiological performance govern the amount of light reflected from leaves at visible and infrared wavebands. Information on leaf optical properties of tropical trees is scarce. Here, we examine leaf reflectance of Mesoamerican trees for three applications: (1) to compare the magnitude of within- and between-species variability in leaf reflectance, (2) to determine the potential for species identification based on leaf reflectance, and (3) to test the strength of relationships between leaf traits (chlorophyll content, mesophyll attributes, thickness) and leaf spectral reflectance. Within species, shape and amplitude differences between spectra were compared within single leaves, between leaves of a single tree, and between trees. We also investigated the variation in a species' leaf reflectance across sites and seasons. Using forward feature selection and pattern recognition tools, species classification within a single site and season was successful, while classification between sites or seasons was not. The implications of variability in leaf spectral reflectance were considered in light of potential tree crown classifications from remote airborne or satellite-borne sensors. Species classification is an emerging field with broad applications to tropical biologists and ecologists, including tree demographic studies and habitat diversity assessments.
Key Words: classification • hyperspectral • interspecific • intraspecific • leaf level • Mesoamerica • tropical trees
An extensive body of literature on leaf optics describes the interaction of light with light-harvesting pigments and leaf structural components (e.g., Gates et al., 1965
; Woolley, 1971
; Gausman, 1985
; Vogelmann, 1993
). Three primary regions emerge from the reflectance spectrum of a leaf, the first (400–700 nm) is predominantly due to light absorption by pigments (especially chlorophyll), the second (700–1200 nm) is influenced by light scattering at air–cell-wall interfaces and characterized by high reflectance and transmittance, and the third (1200–2500 nm) is governed by water absorption (Hoffer and Johannsen, 1969
). Furthermore, scientists have examined perturbations to leaf absorption, reflection, and transmission of light by external influences such as light availability (Gausman, 1984
), disease (Knipling, 1967
), freezing (Woolley, 1971
), stress (Carter, 1993
), and senescence (Gitelson and Merzlyak, 1994
). The spectral properties of different plant types (Gates et al., 1965
) or species (Allen et al., 1970
; Gausman and Allen, 1973
; Knapp and Carter, 1998
) have also been compared. In more recent years, our understanding of leaf optics has been extended to larger-scale applications in the field of remote sensing, for assessing physiological performance of plant canopies (e.g., Gamon et al., 1990
) and distinguishing between land cover vegetation types (e.g., Salem et al., 1995
). In effect, results from leaf-level studies are often used as a basis for establishing techniques for retrieving information from canopy reflectance measured by remote sensors.
Little is known about the leaf optical properties of tropical trees or about the potential for identifying tropical tree species using remote sensing. Among the few exceptions are studies by Lee and Graham (1986)
and Lee et al. (1990)
, who compared leaf optical properties of tropical sun and shade species, as well as Avalos et al. (1999)
, who compared leaf optical properties of tropical dry forest trees and lianas. Fung and Siu (1998)
and Cochrane (2000)
assessed reflectance of leaf or branch samples for differentiating species of tropical trees. More recently, Clark et al. (2005)
was successful in discriminating seven tropical tree species at the leaf, pixel, and crown levels, publishing the first report on automated tropical rain forest tree classification using high spatial resolution hyperspectral imagery, in which spectral reflectance is measured over numerous contiguous bands of narrow width (typically 
10 nm). The study represents an important breakthrough for identifying tropical tree species. Operational species classification could have numerous applications, including monitoring endangered or commercial tree species, characterizing biodiversity in ecologically important habitats such as reserves and biological corridors, and monitoring changes in species composition over time and changes in tree demography associated with global environmental changes (Clark et al., 2005
).
The precept for species discrimination is that interspecies variability in spectral reflectance exceeds intraspecies variability. This presents a challenge for species identification due to inherent similarities in vegetation characteristics among species, including leaf biochemistry and anatomy, which result in similarities in leaf spectral reflectance. After posing the question "How unique are spectral signatures?," Price (1994)
determined that while some are clearly distinct, others may not be separable at all. In reality, despite the advantages of hyperspectral technologies, classification of plant species may be difficult and, at times, unfeasible. A complicating factor is the dynamic nature of plant spectral signatures. Natural cycles such as leaf flush and senescence, as well as environmental factors affecting mineral nutrition, health, light availability, and water supply, may contribute to changes in a species' spectral signature over time or at a given position in the canopy (Gausman, 1984
, 1985
; Carter, 1993
). To date, even the studies from temperate regions have not fully evaluated the seasonal dynamics of change in leaf spectral characteristics on tree classifications, nor do they explore the stability of tree spectral signatures over multiple sites with a range of climatic and edaphic conditions. Intuitively, these factors, which lead to within-species variability in reflectance spectra, could lead to significant confusion for species recognition using hyperspectral data. Although Clark et al. (2005)
tested within-species variability for seven tropical tree species at one site, the degree of within-species variability over a large number of tropical tree species and a range of environmental conditions has not been adequately addressed.
With image-based, automated tropical tree species classification in its preliminary stages of development, we believe it to be constructive at this time to provide a comprehensive assessment of intraspecific variability of leaf spectral reflectance and its potential to influence tree species classifications. The main objective of this paper, therefore, was to evaluate within-species and between-species variability in leaf spectral reflectance for a wide variety of tropical trees found at seven sites in southern Mexico and Central America. We also investigated crown-level spectra of five tree species from Fort Sherman, Panama. In effect, we asked, "How unique are spectral signatures of tropical tree species?" A large part of our study focuses on the leaf level for several reasons: (1) by necessity, because high spatial resolution (<5 m) hyperspectral data was not available for most of our sites at the time of data collection, (2) the use of leaf spectra allows for a more controlled assessment of variability at multiple levels, and (3) spectral reflectance from leaves is usually the dominant contribution to tree crown reflectance. To evaluate variability in leaf reflectance of tropical trees, spectral metrics that describe reflectance shape and amplitude (Price, 1994
) are compared at various levels within and between species. To explore the potential to identify species, we also classified the spectral data from the different sites using forward selection of features and a set of classifiers described in Castro-Esau et al. (2004)
. Lastly, we examined the correspondence between selected leaf traits (chlorophyll content, leaf thickness, and leaf histological features) and leaf spectral reflectance characteristics for one site to further our understanding of differences in leaf spectra between species.
MATERIALS AND METHODS
Study sites
Leaf samples were gathered from seven sites in Mexico, Costa Rica, and Panama (Table 1). Five of the sites support tropical dry forest ecosystems, where our research group is concurrently conducting studies on leaf area index and forest complexity and their ties with remote detection of secondary growth vs. mature forest. These sites are part of an international research consortium called TROPI-DRY that aims to understand the biophysical characteristics of neotropical dry forests and human interactions with these forests. Tropical dry forests undergo significant phenological changes through dry–wet seasonal cycles and typically contain species with varying degrees of deciduousness (Frankie et al., 1974
; Bullock and Solis-Magallanes, 1990
; Avalos and Mulkey, 1999
). In addition, tropical dry forests can present high rates of species endemism (Lott et al., 1987
; Kalacska et al., 2004
).
|
Crown spectra were recorded from a canopy crane at Fort Sherman, Panama, a tropical wet forest site.
Leaf sample collection
At each site, 3–15 sun leaves were collected per tree (Table 2). One to five trees were sampled per species. In general, leaves were collected from the distal or middle portion of the branch. At some sites, it was not logistically feasible to sample more than one tree per species, such as in Parque Natural Metropolitano and Fort Sherman, where we were restricted to trees within reach of a canopy crane and by institutional regulations regarding the size of samples that can be collected from each tree. A full list of the species sampled is provided in Table 3. The samples were acquired using one of the following methods: pole pruner (Mexico and Costa Rica), sling shot (Costa Rica), or canopy crane (Panama). Leaves were collected, placed in plastic bags containing moistened paper towels, labeled, and placed in a larger bag or cooler containing ice. Samples were transported to an indoor laboratory, and leaf spectral reflectance was analyzed the same day. Maintaining leaf moisture prevented significant changes in leaf reflectance during this period (Foley et al., in press
).
|
|
|
Leaf chlorophyll content and thickness
Five 2.0-cm2 leaf cores per tree species, each from a different leaf, were analyzed for chlorophyll content at Parque Natural Metropolitano. Chlorophyll a, chlorophyll b, and total chlorophyll content were determined using the dimethyl sulphoxide extraction method (Hiscox and Israelstam 1979
). A Smartspectro spectrophotometer (LaMotte, Chestertown, Maryland, USA) was used to determine absorbance at 645 and 663 nm. Pigment content was computed using Arnon's (1949)
relationships and converted to a per area basis. In addition, chlorophyll content was nondestructively assessed using a SPAD 502 (Minolta Camera, Osaka, Japan) chlorophyll absorbance meter. The SPAD measures absorbance at 650 and 940 nm and yields a relative chlorophyll content, or "chlorophyll index" value (Markwell et al., 1995
; Richardson et al., 2002
). Leaf thickness was measured for 5–10 leaves per species using a Nikon (Tokyo, Japan) digital micrometer (to nearest 0.001 mm). Five measurements were taken per leaf, avoiding large veins.
Leaf histology
Small pieces of leaf blade (approximately 1 cm long and 2–3 mm wide) were collected at Parque Natural Metropolitano in October 2003 for histological analysis. The leaf pieces were fixed under vacuum for 2 wk in a formalin aceto-alcohol (5% formalin) solution. After fixation, the samples were run through an ethanol processing center (Fisher Histomatic 166, Hampton, New Hampshire, USA) and embedded in paraffin molds. Thin sections (5 µm) were cut from the mounted samples using a microtome. Slides were mounted and stained (Harris' hematoxylin stain), and viewed with a Leica (Wetzlar, Germany) DM RXM light microscope (200x total magnification). Digital photos were taken of the stained samples. The percentage of air space in the spongy mesophyll was determined by classifying the photos for cells vs. air space. Spongy mesophyll thickness and the ratio of spongy mesophyll to total leaf thickness were calculated using a 100-µm scale in each photo, from which a new ruler was constructed with 10-µm increments.
Leaf diffuse reflectance and transmittance
In addition to measurements of bidirectional reflectance of leaves described next, diffuse reflectance and transmittance were recorded using an 1800–12S external integrating sphere (LI-COR, Lincoln, Nebraska, USA) in conjunction with a UniSpec spectrometer. The spectrometer fiber optic was inserted through a port on the sphere. To record diffuse reflectance, the sphere illuminator was directed toward the adaxial side of the leaf sample in the sample port, and the fiber optic was directed to the sphere wall. To record transmittance, the leaf was reversed so that the underside faced the sphere interior. The sphere illuminator was directed through the leaf sample, and the spectrometer fiber optic was directed to the wall. Dark scans were performed regularly, and white reference scans (using barium sulfate) were repeated for every leaf sample. Both diffuse reflectance and transmittance were calculated as a ratio in relation to the white reference data.
Measurements of bidirectional spectral reflectance
Bidirectional leaf spectral reflectance was measured at all sites, using one of two spectrometers: (1) a UniSpec Spectral Analysis System (PP Systems, Amesbury, Massachusetts, USA) or (2) a FieldSpec HandHeld spectrometer (Analytical Spectral Devices, Boulder, Colorado, USA). The 2002 and 2003 data sets employed the FieldSpec HandHeld (FieldSpec HH) and the UniSpec, respectively. All crown spectra (Fort Sherman, Panama) were measured using the UniSpec.
UniSpec spectral analysis system
The Unispec spectral analysis system VIS/NIR relies on a 256-element photodiode array. The spectral range is 350–1100 nm, with a sampling interval of 3.3 nm and a spectral resolution of <10 nm. A bifurcated fiber optic delivers light from an internal 7.0 W halogen lamp via one branch and receives reflected light via the other.
To measure spectral reflectance of leaves, we employed a leaf clip that holds the foreoptic at 60° and maintains a 2.3 mm diameter field of view (FOV). Sample reflectance was measured by comparing leaf reflectance to reflectance of a white standard (barium sulfate). Ten scans were averaged per recorded spectrum. Dark scans and white reference scans were performed frequently to detect instrument drift on the spectra.
At Fort Sherman (Panama), crown spectra were recorded mid-morning under sunny conditions. White reference measurements were taken before measurements for each species as well as intermittently if we sensed the light conditions changed. Care was also taken to avoid taking measurements if the shadow of the gondola fell within the estimated FOV. Spectra were recorded at a height of 5 m above each crown, determined by lowering a measuring tape from the gondola. Using the bare UniSpec fiber-optic probe (40° FOV), the diameter of the viewed crown was approximately 3.6 m. Six measurements were taken from the same position above each crown and averaged.
FieldSpec HandHeld
The FieldSpec HandHeld spectrometer relies on one 512-element photodiode array. Its range is 325–1075 nm, with a sampling interval of 1.6 nm and spectral resolution of 3.5 nm at 700 nm. The FOV of the bare fiber-optic probe is 25°.
To obtain leaf spectral reflectance measurements, an external 50 W halogen lamp was directed at the leaf at an illumination angle of 45°. Leaves were placed against a black minimal (2%) reflectance panel. The FieldSpec HandHeld fiber-optic probe, inserted into a mounting gun attached to a tripod, was positioned directly above the sample. In all cases, height was adjusted so that the FOV had a diameter of approximately 11 mm. Reflectance was determined by standardizing the sample data to white reference (Spectralon, Labsphere, North Sutton, New Hampshire, USA) data under the same illumination and viewing conditions. White reference scans were recorded frequently. Ten scans were averaged per reflectance spectrum.
Spectral analysis
Spectra collected with the two spectrometers were analyzed separately rather than collectively due to differences between spectra measured for the same objects as described in Castro-Esau et al. (in press)
.
D and 
Two spectral metrics described by Price (1994)
enabled comparisons between pairs of spectra as follows:
D represents the difference in amplitude between two spectra and is computed as the root mean square difference between a pair of spectra (S1 and S2) averaged over the spectral range of interest (
a to b). The metric determines the angle between two spectra and may be interpreted as the difference in shape between the pair of spectra. It is computed as a vector dot product for the pair of spectra, with amplitude dependence removed.
The spectral range used to calculate D and was 445.9–949.6 nm for both the UniSpec and FieldSpec Handheld. In this region, noise from both instruments was low. To compute D and for the same spectral range and interval, FieldSpec HandHeld data were interpolated to 3.3-nm intervals to match the sampling intervals of the UniSpec. No smoothing was performed.
D and were compared at the following levels for leaf spectra (numbers 1–6) or crown spectra (number 7):
were calculated where multiple measurements were made per individual leaf. For example, six measurements per leaf were recorded for the Parque Natural Metropolitano March 2003 data set (Table 2). Therefore, D and were assessed for all 15 possible pairs of the six spectra (called pairwise combinations hereafter). were then calculated on all pairwise combinations of spectra from those leaves. were computed for all pairwise combinations of the mean tree spectra. were made between mean tree spectra of the same species measured in the dry season and wet season. These comparisons were possible at the Parque Natural Metropolitano and Fort Sherman (Panama) sites only where sunlit leaves were collected from the same trees and measured with the same spectrometer in both seasons (Table 2). This comparison was not performed for Los Inocentes (Costa Rica) since the November 2002 and March 2003 data sets were obtained using different spectrometers. comparisons were possible for species sampled at more than one site. There were seven such species, all measured with the FieldSpec HandHeld: Bombacopsis quinata (Malvaceae), Bursera simaruba (Burseraceae), Byrsonima crassifolia (Malpighiaceae), Cedrela odorata (Meliaceae), Gliricidia sepium (Fabaceae/Papilionoideae), Guazuma ulmifolia (Sterculiaceae), and Hymenaea courbaril (Fabaceae/Caesalpinoideae). were computed for all pairwise combinations of the five mean crown spectra at Fort Sherman, Panama.
A Wilcoxon rank sum test was used to determine whether levels 3 and 6 differed significantly. This test indicated whether comparisons of individual conspecific tree crowns differed from comparisons among species. The test was performed using mean values of D and for comparisons of conspecific trees (level 3) and comparisons of heterospecific species pairs (level 6).
Classification
Beyond examination of shape and amplitude differences between spectra, we determined whether it was possible to discriminate between species based on reflectance spectra. Steps involved in the classification included feature selection, classifier training, and classifier testing.
Forward stepwise feature selection reduced the number of variables entered into the classification. Stepwise feature selection has proven useful for studies involving species classification from hyperspectral data (van Aardt and Wynne, 2001
; Clark et al., 2005
), and initial testing of our data sets led to higher classification accuracies than using principal components analyses as described in our earlier work (Castro-Esau et al., 2004
) for discriminating liana and tree species. The approach for forward feature selection involves selecting the single best feature and consecutively adding those features that improve performance the most. In this case, the criterion used for the procedure was first nearest neighbor leave-one-out classification performance (Duin, 2000
). By using alternate spectra, we split the data array in half into training and testing sets. Feature selection was performed using the training set only.
Spectra were classified to species using a selection of supervised parametric and nonparametric classifiers, previously described in Castro-Esau et al. (2004)
. The classifiers are part of a pattern recognition toolbox, PRTools, developed by Duin (2000)
for use within Matlab (MathWorks, Natick, Massachusetts, USA). Parametric classifiers included a logistic linear classifier (loglc) and a normal density-based quadratic classifier (qdc). Nonparametric classifiers included a decision tree classifier (treec), a neural network classifier (based on back-propagation and Levenberg-Marquardt gradient descent) (lmnc), and a k-nearest neighbor classifier (knnc).
Classification involved several steps. Initially, training data were labeled according to species and used to train the classifiers. Secondly, the trained classifiers classified the test data set to validate the robustness of the model (Defernez and Kemsley, 1997
). Error estimation was determined for both the training and testing data, which constituted the percentage of incorrectly classified spectra in each case. For each data set, classifications were run several times using multiples of 10 selected features. The highest classification accuracies were reported. The classifications involved all leaf spectra for each species (a minimum of five spectra, but in most cases 10+). Classifications were run on noninterpolated data in the range 445.9–949.6 nm for the UniSpec and 445.5–950.7 nm for the FieldSpec HandHeld.
A final analysis was run to determine the potential for classifying increasing numbers of tropical forest tree species. A data set of 50 species with 10 spectra each was compiled from Parque Natural Metropolitano, Fort Sherman, and Chamela, all using UniSpec 2003 wet season data. Classifications were run on 10000 sets each of 5, 10, 15, 20, 25, 30, 35, 40, and 45 species, where each set of species was chosen randomly from the 50 total species. A single classification was run on the full set of 50 species. For this analysis, we used the logistic linear classifier and 30 previously selected features, the top 10 features from the within-site classifications for each of the three sites. Ten thousand repetitions were considered a representative sample of all possible species combinations since variation around mean classification errors stabilized by this point, as determined by testing with greater numbers of repetitions.
RESULTS
D and
With respect to the goal of classifying trees according to species, it is evident that the magnitude of variability in amplitude (D) and shape () within each species presents a challenge (Figs. 2 and 3). Levels of comparison to the left of the between-species category are measures of intraspecific variability. While there is a general increase in mean D (Fig. 2) and (Fig. 3) with increasing level of comparison or "scale" from within-leaf measurements through to between-species measurements (with the exception of the between-sites category for the FieldSpec HH data), there is a high degree of overlap between each of these levels. Even so, a significant difference (P < 0.05) was found between level 3 (between tree) and level 6 (between species) with the Wilcoxon rank sum test, indicating that for one site at one date, variability between species is greater than variability between individuals of the same species. At the crown level, amplitude differences between all possible pairwise combinations of spectra of the five species were much higher than for any other category (Fig. 2a). This was particularly evident in the near-infrared region (Fig. 4).
|
|
|
|
|
Classification
Species from one site and season were classified with high accuracy (Table 5). Not unexpectedly, the highest classification error accompanies the site for which the largest number of species was sampled (Chamela, Mexico, 27 species), indicating a potential problem for complex tropical dry forest ecosystems in which species density can exceed 50–70 species per ha (Kalacska et al., 2004
). Still, accuracy on the test data was 80%.
|
Accuracy was dependent on number of features and classifier. Highest classification accuracies were obtained with 20–40 features, beyond which classification accuracy eroded due to overfitting. The frequency at which wavebands in various regions of the spectrum were selected for the various species-level classifications, based on all classified sites and the top 20 selected features per site, was higher for visible (400–700 nm) bands than for near-infrared (700–950 nm) bands (Fig. 6). Classifiers also affected accuracy. For example, using 40 wavebands and a logistic linear classifier produced the lowest overall classification error of 2.67% for the Fort Sherman site (Table 5). The second lowest error for the same site, however, was 14.67% for a k-nearest neighbor classifier (k = 2; not shown). In general, however, reasonably accurate classifications by multiple classifiers at each site lends support for the inherent distinguishability of the species. For four of the six sites, the species-level classification using the logistic linear classifier produced the lowest classification errors.
|
The classification analysis using repeated classifications on random sets of species in multiples of 5, up to a total of 50 species, indicated that, within this range, mean classification error on the test data rises linearly with number of species (Fig. 7). Classification error on the training data was 0% in all cases. For the test data, it ranged from 13.79% for five species to 22.40% for 50 species. Variability was greater with fewer species since, from the random selection of species, entirely different sets of species were selected often. With a higher number of species, the same species would have entered into multiple classifications. It was noted that, below five species, the linear trend shown in Fig. 7 was not followed (e.g., for two classes, test data error was only 3.30%).
|
|
|
|
). To test these relationships, chlorophyll data from individual samples (N = 5 samples per species for six species) were paired with bidirectional reflectance data for the same sample. (UniSpec spectral reflectance measurements were taken of the leaf cores that were then designated for chlorophyll analysis.) Leaf thickness was significantly correlated (P < 0.05) with bidirectional reflectance at 532, 548.5, 571.6, and 706.2 nm (not shown).
|
Leaf spectral reflectance is a rich data source for understanding differences in leaf traits among species and for assessing the potential for species discrimination. Differences in leaf traits are the basis for differences in leaf optical properties that permit species discrimination at the leaf level, which within sites and within seasons, appears promising when a limited number of species are present. Because classification accuracy degrades with increasing numbers of species, it is more likely that, for species-rich tropical forests, a portion of species with distinctive reflectance characteristics will be distinguishable from the larger community. At the crown scale, in the first tropical rain forest tree discrimination study using high spectral and spatial resolution imagery, Clark et al. (2005)
also found good potential (92% accuracy) for analyzing a limited number of emergent tree species (seven) from a single-date (1998) HYDICE image of La Selva, Costa Rica, using a linear discriminant analysis and 30 wavebands. These efforts represent the initial stages of operational tropical tree species identification by hyperspectral remote sensors, which will provide an important tool to tropical biologists and ecologists for a variety of applications.
D and (leaf level)
Use of Price's (1994) metrics D and show that variability in a species' reflectance begins within individual leaf blades (Figs. 2 and 3). Multiple spectral measurements taken over the leaf blade show variation, particularly over leaf veins (we avoided the main vein). Similarly, multiple measurements from the SPAD chlorophyll absorbance meter of the same leaf also vary, indicating uneven chlorophyll distribution across the leaf blade (also documented in other species; e.g., Hew et al., 1998
). This variation increases when comparing multiple leaves from the same tree and is compounded further when examining spectra of leaves of different conspecific trees from a single site. Greater differences yet are seen for between-season and between-site spectra of the same species (Fig. 5).
Classification
Classification for each site was successful despite high intraspecific variability, with greater than 80% test data accuracy at all sites. These results may be attributed to the selection of a set of highly discriminating features (wavebands) from the spectral range available and the use of a variety of powerful tools for pattern recognition.
The most frequently selected wavebands in our analyses were in the blue region (400–499 nm), which is strongly influenced by absorption of chlorophylls and carotenoids (Jensen, 2000
). The blue-green edge leading to the green peak (500–549 nm) and chlorophyll absorption leading to the red edge (650–699 nm) were also important (Fig. 6). This finding is similar to Fung et al. (1999)
, who used stepwise linear discriminant analysis for feature selection and found selected bands to lie mainly in the green peak and red edge regions. Overall, however, the bands selected depend on the data used and therefore differ from one suite of species to the next to optimize separability in each case. Wavebands beyond the range 400–950 nm, related to nitrogen and lignin concentrations, as well as O-H stretching, may be useful in future tree discrimination studies, as was the case for Martin et al. (1998)
, van Aardt and Wynne (2001)
, and Clark et al. (2005)
.
The logistic linear classifier, a parametric classifier, provided the best species-level classification results at four of six sites (Table 5). The k-nearest neighbor classifier, a nonparametric classifier, produced the second best results overall. The quadratic classifier generally performed the worst, possibly due to correlations between features. Of the nonparametric classifiers, the decision tree classifier generally performed poorly, likely due to complex trees and the tendency toward overfitting. The decision tree classifier is not well suited to features that have a large number of possible values (Mitchell, 1997
), as was also observed when discriminating between lianas and trees (Castro-Esau et al., 2004
).
The spectrometer used for data collection may possibly influence classification accuracy as well. Instruments with higher spectral resolution will better capture fine spectral details that may be used for discriminating species. In this case, the FieldSpec HandHeld had a finer spectral resolution (3.5 nm at 700 nm) than the UniSpec (<10 nm), but there was no obvious advantage observed in our classification results (Table 5).
Based on our findings, which are at this point mainly limited to the level of the leaf, the potential for using a database of spectral signatures of species for classifying across sites and seasons appears low. The leaf-level spectral signature of a species from one site did not serve effectively to classify the same species at a second site (Fig. 5). Differences in leaf age, phenology, soil mineral nutrition, moisture availability, temperature, and other stresses (Carter, 1993
) could certainly lead to the differences observed in spectral reflectance of the same species over multiple sites with varying edaphic and climatic conditions. In a similar way, progressively increasing stress with increasing elevation has been noted in the spectral reflectance properties of Picea rubens, Abies balsamea (Richardson et al., 2001
), and Betula papyrifera (Richardson and Berlyn, 2002
). Another possible cause of intersite differences in conspecific leaf spectra is that genetic differences between populations can contribute to spectral differences. Winter (1998)
, for instance, was able to distinguish between two different varieties of avocado (Persea americana) in an airborne TRWIS III image. Differences in the spectral characteristics of different genotypes of crops such as barley (e.g., Fetch et al., 2004
) have also been detected. However, more investigation is required in this area with respect to tropical trees.
Differences in leaf optical properties due to seasonality (wet/dry cycles) could be detrimental or beneficial to tree species identification. Temporal phenomena such as flowering, leaf flush, or senescence could be confusing if not well understood, but could become useful in detecting particular species at particular times of the year if understood. In a similar way, the timing of imagery is important in crop discrimination studies (Collins, 1978
; Congalton et al., 1998
; Murakami et al., 2001
). Thus, a creative combination of image processing and classification with in-depth knowledge of tree ecology and phenology will likely be necessary for a successful species-level classification.
Based on our leaf-level classification analysis of 50 species, there is potential for classification of up to 20 species with approximately 85% test data accuracy and approximately 80% test data accuracy for up to 45 species (Fig. 7). Fung et al. (1999)
achieved a similar level of overall accuracy (89%) for 25 subtropical tree species using a set of 13 bands selected using hierarchical clustering (as compared to 83.2% test data accuracy and 91.6% overall accuracy for 25 species, based on our data). For our analysis, five spectra were used for training classifiers and five for testing the classifiers. With greater amounts of training data, such as using a cross-validation approach, classification accuracies could have been higher yet.
Provided the linear relationship in Fig. 7 continues, we may be able to classify test data for 75 species with 73% accuracy, and 100 species with 69% accuracy. Classification accuracy erodes further beyond 100 species. At the leaf level, therefore, it would not be possible to classify all species in species-rich neotropical forests where the number of tree species can exceed 100 even in 0.1-ha plots (Gentry, 1991
), although potential for discrimination would likely be improved using a spectrometer with a greater range (e.g., to 2500 nm). Also, it remains to be determined whether a similar trend holds at the crown level. If it does, classification could still be successful for less species-rich tropical forests or for detecting smaller numbers of tree species with distinct features from the remainder of the canopy.
Spectra and leaf traits
While this paper does not attempt to explore a comprehensive suite of leaf traits, the selected traits measured for species at Parque Natural Metropolitano (chlorophyll content, leaf thickness, percentage of intercellular space, spongy mesophyll width) provide an illustration of leaf feature variability among nine tropical dry forest tree species, which supports the differences observed in leaf reflectance for these same species (Table 6). Where two species differ in one or more of these traits, the resultant differences in spectral reflectance at one or multiple wavebands provide the opportunity for their discrimination.
Pigment content plays a large role in visible reflectance. The range of total chlorophyll content for the selected species at Parque Natural Metropolitano was 496-1055 µmol/m2 and for reflectance at 550 nm, 9.10–15.36% (Table 6). Overall, visible wavebands (400–700 nm) were key to species identification and were emphasized in feature selection (Fig. 6). Leaf thickness and internal leaf morphology appear to control near-infrared reflectance (Gates et al., 1965
; Knipling, 1970
; Knapp and Carter, 1998
). From our samples, histological analysis of leaf thin sections (Fig. 9) can provide unique insight into the relative magnitude of reflectance in the near-infrared. The most striking link encountered was between the low percentage of intercellular space, spongy mesophyll width, and leaf thickness of Astronium graveolens and the species' uniquely low near-infrared reflectance. The leaf thin section of this species revealed a very thin, compact spongy mesophyll layer, allowing little opportunity for interactions of light at air–cell-wall interfaces. In fact, this same species transmitted more light through the leaf than any other species sampled. Stronger correlations with near-infrared bands might have been observed using the ratio of mesophyll cell surface area exposed to intercellular spaces per unit leaf surface area (Slaton et al., 2001
), which provides a better description of air–cell-wall interfaces in the mesophyll than does the percentage of air space (Knipling, 1970
). Water content was not studied here, but would probably correlate with leaf thickness, intercellular space, and infrared reflectance and transmittance (Sims and Gamon, 2003
).
In general, accurate classification is closely associated with the presence of distinctive leaf features within the spectral range considered, as seen from the example of Parque Natural Metropolitano. Distinctive leaf spectral features that facilitate species classification will be tied to leaf traits such as (1) low or high levels of chlorophyll per unit area, (2) thick or thin leaves, and/or (3) low or high percentage of air space in the spongy mesophyll relative to other species.
Crown spectra
Although this discussion focused mainly on leaf spectra and their discrimination, the primary application for this study will be towards tree crown identification. How then, can knowledge about intraspecific variability in leaf reflectance be applied to future crown-level tree classifications from airborne or satellite-borne imagery? First, our results will provide a partial basis for interpreting variability in tree crown reflectance spectra within and between sites and seasons, since leaf reflectance often comprises the major component of an overall crown reflectance spectrum (with exceptions, especially in dry-season conditions in which many trees are deciduous). Unlike leaf spectra, tree crown spectra obtained from airborne or satellite-borne sensors are influenced by leaf area index (leaf density), leaf angle distribution, crown shape and shading, and background signals such as tree bark and soil. These additional factors may help or hinder tree crown classification. For species that have unique and predictable combinations of these factors, their overall crown spectral signatures may be even more distinguishable than their leaf spectral signatures. However, this has not been substantiated yet, and recent work (J. L. Zhang, B. Rivard, G. A. Sánchez-Azofeifa, and K. L. Castro-Esau, unpublished manuscript) indicates that intra-crown variation within species may pose a challenge. In a tropical environment, visual inspection of the crown spectra we recorded at Fort Sherman, Panama, indicate that those five crowns would be distinguishable without difficulty, mainly due to large differences in spectral amplitude and, to a lesser extent, differences in shape (Figs. 1 and 4).
Classification at the canopy level is clearly the next step in determining whether tropical tree species are differentiated and requires airborne or satellite-borne imagery with a combination of both high spectral and spatial resolution. Clark et al. (2005)
presented pioneering work in hyperspectral tropical rain forest tree species discrimination at leaf to crown scales, which will likely lead to a burgeoning of research in this field. Additional efforts so far, using aerial photographs (Trichon, 2001
) and high spatial resolution multispectral data (reflectance measured over a small number [four] of broad bands, each typically >50 nm wide) such as IKONOS (1 m and 4 m resolution) and Quickbird (0.7 m and 2.8 m resolution; Read et al., 2003
; Wang et al., 2004
), also support the possibility of separating a portion of the total number of tropical species. Combined with LiDAR (Light Detection And Ranging, an active remote sensor that measures return time for laser pulses), which can provide information on tree location and height as well as crown diameter and shape, the possibilities may extend even further (Gillespie et al., 2004
). As these types of data become increasingly available for tropical forests, scientists may draw upon techniques developed for automated tree isolation (Wulder et al., 2000
; Leckie et al., 2003
) and for tree species classifications in ecosystems with more limited species diversity, such as arid environments (Lewis, 2000
; Lewis et al., 2001
) and temperate forests (Martin et al., 1998
; Key et al., 2001
; Roberts et al., 2004
). The ability to accurately map tree species in tropical ecosystems will represent a significant advance that will facilitate ecosystem characterization, tree demographic studies, mapping endangered or endemic species, identifying important food sources for wildlife, and quantifying carbon pools and carbon sequestration rates.
FOOTNOTES
1 The authors thank M. Samaniego, L. Coronado-Chacón, M. Kachmar, K. Wolbert, M. Kalacksa, M. Castro, and L. Cheng for their field, laboratory, and/or technical assistance; and two anonymous reviewers who were helpful in improving the manuscript; and acknowledge the Panamanian Autoridad Nacional del Ambiente (ANAM) for collection and export permits, the Ministerio del Ambiente y Energía (MINAE) of Costa Rica, the SEMANART of Mexico and the extraordinary support of the personnel of the Santa Rosa National Park in Costa Rica. An Institutional Grant from the Tinker Foundation to the University of Alberta, the Inter-American Institute for Global Change Research (IAI) and the Canadian International Development Agency (CIDA) provided funds for this work. K.C. was funded by an NSERC PGS-B scholarship and a Smithsonian Tropical Research Institute (STRI) Supplementary Research Award. 
5 Author for correspondence (arturo.sanchez{at}ualberta.ca
), phone: 1-780-492-1822; fax: 1-780-492-2030.
LITERATURE CITED
Allen W. A. Gausman H. W. Richardson A. J. Wiegand C. L.. 1970. Mean effective optical constants of thirteen kinds of plant leaves. Applied Optics 9: 2573-2577.[Web of Science]
Angiosperm Phylogeny Group (APG II).. 2003. An update of the Angiosperm Phylogeny Group classification for the orders and families of flowering plants: APG II. Botanical Journal of the Linnean Society 141: 399-436.[CrossRef]
Arnon D. I.. 1949. Copper enzymes in isolated chloroplasts: polyphenoloxidase in Beta vulgaris. Plant Physiology 24: 1-15.
Avalos G. Mulkey S. S.. 1999. Seasonal changes in liana cover in the upper canopy of a neotropical dry forest. Biotropica 31: 186-192.[Web of Science]
Avalos G. Mulkey S. S. Kitajima K.. 1999. Leaf optical properties of trees and lianas in the outer canopy of a tropical dry forest. Biotropica 31: 517-520.[CrossRef][Web of Science]
Bullock S. H. Solis-Magallanes A.. 1990. Phenology of canopy trees of a tropical deciduous forest in Mexico. Biotropica 22: 23-35.
Carter G. A.. 1993. Responses of leaf spectral reflectance to plant stress. American Journal of Botany 80: 239-243.[CrossRef][Web of Science]
Castro-Esau K. L. Sánchez-Azofeifa G. A. Caelli T.. 2004. Discrimination of lianas and trees with leaf-level hyperspectral data. Remote Sensing of Environment 90: 353-372.[CrossRef][Web of Science]
Castro-Esau K. L. Sánchez-Azofeifa G. A. Rivard B.. In press. Comparison of spectral indices obtained using multiple spectroradiometers. Remote Sensing of Environment..
Clark M. L. Roberts D. A. Clark D. B.. 2005. Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sensing of Environment 96: 375-398.[CrossRef][Web of Science]
Cochrane M. A.. 2000. Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing 21: 2075-2087.[CrossRef][Web of Science]
Collins W.. 1978. Remote-sensing of crop type and maturity. Photogrammetric Engineering and Remote Sensing 44: 43-55.[Web of Science]
Congalton R. G. Balogh M. Bell C. Green K. Milliken J. A. Ottman R.. 1998. Mapping and monitoring agricultural crops and other land cover in the Lower Colorado River Basin. Photogrammetric Engineering and Remote Sensing 64: 1107-1113.[Web of Science]
Defernez M. Kemsley E. K.. 1997. The use and misuse of chemometrics for treating classification problems. Trends in Analytical Chemistry 16: 216-221.[CrossRef]
Duin R. P. W.. 2000. PRTools, version 3.0: a Matlab toolbox for pattern recognition Pattern Recognition Group, Delft University of Technology, Delft, Netherlands.
Fetch T. G. Jr Steffenson B. J. Pederson V. D.. 2004. Predicting agronomic performance of barley using canopy reflectance data. Canadian Journal of Plant Science 84: 1-9.[Web of Science]
Foley S. L. Rivard B. Sánchez-Azofeifa G. A. Calvo-Alvarado J.. In press. Foliar spectral properties following leaf clipping and implications for handling techniques. Remote Sensing of Environment..
Frankie G. W. Baker H. G. Opler P. A.. 1974. Comparative phenological studies of trees in tropical wet and dry forests in the lowlands of Costa Rica. Journal of Ecology 62: 881-899.[CrossRef]
Fung T. Siu W. L.. 1998. Hyperspectral data analysis for subtropical tree species recognition. Proceedings of symposium on IEEE International Geoscience and Remote Sensing, 1998, Sensing and Managing the Environment, vol. 3, 1298–1300 IEEE International Geoscience and Remote Sensing, Seattle, Washington, USA.
Fung T. Ma F. Y. Siu W. L.. 1999. Band selection using hyperspectral data of subtropical tree species. Proceedings of Asian Conference on Remote Sensing, 1999, Hong Kong, China, available at http://www.gisdevelopment.net/aars/acrs/1999/ps3/ps3055pf.htm.
Gamon J. A. Field C. B. Bilger W. Bjorkman O. Fredeen A. L. Peñuelas J.. 1990. Remote-sensing of the xanthophyll cycle and chlorophyll fluorescence in sunflower leaves and canopies. Oecologia 85: 1-7.
Gates D. M. Keegan H. J. Schleter J. C. Weidner V. R.. 1965. Spectral properties of plants. Applied Optics 4: 11-20.
Gausman H. W.. 1984. Evaluation of factors causing reflectance differences between sun and shade leaves. Remote Sensing of Environment 15: 177-181.[CrossRef][Web of Science]
Gausman H. W.. 1985. Plant leaf optical properties in visible and near-infrared light. Graduate Studies no. 29 Texas Tech Press, Lubbock, Texas, USA.
Gausman H. W. Allen W. A.. 1973. Optical parameters of leaves of 30 plant species. Plant Physiology 52: 57-62.
Gentry A. H.. 1991. The distribution and evolution of climbing plants. In F. E. Putz and H. A. Mooney [eds.], The biology of vines Cambridge University Press, Cambridge, UK.
Gillespie T. W. Brock J. Wright C. W.. 2004. Prospects for quantifying structure, floristic composition and species richness of tropical forests. International Journal of Remote Sensing 25: 707-715.[CrossRef][Web of Science]
Gitelson A. Merzlyak M. N.. 1994. Spectral reflectance changes associated with autumn senescence of Aesculus hippocastanum L. and Acer platanoides L. leaves: spectral features and relation to chlorophyll estimation. Journal of Plant Physiology 143: 286-292.[Web of Science]
Hew C. S. Soh W. P. Ng C. K. Y.. 1998. Variation in photosynthetic characteristics along the leaf blade of Oncidium goldiana, a C-3 tropical epiphytic orchid hybrid. International Journal of Plant Sciences 159: 116-120.[CrossRef]
Hiscox J. D. Israelstam G. F.. 1979. Method for the extraction of chlorophyll from leaf tissue without maceration. Canadian Journal of Botany 57: 1332-1334.[Web of Science]
Hoffer R. M. Johannsen C. J.. 1969. Ecological potentials in spectral signature analysis. In P. L. Johnson [ed.] Remote sensing in ecology 1-16 University of Georgia, Athens, Georgia, USA.
Jensen J. R.. 2000. Remote sensing of the environment: an earth resource perspective Prentice-Hall, Upper Saddle River, New Jersey, USA.
Kalacska M. Sánchez-Azofeifa G. A. Calvo-Alvarado J. C. Quesada M. Rivard B. Janzen D. H.. 2004. Species composition, similarity and diversity in three successional stages of a seasonally dry tropical forest. Forest Ecology and Management 200: 227-247.[CrossRef][Web of Science]
Key T. Warner T. A. McGraw J. B. Fajvan M. A.. 2001. A comparison of multispectral and multitemporal information in high spatial resolution imagery for classification of individual tree species in a temperate hardwood forest. Remote Sensing of Environment 75: 100-112.[CrossRef][Web of Science]
Knapp A. K. Carter G. A.. 1998. Variability in leaf optical properties among 26 species from a broad range of habitats. American Journal of Botany 85: 940-946.[Abstract]
Knipling E. B.. 1967. Physical and physiological basis for differences in reflectance of healthy and diseased plants. Proceedings of Workshop on infrared color photography and diseased plants Florida Department of Agriculture, Winter Haven, Florida, USA.
Knipling E. B.. 1970. Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation. Remote Sensing of Environment 1: 155-159.
Leckie D. G. Gougeon F. A. Walsworth N. Paradine D.. 2003. Stand delineation and composition estimation using semi-automated individual tree crown analysis. Remote Sensing of Environment 85: 355-369.[CrossRef][Web of Science]
Lee D. W. Graham R.. 1986. Leaf optical-properties of rain-forest sun and extreme shade plants. American Journal of Botany 73: 1100-1108.[CrossRef][Web of Science]
Lee D. W. Bone R. A. Tarsis S. L. Storch D.. 1990. Correlates of leaf optical-properties in tropical forest sun and extreme-shade plants. American Journal of Botany 77: 370-380.[CrossRef][Web of Science]
Lewis M.. 2000. Discrimination of arid vegetation composition with high resolution CASI imagery. Rangeland Journal 22: 141-167.
Lewis M. Jooste V. de Gasparis A. A.. 2001. Discrimination of arid vegetation with airborne multispectral scanner hyperspectral imagery. IEEE Transactions on Geoscience and Remote Sensing 39: 1471-1479.[CrossRef][Web of Science]
Lott E. J. Bullock S. H. Solís-Magallanes J. A.. 1987. Floristic diversity and structure of a tropical deciduous forest of coastal Jalisco. Biotropica 19: 228-235.[CrossRef][Web of Science]
Markwell J. Osterman J. C. Mitchell J. L.. 1995. Calibration of the Minolta SPAD-502 leaf chlorophyll meter. Photosynthesis Reseach 46: 467-472.[CrossRef]
Martin M. E. Newman S. D. Aber J. D. Congalton R. G.. 1998. Determining forest species composition using high spectral resolution remote sensing data. Remote Sensing of Environment 65: 249-254.[CrossRef][Web of Science]
Mitchell T. M.. 1997. Machine learning McGraw-Hill, New York, New York, USA.
Murakami T. Ogawa S. Ishitsuka N. Kumagai K. Saito G.. 2001. Crop discrimination with multitemporal SPOT/HRV data in the Saga Plains, Japan. International Journal of Remote Sensing 22: 1335-1348.[CrossRef][Web of Science]
Price J. C.. 1994. How unique are spectral signatures?. Remote Sensing of Environment 49: 181-186.[CrossRef][Web of Science]
Read J. M. Clark D. B. Venticinque E. M. Moreira M. P.. 2003. Application of merged 1-m and 4-m resolution satellite data to research and management in tropical forests. Journal of Applied Ecology 40: 592-600.[CrossRef][Web of Science]
Richardson A. D. Berlyn G. P. Gregoire T. G.. 2001. Spectral reflectance of Picea rubens (Pinaceae) and Abies balsamea (Pinaceae) needles along an elevational gradient, Mt. Moosilauke, New Hampshire, USA. American Journal of Botany 88: 667-676.
Richardson A. D. Berlyn G. P.. 2002. Spectral reflectance and photosynthetic properties of Betula papyrifera (Betulaceae) leaves along an elevational gradient on Mt. Mansfield, Vermont, USA. American Journal of Botany 89: 88-94.
Richardson A. D. Duigan S. P. Berlyn G. P.. 2002. An evaluation of noninvasive methods to estimate foliar chlorophyll content. New Phytologist 153: 185-194.[CrossRef][Web of Science]
Roberts D. A. Ustin S. L. Ogunjemiyo S. Greenberg J. Dobrowski S. Z. Chen J. Q. Hinckley T. M.. 2004. Spectral and structural measures of northwest forest vegetation at leaf to landscape scales. Ecosystems 7: 545-562.[Web of Science]
Salem B. B. Elcibahy A. Elraey M.. 1995. Detection of land-cover classes in agroecosystems of northern Egypt by remote-sensing. International Journal of Remote Sensing 16: 2581-2594.[CrossRef][Web of Science]
Sims D. A. Gamon J. A.. 2003. Estimation of vegetation water content and photosynthetic tissue area from spectral reflectance: a comparison of indices based on liquid water and chlorophyll absorption features. Remote Sensing of Environment 84: 526-537.[CrossRef][Web of Science]
Slaton M. R. Hunt E. R. Smith W. K.. 2001. Estimating near-infrared leaf reflectance from leaf structural characteristics. American Journal of Botany 88: 278-284.
Trichon V.. 2001. Crown typology and the identification of rain forest trees on large-scale aerial photographs. Plant Ecology 153: 301-312.[CrossRef][Web of Science]
van Aardt J. A. N. Wynne R. H.. 2001. Spectral separability among six southern tree species. Photogrammetric Engineering and Remote Sensing 67: 1367-1375.[Web of Science]
Vogelmann T. C.. 1993. Plant-tissue optics. Annual Review of Plant Physiology and Plant Molecular Biology 44: 231-251.[CrossRef][Web of Science]
Wang L. Sousa W. P. Gong P. Biging G. S.. 2004. Comparison of IKONOS and QuickBird images for mapping mangrove species on the Caribbean coast of Panama. Remote Sensing of Environment 91: 432-440.[CrossRef][Web of Science]
Winter E. M.. 1998. Classification of vegetation types using a high spectral and spatial resolution hyperspectral sensor. In H. Fujisada [ed.] Proceedings of EUROPTO Conference on Sensors, Systems, and Next-Generation Satellites II, vol. 3498 216-222 The International Society for Optical Engineering (SPIE), Barcelona, Spain.
Woolley J. T.. 1971. Reflectance and transmittance of light by leaves. Plant Physiology 47: 656-662.
Wulder M. Niemann K. O. Goodenough D. G.. 2000. Local maximum filtering for the extraction of tree locations and basal area from high spatial resolution imagery. Remote Sensing of Environment 73: 103-114.[CrossRef][Web of Science]
CiteULike Complore Connotea Del.icio.us Digg Facebook Reddit Technorati Twitter What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
