Identifying tropical Ecuadorian Andean trees from inter-crown pixel distributions in hyperspatial aerial imagery

Authors


corresponding author, m.r.peck@sussex.ac.uk

Abstract

Question

Identification of tropical tree species from low-cost, very high-resolution (VHR) proximal canopy remote sensing imagery has great potential for improving our understanding of tropical forest ecology. We investigated whether inter-crown pixel information from VHR imagery could be used for taxonomic identification of trees and characterization of successional stages of tropical Andean mountain forest.

Location

Santa Lucia cloud forest reserve, NW Ecuadorian Andes.

Methods

We gathered digital camera imagery (0.05-m spatial resolution), using a remote-controlled helicopter platform, from primary and secondary forest and identified visible crowns to species before extracting digital crown samples (2-m radius). Using an object-based approach, histogram descriptors and diversity metrics of pixel intensity in red, green and blue (RGB) bands were calculated for crowns, and patterns in similarity explored using ordination. Predictive models were developed and validated using four decision tree models (CHIAD, Exhaustive CHIAD, CRT and QUEST).

Results

Aerial imagery represented 54% of families, 53% of genera and 56% of species sampled from the ground. Ordination (redundancy analysis) confirmed that inherent continuities, based on crown metrics, correlated with traditional species, genus and family groupings (< 0.05). Data were best described by histogram means in the green band. The best predictive model (CRT) generated a 47% probability of correct species identification for 41 species – with similar success at genus and family level. Predictive ability was highly species-specific, ranging from zero for some taxa to 93% for Cecropia gabrielis Cuatrec.

Conclusions

From the crown metrics tested, we found the mean pixel intensity in the green band was most effective in predicting species and species grouping of tropical mountain trees. This metric integrates species-specific differences in leaf density of crowns and reflectance in the green waveband. High predictive success for indicators of primary (Cornus peruviana J.F. Macbr.) and secondary forest (Cecropia gabrielis Cuatrec.) shows that VHR imagery can be used to identify species from pixel information to provide ecological information on successional status. Further research is needed to develop pattern and textural metrics specifically for hyperspatial digital imagery to identify tree species from crown imagery in diverse tropical forests.

Nomenclature
Gentry

(1996)

Introduction

Collecting ecological data in tropical rain forests is difficult since they feature high biodiversity and are difficult to access (Gentry 1992; Condit et al. 2000). It is thus often the case that ecological surveys can be time-consuming, labour intensive and therefore expensive (Trichon & Julien 2006). Recent work in the tropical forest environment has begun to focus on rapid inventory methods (Sutton 2001) that allow conservationists, environmental scientists, land managers and carbon market-related studies to quickly, accurately and efficiently gain an understanding of the ecology of large tracts of tropical forest in relatively little time (Leckie et al. 2003). To this end, remote sensing using satellite imagery and airborne sensors has generated high expectations for biodiversity assessment, however it is still in relatively early stages of development (Nagendra & Rocchini 2008). Most work on the use of imagery with high-spatial resolution to identify forest canopy species comes from temperate regions – as the tropics offer greater challenges due to higher species number and canopy complexity, and frequent cloud cover. In spite of the difficulties, significant advances in aerial photographic survey techniques have been made over the last two decades, and there is a large body of work that aims to discriminate tree species based on their biochemistry and spectral properties (e.g. Asner & Martin 2009, 2011). Access to high-resolution photographs of larger areas of forest has resulted in progress in individual crown canopy delineation (Brandtberg & Walter 1998; Culvenor 2002; Ke et al. 2010) and the identification of dominant emergents, but there has not been a particularly large body of work focused on the taxonomic identification of canopy species, least of all in tropical forests where the need for biodiversity assessment and monitoring is perhaps most critical. The studies in tropical forests using high-resolution imagery have mainly focused on the development of taxonomic identification techniques for identification of crowns by eye (Trichon 2001; Trichon & Julien 2006; Gonzalez-Orozco et al. 2010). Trichon (2001) investigated the identification of tropical forest tree crowns in non-digital photographs. Stereoscopic photographs were analysed manually and crowns assessed using a number of typology ‘keys’ based on physiological traits such as crown size, colour, architecture and phenology with images (of 0.3-m resolution) taken from 100 m above the forest canopy using a hot-air ship. Trichon found that this method enabled accurate determination of the 12 dominant species in the photographs due to striking and consistent characteristics particular to each taxonomic group. In a follow-on paper, Trichon & Julien (2006) highlighted the issue of intra-species and intra-crown variation in crown appearance and their impact on analysis success rate. It is notable that in this study the authors found that only approximately 25% of canopy trees were identifiable in photographs. This was due to the small diameter of many crowns and the overall heterogeneity of the forest canopy that would not allow isolation of particular crowns. For the Ecuadorian Amazon, Gonzalez et al. (2010) developed ‘aerial taxonomy’ keys for crowns in high-resolution imagery, resulting in an overall identification accuracy of 50% – with three taxa showing identification success of over 70%. All these studies were based on methods to guide human observers. Hence, there is a need to develop automated methods for crown identification from high-resolution imagery. Little information is currently available on the potential of algorithms to identify crowns in tropical canopies that could lead to automated methods for scanning aerial imagery in the tropics. One of the main reasons for this research gap is that imagery is often not available at the pixel resolution required for an analysis of within-crown patterns. The 41-cm resolution of GEOEYE-1 is the highest commercially available spatial resolution. Current studies thus had to collect imagery of sufficient resolution independently, using a variety of methods that included low-altitude aerial photography and helium balloon camera platforms (Trichon & Julien 2006; Gonzalez-Orozco et al. 2010). The perception that higher resolution imagery will lead to improved identification has also been challenged by Nagendra & Rocchini (2008), who suggest that high-pixel resolution imagery may actually make species identification more complicated due to increased variability introduced from shadow and pattern.

In this paper we investigate whether pixel distribution information from very high-resolution (VHR) digital photographic imagery (0.05-m pixel resolution) collected with proximal remote sensing techniques (a remote control helicopter photographic platform) in the tropical Ecuadorian Andean cloud forests provides information that enables the classification of crowns into existing taxonomic groupings. Our main aim was to investigate relationships in pixel heterogeneity in the red, green and blue spectral bands (RGB) of very high-resolution digital imagery of canopy crown species. We aimed to determine whether the distribution patterns of pixels in RGB are sufficiently similar to those of another from the same species, genus or family so that a pattern signature for each species group can be identified.

Methods

Study area

Fieldwork was conducted at the Santa Lucia cloud forest Reserve in the Western Ecuadorian Andes, Pichincha Province. The reserve is located just north of the equator (0°7′30″ N, 78°40′30″ W), approximately 60 km NW of the capital, Quito (Fig. 1). The reserve ranges from an altitude of 1300 m a.s.l. to over 2600 m a.s.l. covering an area of 756 ha. Topography is defined by steep sloping valley systems of varying aspect. The forest in the study area is lower montane rain forest, commonly referred to as cloud forest. It lies within the tropical Andes biodiversity hotspot (Myers et al. 2000), exhibiting high plant species diversity. Annual rainfall in the project area ranges from 1500 to 2800 mm, and average annual temperature is 16 °C (Rivas-Martinez & Navarro 1995). Based on satellite image analysis, ca. 80% of the reserve area is primary forest, with the remainder a mixture of regenerating secondary forests (of between 15 and 25 yr old) and silvopasture. Due to cloud cover, habitats such as cloud forest often lack any clear satellite imagery. Hence, local habitat assessments often require collection of independent aerial imagery.

Figure 1.

Location of the study site in the Ecuadorian Andes.

Image acquisition

A Canon PowerShot A650 12.1 megapixel digital camera (Canon, New York, NY, USA) with a remotely operated shutter was mounted on a gimballed platform on a remote-controlled helicopter (Fig. 2).

Figure 2.

Camera platform mounted below remote-controlled helicopter.

The helicopter was sent out on pre-determined flight paths to capture proximal canopy remote sensing imagery across the survey area. Two sites were chosen to allow the operator of the helicopter camera platform a view of the plots during image acquisition and to ensure the helicopter remained within range of the radio-control device. We raised geomarkers of foil balloons into the canopy and placed reflective foil markers at various prominent or significant locations in the sampling area and recorded their coordinates using a GPS (Garmin CSx60, Garmin International Inc., Olathea, KS, USA). The flights took place between 11:00 and 14:00 hr on the 2 and 3 Oct 2008 to ensure that light conditions were similar for each group of images. Images were assessed for clarity and resolution. A maximum pixel size of 5 cm was established as a threshold for subsequent analysis. Laminated printouts of individual high-quality images were used in the field to ‘ground-truth’ crowns in the imagery with crowns in the field that were identified to species level. For the primary forest site, 17 aerial images were collected, covering a sampling area of 2.15 ha. Imagery for the secondary site covered an area of 2.47 ha for analysis. The composite of georeferenced imagery is shown in Fig. 3.

Figure 3.

Georeferenced composite of all aerial images used in the analysis for the secondary (a) and primary (b) forest sites.

Fieldwork to identify tree crowns in the imagery

We collected data from the primary and regenerating secondary forest site during two fieldwork periods, 20 June–3 Aug 2008 and 17 June–1 Sept 2009. GPS coordinates lacked the resolution to identify individual trees so we used laminated printouts of aerial imagery to match visible crowns to trees in the field. Each selected tree was identified and tagged with an individual number. For each tree we took the following measurements: height and crown diameter, diameter at breast height (DBH), slope angle, aspect, phenology (flowering status of the tree crown), GPS coordinates and altitude. Where possible, we identified trees to species in the field; where this was not possible we took samples from the canopy using a catapult and saw system for identification at the National Herbarium of Ecuador (QNCE) in Quito. We identified potential canopy indicator species of forest successional stage by their relative abundance in the two forest types for species that represented a minimum of 5% of total samples.

Processing of the images

Raw data set

Aerial imagery was imported into ESRI Arcmap Version 9.2 (ESRI, Redlands, CA, USA) and georeferenced using the geomarkers. We georeferenced each photographic image individually rather than use the composite image (Fig. 3) to minimize image distortion. A least-squares affine transformation was applied and only images with a root-mean square error (RMSE) of <0.15 m were analysed. Once georectified, the individual red, green and blue bands that made up each image were imported into the GIS, allowing us to sample each reflectance band. A circular ‘cut-out’ of 2-m radius was generated to sample the three spectral layers for each crown (Fig. 4). Crowns of <4 m in diameter were rejected. A histogram of pixel intensity (ranging from 0 to 255) was generated for each crown and each band. All data were subsequently reduced to 25 bins, and each bin standardized by dividing by total pixel number (Fig. 5); this formed the raw data set.

Figure 4.

Sampling of canopy crowns in secondary forest from band 3 (Red) using 2-m radius extractions.

Figure 5.

From left to right; crown samples from band 3 (Blue) at 0.05-m pixel resolution, histogram of 256 bins exported from GIS software and final histogram of raw data (with histograms for band 1 and 2 also shown).

For each sample and band, crown characteristics were calculated from the frequency distribution. We calculated the mean, standard deviation, skew and kurtosis and, using the CANOCO software package (CANOCO for Windows® Version 4.5; Microcomputer Power, Ithaca, NY, US), landscape diversity indices for the crown pixel distributions: number of samples (s), log number samples (ln s), H Shannon index inline image where pj is the abundance of the j-th pixel intensity, N2 diversity inline image where pi is the fraction of the i-th pixel intensity, N1 richness (N1 = eH), and N2/N1 evenness (E = N2/N1).

Standardized data set

Preliminary ordination on the raw data set identified the mean values of histograms as the most significant descriptors between images and within crowns. As a result, crown samples from different individual aerial photographs that make up the composite data set were standardized by adjusting their means to a global image mean (determined as the mean of all image pixel histograms).

Analyses

We explored the raw and standardized data sets for comparative continuities in composition using linear redundancy analysis in CANOCO (Plant Research International, Wageningen, The Netherlands). We then used the classification and regression tree modelling procedures in SPSS (SPSS statistics, v. l 17.0.0, 2008; SPSS Inc., Chicago, IL, USA) to create tree-based classification models to predict crown taxonomic groups at species, genus and family level. Classification and regression tree modelling are used in exploratory data analysis to study the relationships between a dependent measure and a large series of possible predictor variables, which themselves may interact. We chose tree-based classification methods as they generate decision tree outputs that closely resemble the traditional taxonomic keys used in botanical identification. For a model that provides a good fit, the decision tree generated allows a user (or algorithm) to identify the species of a new crown by following dichotomous or multi-way branches (nodes). Splitting decisions are based on values of crown image features that best explain the split at each level. We applied and compared the predictive ability of four tree growing methods: CHIAD, Exhaustive CHAID, CRT and QUEST. CHIAD (chi-squared automatic interaction detector) uses multi-way splits based on adjusted significance testing (Bonferroni testing, P < 0.05; Kass 1980). At each step the predictor variable with the strongest interaction with the dependent variable is chosen for the split. Exhaustive CHIAD is a modification of the basic CHAID algorithm that performs a more thorough merging and testing of predictor variables by examining all possible splits for each predictor (we used the default values of 100 iterations with minimum change in expected cell frequencies at 0.05). CRT (classification and regression tree) analysis produces a binary decision tree output using the Gini measure of impurity to split the data into segments that are as homogeneous as possible with respect to the dependent variable. Terminal nodes are defined as those in which the dependent variables (species/genus or family, dependent on level of analysis) are the same. QUEST, (quick, unbiased and efficient statistical tree) is a fast binary-split decision tree algorithm that employs a modification of recursive quadratic discriminant analysis (Shih 1999). QUEST splits levels at nodes based on multiple predictors at a defined significance level (we used the default value of P < 0.05). Nisbet et al. (2009) provide a detailed description of the applied techniques. All models were set to a minimum of a single crown in parent and child nodes and were cross-validated using ten sample folds.

Results

Data exploration – ordination

For the raw data set, the nominal environmental descriptors (species, genus and family groupings) were significantly correlated with the first, and all, canonical axes (RDA Monte Carlo permutation test, P < 0.001). The sum of all canonical eigenvalues (variance explained) ranged from 31.8% for species to 26.2% for genera and 16.6% for family groupings, indicating that greater variability was introduced when grouping species into higher taxonomic units. For species, the first axis explained 26.7% of the variance, with subsequent axes explaining the remaining 5.1%. For genera, 21.9% is explained by first axis with 4.3% explained by remaining axes, and for family taxonomic groupings 11% is explained by first axis with 5.6% explained by subsequent axes. The biplot for species in Fig. 6 identifies the mean value of the histogram of pixel intensity, in band 2 (green), as best correlated with the first axis – and explaining most multidimensional patterns in the data set. It is interesting to note that differences between individual aerial images are correlated with the second axis and not significantly explained by either means or standards deviations in histograms.

Figure 6.

Redundancy analysis (RDA) biplot showing the ‘image features (B1_mean – red band histogram mean; B2_mean – green band histogram mean; B3N2div – N2 diversity index for blue band) with >19% first axis fit and nominal environmental variables (species groups and image number) outside correlation range −0.1 to 0.1. Green pixel intensity mean (B2_MEAN) is the best explanatory variable. Species 22 (Miconia clathrantha Triana ex Cogn.) is characterized by high mean values for pixel distribution in band 2 (green) and species 16 (Faramea oblongifolia Standl.) with low mean values.

For the standardized data set, the environmental variables (aerial image, species, genus and family) were all again significantly correlated with the first, and all, canonical axes (RDA, Monte Carlo permutation test, P < 0.001 for species and genera, and P = 0.46 for family groupings). The standardization procedure improved the explanatory power of species, genus and family level groupings. For species, all axes of the standardized data set now explained 40.5% of total variance (improving on the raw data set), with the first axis explaining 36.3%. For genera, the first axis explained 30.1% (with remaining axes explaining an additional 3.6%). For family, the first axis explained only 7.4% of variance (and remaining axes added 3.4%). The species biplot (Fig. 7) shows that the mean histogram value for band 2 (green) is still the best descriptor for the first axis. Standardization increased the amount of variance explained by environmental variables, with the second axis now explaining only 1.5% of remaining variance.

Figure 7.

RDA biplot for standardized data set showing image features (B1_mean – red band histogram mean; B2_mean – green band histogram mean; B3_mean – blue band histogram mean; B1_StdDe – Red band histogram standard deviation; B3_StdDe – Blue band histogram standard deviation; B1N1rich – Red band richness index; B3N1rich – Blue band richness index; B1N2div – N2 diversity index for red band; B3N2div – N2 diversity index for blue band; B3Shan – blue band Shannon index) with >19% first axis fit and nominal environmental variables (species groups and image number) outside correlation range −0.1 to 0.1. Once again, green pixel intensity mean (B2_MEAN) is the best explanatory variable for data set structure.

Species from groundwork and aerial imagery

Data for 1048 trees were collected from the fieldwork on the ground, with trees representing 73 species, 58 genera and 39 families. All samples were collected between altitudes of 1786 m and 1966 m. Species numbers in primary and secondary forest were 29 and 22, respectively. Effective species number (ESN = eH) computed using EstimateS (Version 8.2, R. K. Colwell, http://viceroy.eeb.uconn.edu/estimates) for primary forest was 17.5 and 11.8 for secondary, showing increased diversity for the primary site.

A total of 196 individual tree crowns (of minimum radius 2 m) were identified and sampled from aerial imagery. With individual crowns represented across multiple images, a total of 396 individual crown samples were used in the analysis. Crown samples in aerial imagery represented 41 species, 31 genera and 21 families. Species abundance differs between those identifiable in aerial imagery and species represented on the ground (Fig. 8). In the plots, the five most abundant species based on ground surveys, in descending order, are; Critoniopsis occidentalis (Cuatrec.) H. Rob., Saurauia tomentosa var. spruce (Sprague) Soejarto, Sapium stylare Müll. Arg., Alnus acuminata Kunth, and Miconia clathrantha Triana ex Cogn; and in the aerial images, Miconia clathrantha Triana ex Cogn., Critoniopsis occidentalis (Cuatrec.) H. Rob., Sapium stylare Müll. Arg., Miconia brevitheca Gleason and Cecropia gabrielis Cuatrec. Aerial imagery only captures four crowns of S. tomentosa although it is represented by 90 samples on the ground, and only one crown of A. acuminata, represented by 79 individuals on the ground. A total of 32 species, collected from the plots, are not included in the crown analysis as they were not detected in the aerial imagery.

Figure 8.

Species abundance on the ground (dark column) and in the canopy (light column) in order of decreasing abundance in the canopy.

The DBH of crown samples ranged from 10 to 111 cm, with an average of 35.5 cm. From indicator ratio values (number of individuals in primary forest + 1/number of individuals in secondary forest + 1), the presence of Clusia crenata Cuatrec. (8.7), Cornus peruviana J.F. Macbr (21.0) and Myrcia fallax Rich. (22.0) in the canopy is principally associated with primary forest habitats. Secondary forest lacked clear dominant canopy species, with Cercropia gabrielis Cuatrec. (1.8) providing some information, as it was almost twice as common in secondary forest than primary forest.

Predictive models

Using the validation data sets, the probability of correctly classifying a crown using the four predictive models is shown in Table 1.

Table 1. The probability of correct species or species group classification of a crown for each predictive model. Classification success is shown for all crowns in the raw and standardized data set and for raw and standardized data in a data set where each species is represented by a minimum of nine crowns . Classification success in bold signifies most successful decision tree building method for each category
GroupingDecision TREE building methodClassification success % (number of categories)
All crowns: raw dataAll crowns: sandardized dataGroups with 9+ crown replicatesGroups with 9+ crown replicates: standardized
SpeciesCHIAD33.1 (41)34.3 (41)44.8 (12)53.5 (12)
Exhaustive CHIAD45.5 (41)39.1 (41)55.2 (12)53.5 (12)
CRT 47.1 (41) 47.0 (41) 61.2 (12) 59.8 (12)
QUEST46.7 (41)34.0 (41)44.2 (12)39.9 (12)
Mean of all models43.138.651.351.7
GenusCHIAD34.8 (31)39.6 (31)49.5 (11)58.2 (11)
Exhaustive CHIAD 51.5 (31) 44.2 (31)62.5 (11) 64.2 (11)
CRT48.7 (31) 49.5 (31) 65.0 11) 62.0 (11)
QUEST36.6 (31)33.8 (31)45.4 (11)49.7 (11)
Mean of all models42.941.854.352.5
FamilyCHIAD41.2 (23)40.9 (31)48.6 (9)57.6 (9)
Exhaustive CHIAD49.7 (23)48.5 (31)57.1 (9)53.8 (9)
CRT 52.8 (23) 51.8 (31) 66.2 (9) 62.0 (9)
QUEST38.9 (23)38.1 (31)46.4 (9)48.4 (9)
Mean of all models45.644.855.255.4

Following validation, the highest classification success was nearly always associated with predictive models generated with the CRT model, although the improvement over other models was small (Table 1). Predictive success varied from ca. 50% for the complete data set to just over 60% for a data set that only included species with nine representative crowns or more. There was little actual improvement in prediction between the data sets, as the random probability of correct species allocation in the data set with a minimum of nine crowns per species increased to 8.3% (from 2.4% in the complete data set) as species groups reduced from 41 to 12.

There was no significant improvement in classification success through higher taxonomic grouping from species to genus to family level. In fact, there was an underlying reduction in success as, again, the number of categories decreases from 41 species to 31 genera and 23 families – increasing the random probability of correct assignment from species to family.

Predictive success at species level

All crowns of eight species, consisting of individuals that represented at least 5% of the canopy crown, were correctly assigned to their species categories at rates higher than by chance for both the raw and standardized data sets (Fig. 9a). Using our whole image standardization approach, the aerial data set only improved species prediction for half of the species. From the raw data set, crown samples of Cecropia gabrielis Cuatrec could be identified with the greatest certainty. Cornus peruviana J.F. Macbr, Miconia brevitheca Gleason and Miconia clathrantha Triana ex Cogn could also be identified with high probability (over 70%).

Figure 9.

Ratio of correctly predicted crown species during validation to random probability of assignment to correct species. Dark histogram bar represents the raw data set and light bar the standardized data set for species (a), genus (b) and family (c) for individual groupings representing a minimum of 5% of crowns in the data set.

Predictive success at genus level

All crowns in the eight genera representing at least 5% of the canopy could be predicted at rates higher than random (Fig. 9b). As with species data, standardization did not improve predictive success of the model for all genera, with only four of the eight genera showing improved rates of prediction. For the standardized data set, crowns from the genus Miconia were most successfully identified (81%), followed by Cecropia for the raw data set, then Critoniopsis in the standardized data set.

Predictive success at family level

With the exception of Clusiaceae, in the standardized data set all crowns could be identified to family at rates higher than random for families representing over 5% of crowns (Fig. 9c). Standardization of imagery only improved classification for two families, Asteraceae and Euphorbiaceae, with all other families showing no change or a reduction in correct classification. Euphorbiaceae and Lauraceae were predicted with over 80% success, with Melastomataceae and Urticaceae predicted with 74% and 75% success, respectively.

Discussion

The collection and use of proximal remote sensing imagery is particularly relevant in areas where extensive and regular cloud cover makes standard satellite imagery redundant and for localized ecological assessments. However, the collection of very high-resolution imagery is fraught with practical challenges, particularly when applying low-budget technology, as applied here, in topographically varied terrain such as the Ecuadorian Andes. A major challenge is to reduce or standardize image quality during the acquisition process, as many factors can contribute to variance in image characteristics. For example, the angle of incident light, or sun height angle, has been shown to play a significant role in influencing image quality (Kleman 1987; Barbier et al. 2011), and in our study the only control over this was to ensure images were collected at the same time each day on forests of the same aspect, to attempt to achieve similar light conditions (although with further study imagery taken at varying sun angles might provide a innovative means for defining crown shape through changes in shadow form). Similarly, no account was taken of the vertical sensor angle of the camera platform at the moment of image capture, although small changes in angle have been modelled and shown to contribute less to image variance than sun angle (Barbier et al. 2011). The extreme topography of our study site can also amplify variance in image parameters through image distortion, particularly when attempting to generate composite georeferenced images. We attempted to minimize the propagation of these errors by analysing individual images rather than working from a composite data set (Fig. 3). These effects could be evaluated, and possibly standardized, in further studies by assessing the outputs of a particular group of crowns in a number of different images, taken at different times in the day and from varying camera angles. Another factor influencing collected imagery is vibration from the helicopter camera platform, which at fine levels is hard to quantify.

Our crown analysis methodology captured approximately 16% of the trees identified from ground-based survey (>10 cm DBH). This is slightly less than the quarter observed by Trichon & Julien (2006) and 28% by Herwitz et al. (1998), due in part to the protocol adopted, where crowns of <4-m diameter were excluded from our analysis and possibly also due to greater crown overlap on steep Andean forest slopes. Aerial data sets represented 54% of families, 53% of genera and 56% of taxonomic groups sampled from the ground, a value slightly higher than the 45% seen in imagery collected by Trichon & Julien (2006). The limitation of aerial imagery is that it only captures the larger trees, with many sub-canopy species unrepresented in sampling. For these Andean montane forests, some 13% of canopy samples were associated with a DBH of <20 cm. In work from the tropical lowlands, canopy species were only represented by trees with DBH > 20 cm (Trichon 2001; Trichon & Julien 2006). This simply reflects the structural difference of forest types at higher altitudes, with lower overall canopy height than lowland rain forest (Gentry 1988).

Ordination, based on the range of pixel descriptors used, reveals inherent structure between crowns and shows that existing species-, genus- and family-level groupings correlate significantly to this pattern. We found that it was the mean value of the pixel histogram, for the green spectral band, that provided the best explanation of species groupings. It is well known that terrestrial plants all tend to reflect at high levels in the green band of visible light. This has often been considered a suboptimal feature of inherited photosynthetic apparatus due to evolutionary ‘lock-in’ from the lineage of green algae, although there is also evidence that plants often experience too much light in this frequency (or are limited by other resources), with no selective advantage to absorbing more green light (Kiang et al. 2007). The higher reflectance of leaves in this band compared to the red and the blue wavebands provides information on both leaf density and species-specific reflectance (Castro-Esau et al. 2006). At the crown scale, it is leaf density and three-dimensional arrangement, together with non-photosynthetic crown components, that influence the intensity and wavelength of reflected light (Clark et al. 2005). Differences in green band mean pixel intensity between crowns is also influenced by species-specific absorbance in green wavelengths. Studies using hyperspectral imagery at the leaf and crown scale have confirmed that the most effective visible wavebands for species classification lie in the visible blue and blue-green edge leading to the green peak (Clark et al. 2005; Castro-Esau et al. 2006). We suggest that it is this combination of information from leaf density, or leaf area index (LAI), and species-specific reflectance that enables the mean pixel intensity values in the green band to best predict taxonomic groupings. Although our data are limited to three broad visible spectra recorded by digital CCD cameras, i.e. RGB, it is clear that spectral information within the green waveband plays a dominant role in species group classification at the crown level. Spectral information is clearly important in classification, and with extra spectral information likely to improve our ability to classify imagery. For example, the use of sensors capable of adding the near-infrared part of the spectrum (red-edge effect) and shortwave infrared could improve the classification results markedly (Clark et al. 2005).

We observed that a value integrating pixel information across the crown, the mean of green pixel histogram, was more powerful than pixel diversity information. This lends support to the argument of Nagendra & Rocchini (2008) that hyperspectral information at resolutions below crown size might actually hinder classification by introducing greater variability. In reality, their conclusion probably reflects the fact that computers still lack the capacity to recognize complex shape patterns. This is where hyperspatial imagery can contribute to increased probability of species identification by adding a structural face to the analysis, such as textural properties of crowns (Trichon 2001; Gonzalez-Orozco et al. 2010). In this study, we tested simple textural descriptors such as the standard deviation (Meyer et al. 1996) and landscape diversity and evenness indices of crown pixel intensities, providing broad textural descriptors of data spread and evenness. These crown pixel pattern and textural descriptors failed to describe species groupings more effectively than pixel distribution means. At this stage, we should be careful in concluding that hyperspatial approaches have less potential than hyperspectral ones (Nagendra & Rocchini 2008). What is required is an effective suite of algorithms for texture and structure that mimic human ability to separate patterns in imagery, such as that in Fig. 5. Our future work will screen a range of textural and pattern descriptors to identify information-rich classifiers for improved crown-level species identification. In particular, we will focus on approaches that reintroduce two-dimensional information, such as multivariate image analysis (MIA) and wavelet texture analysis (WTA) (Bharati et al. 2004).

Weaknesses of a crown classification method such as ours, that combine elements of spectral information (species-specific green waveband reflectance) and hyperspatial pixel information (LAI), include physicochemical and temporal changes that can influence leaf size, density and chemistry. Within-species crown metrics are likely to increase in variance due to seasonal changes associated with phenological events, such as leaf turnover and flowering, differences in soil type and moisture availability, temperature and other stressors. Based on spectral signatures at the leaf level, Castro-Esau et al. (2006) suggest that the potential for classification of species across sites and seasons was low. More recent work taking a spectranomics approach that links leaf chemistry and spectral signatures is more optimistic. Despite clear substrate effects, Asner & Martin (2011) found that phylogeny remained the most important variable in defining leaf chemistry and ability to classify crowns to species using hyperspectral imagery. In contrast, information on how season and substrate impact crown pattern in hyperspatial imagery is totally lacking – principally due to a lack of temporal imagery.

Using our suite of object-based crown metrics, the predictive success of decision trees generated with the range of tree-growing methods tested was similar. The most successful predictive TREE model (CRT) generated a 47% probability of correct species identification for crowns in our aerial sample containing 41 species. This compares with 87% overall tree species identification by eye in the study of Trichon & Julien (2006) for 12 species in Amazonian forest of French Guiana, and 75% accuracy for 24 species in a study by Myers (1982) in the northern Queensland rain forests. Dichotomous keys developed and tested by Gonzalez-Orozco et al. (2010) for canopy species in the Ecuadorian Amazon showed a successful identification rate closer to ours, with a value of over 50% for five of the ten taxa studied, and over 70% success for three taxa. In our study, as in all other studies, the probability of correctly identifying a crown was species-specific, varying from 0% for a number of species (generally singletons or species represented by just a few samples) to over a 93% for Cecropia gabrielis Cuatrec. High rates of success were also seen for crown samples of Cornus peruviana J.F. Macbr. Both C. gabrielis and C. peruviana provide ecological information on successional status, so with high rates of identification success, we have seen that using relatively simple pixel descriptors can provide important ecological information on the successional status of Andean mountain forest within this altitudinal range. Grouping species into traditional genus and family levels did not result in an actual improvement in overall predictive success, as grouping species introduced greater variability. Again, crowns from some groups were preferentially predicted. Future work should investigate the potential of the range of alternative classification methods for multivariate crown metrics. Algorithms such as the random forest (or random forests) ensemble classifier need exploration and might improve crown classification success as they make greater use of crown information (Kampichler et al. 2010).

A successful method of remote identification of tree species within a tropical forest canopy would allow rapid landscape-scale analysis of a number of forest characteristics that have been studied at smaller scales for many years, for example seed dispersal and flowering patterns (Janzen 1970; Hubbell 1980), and the mapping of distributional and diversity patterns (Gentry 1992). An understanding of forest composition would, in turn, enable more accurate assumptions to be made regarding habitat suitability (Hyde et al. 2005) in terms of quality, keystone species, fragmentation and their respective effects on the biodiversity of the forests (Hill & Curran 2003; Leigh et al. 2004). The increased knowledge in these areas would enable more informed conservation planning and management (Margules & Pressey 2000; Pouliot et al. 2002). In particular, the emerging need to assess carbon stocks as part of Reduced Emissions from Deforestation and Degradation (REDD+) mechanisms would also benefit from research in this area, and applications are already under investigation using high-resolution satellite images (Gonzalez et al. 2010). A clear role for crown-level identification exists in monitoring degradation, for example, by identifying selective illegal logging of high-value or endangered timber species. The ability to survey the changing composition of large tracts of forest would enable long-term monitoring studies to investigate impacts and adaptation to climate change.

Our work has shown that relatively simple pixel distribution descriptors for the three visible (RGB) bands in VHR aerial imagery can provide information to predict traditional taxonomic species and species groupings of trees. The technique works particularly well for a few species that have distinctive leaf densities and mean crown intensities in the green band. The complexity and variability in crown patterns and spectra mean that it is likely that an optimized canopy crown species recognition system will need to combine information from a range of sources: pattern information from hyperspatial imagery, crown shape information from LIDAR and spectral information from hyperspectral sensors. To address this, future work should address the current research gap that exists in understanding and developing effective pattern and textural descriptors for species-level classification from crown imagery in very high-resolution aerial imagery.

Acknowledgements

We would like to thank all the Earthwatch volunteers who spent muddy hours collecting the botanical data, staff at the Santa Lucia cloud forest reserve for their never-ending hospitality and Santiago Martínez, our remote control helicopter pilot. We thank the Earthwatch Institute and the Holly Hill Trust who funded all project work and the Ecuadorian Ministry of Environment for collection permits. We would also like to thank our anonymous reviewers whose thoughtful input has greatly improved this manuscript.

Ancillary