Extracting urban vegetation characteristics using spectral mixture analysis and
decision tree classifications
Thoreau Rory Tooke a,
⁎, Nicholas C. Coops a
, Nicholas R. Goodwin a
, James A. Voogt b
a
Department of Forest Resource Management, 2424 Main Mall. University of British Columbia, Vancouver, Canada V6T 1Z4
b
Department of Geography, 1151 Richmond St., University of Western Ontario, London, Canada N6A 5C2
a b s t r a c ta r t i c l e i n f o
Article history:
Received 11 March 2008
Received in revised form 8 October 2008
Accepted 11 October 2008
Keywords:
Urban
Vegetation
High spatial resolution
Quickbird
Spectral mixture analysis
Decision tree classification
Multispectral data
Urban vegetation cover is a critical component in urban systems modeling and recent advances in remote
sensing technologies can provide detailed estimates of vegetation characteristics. In the present study we
classify urban vegetation characteristics, including species and condition, using an approach based on
spectral unmixing and statistically developed decision trees. This technique involves modeling the location
and separability of vegetation characteristics within the spectral mixing space derived from high spatial
resolution Quickbird imagery for the City of Vancouver, Canada. Abundance images, field based land cover
observations and shadow estimates derived from a LiDAR (Light Detection and Ranging) surface model are
applied to develop decision tree classifications to extract several urban vegetation characteristics. Our results
indicate that along the vegetation-dark mixing line, tree and vegetated ground cover classes can be
accurately separated (80% and 94% of variance explained respectively) and more detailed vegetation
characteristics including manicured and mixed grasses and deciduous and evergreen trees can be extracted
as second order hierarchical categories with variance explained ranging between 67% and 100%. Our results
also suggest that the leaf-off condition of deciduous trees produce pixels with higher dark fractions resulting
from branches and soils dominating the reflectance values. This research has important implications for
understanding fine scale biophysical and social processes within urban environments.
© 2008 Elsevier Inc. All rights reserved.
1. Introduction
As our understanding of urban systems has evolved, researchers
have become increasingly aware of the importance of detailed land
surface characteristics to many processes established in social and
physical geographic sciences. These land cover features include both
natural and anthropogenic attributes and are characterized as being in
a state of constant change due to the pervasive influence of human
activity (Ben Dor, 2006). Urban meteorology and hydrology provide
examples of disciplines which apply spatial land cover information to
explain biophysical processes. Specifically, the impervious surfaces of
urban areas represent an essential component of macro-scale models
representing established phenomena including urban heat island
effects (Oke, 1982). Spatial variation in pervious and impervious
surface composition has since been demonstrated to affect surface
thermal and moisture conditions; attributes that are key determinants
of urban climate (Grimmond et al., 1996; Voogt & Oke, 1997).
Understanding the more complex relationships among land
surface characteristics and urban climates requires that meteorologists
incorporate a wide range of features beyond the basic division
between impervious and pervious surfaces. Detailed land cover
characteristics including surface albedo, shade, and vegetation
condition inform meteorological studies at local (e.g. 102
to
5×104
m) and micro- (e.g. 10−2
to 103
m) scales (Sawaya et al., 2003;
Mueller & Day, 2005). Vegetation is of particular interest as it presents
a versatile resource for effectively managing and moderating a variety
of problems associated with urbanization. The spatial distribution and
abundance of urban vegetation, for example, is recognized as a key
factor influencing numerous biophysical processes of the urban
environment, including air and water quality, temperature, moisture,
and precipitation regimes (Avissar, 1996; Grimmond et al., 1996;
Nowak & Dwyer, 2000). Detailed vegetation characteristics, such as
the structure of plant canopies and their physiological condition also
exert a strong influence on more complex processes such as urban
wind flow and rates of transpiration (Avissar, 1996; Wang et al., 2008).
In addition, vegetated areas such as gardens, parks, and forests have
been related to positive social outcomes including reductions in crime
(Kuo & Sullivan, 2001), health benefits (Coen & Ross, 2006), and
advanced childhood development (Taylor et al., 1998). Given the
associations between vegetated land cover and the biophysical and
social processes of urban systems there exists an ongoing demand for
effective urban vegetation mapping and classification techniques.
Remote Sensing of Environment 113 (2009) 398–407
⁎ Corresponding author. Tel.: +1 604 822 4148; fax: +1 604 822 8645.
E-mail address: thoreau@interchange.ubc.ca (T.R. Tooke).
0034-4257/$ – see front matter © 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2008.10.005
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
Mapping detailed land cover attributes within urban environments
has been primarily reliant on conventional cadastral information
from municipal agencies. However, the high cost and time
consuming nature of interpreting this data, as well as difficulties in
accessing data, can restrict the capacity for quantitative studies of
vegetation impacts on biophysical and social processes in urban areas.
In addition, cadastral information is often limited to areas of public
access, resulting in large gaps of detailed land cover information
across cities. In contrast, remote sensing imagery can provide
information that is well suited to extensive mapping of vegetated
surfaces and recent developments in high spatial resolution sensors
(e.g. b5 m) such as IKONOS and Quickbird have further enabled
detailed analysis of urban areas. Herold et al. (2004) suggest that the
visible region of the electromagnetic spectrum provides the most
prominent spectral information required for separating urban land
cover materials. As a result, high resolution broadband sensors with
multiple channels positioned in this region of the spectrum can begin
to resolve some of the detailed land cover components necessary for
informing current microclimate (Noilhan & Mahfouf, 1996; Voogt &
Oke, 1997) and ecological models (Zipperer et al., 1997).
Critical for the interpretation of high spatial resolution remote
sensing imagery in urban environments is the development of accurate
remote sensing classification techniques. Traditional supervised or
unsupervised classifications assign each pixel to a single class and as a
result, these classifications can significantly underestimate or overestimate
land cover types in urban environments as pixels often contain
a mixture of cover types. For example, research by Thomas et al. (2003)
compared high resolution urban mapping methods and found that
traditional supervised and unsupervised spectral classification methods
resulted in map accuracies of around 50%. Urban environments alsotend
to contain fine scale heterogeneous land covers with narrow linear
patterns (Zipperer et al., 1997; Collinge, 1998) that are not always
captured within a single image pixel. Due to the inability of traditional
classification algorithms to account for mixed pixels, techniques better
suited to heterogeneous environments have been developed. Spectral
mixture analysis (SMA), in particular, has been used to classify urban
vegetation cover (Small, 2001; Small & Lu, 2006). This approach divides
pixels into representative fractions of land cover that combine at the
instantaneous field of view (IFOV) of the sensor.
In the past decade SMA has developed as the primary method for
extracting multiple urban land covers from a single pixel value
(Kresller & Steinnocher, 1996; Small, 2001; Rashed et al., 2001). Early
urban land cover classification has been theorized according to Ridd's
(1995) V–I–S (vegetation–impervious surface–soil) classification
scheme. This scheme provides a conceptual model that divides
urban environments into three classes: vegetation, impervious
surface, and soil. This approach remains problematic in a remote
sensing context as it represents features that cannot necessarily be
distinguished on the basis of reflectance values alone (Phinn et al.,
2002, Powell et al., 2007). As a result, Small (2001) developed a more
applicable model that establishes substrate, vegetation, and dark
(SVD) features of the urban environment as components for SMA.
These pure endmembers represent features at the apexes of the urban
mixing space, yet it remains unclear whether more detailed vegetation
characteristics including trees and vegetated ground cover can
also be quantified in terms of their separability along the mixing line
between the dark and vegetation endmembers (Small & Lu, 2006).
Although higher order vegetation details including species and
condition do not produce distinguishable pure pixels in three
endmember mixture models, they represent physically and structurally
distinct land cover features whose extraction at high spatial
resolutions can inform micro- and local scale urban process models
and consequently represents the central focus of the following
research.
The objective of this study is to develop a technique to extract
vegetation species and condition information using sub-pixel abundance
values from high spatial resolution multispectral imagery. We
produce fractions of vegetation, high albedo substrate, and dark
features by applying spectral mixture analysis to a Quickbird image
over the city of Vancouver, Canada. Shadow estimates from a LiDAR
(Light Detection and Ranging) hillshade model in addition to field
based observations of vegetation condition and species were collected
Fig. 1. Study area over Vancouver, British Columbia showing extent of Quickbird imagery and LiDAR dataset.
399T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
and provided training data for decision tree classifications. These
parameters are used in conjunction with the SMA derived fractions of
vegetation, high albedo and dark features to quantify the separability
of various vegetation elements within the urban environment.
Discussion of the results focuses on issues which may impede our
procedure and considerations regarding the application of this
technique for modeling various fine scale urban biophysical and social
processes.
Fig. 3. Comparison between Quickbird multispectral image and LiDAR (Light Detection and Ranging) derived surface model depicting shadow for a,b) the central business district and
b,c) a residential neighbourhood.
Fig. 2. Field plot classification schematic representing increasing levels of vegetation detail.
400 T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
2. Methods
2.1. Study area
The City of Vancouver (49° 15′N, 123° 6′W) on the mainland
western coast of Canada is located within the larger urban region of
metropolitan Vancouver and covers a 114 km2
area. Vegetation
including various evergreen needleleaf and deciduous broadleaf tree
species, shrubs, and grasses comprises a large portion of the city's
surface area and due to the temperate climate of the region much of
the vegetation remains green for a majority of the year (Straley, 1992).
Areas of manicured grass exist throughout the city on private lots and
parks, while wild native grasses are less prevalent and tend to be
found in designated protected areas. Trees are also abundant
throughout Vancouver with native evergreen needleleaf species
dominant in urban parks and deciduous broadleaf species dominant
along streets and in residential areas.
2.2. Remotely sensed data
A Quickbird multispectral image was acquired on March 29th 2007
over the study area, capturing a wide range of land cover types
including residential, commercial, industrial, and forest (Fig. 1). The
image has a spatial resolution of 2.4 m with four spectral bands (blue,
450–520 nm; green, 520–600 nm; red, 630–690 nm; and nearinfrared,
760–900 nm) and a 0.6 m panchromatic band. The
Quickbird multispectral image was initially calibrated to at-sensor
radiance and atmospherically corrected to estimate surface reflectance
using a dark-object subtraction technique (Chavez, 1988).
Digital orthorectified aerial photographs acquired in 2004 with a
spatial resolution of 0.2 m were also available to provide additional
land cover details.
Airborne LiDAR data was acquired in March 2007 by Terra Remote
Sensing (Sidney, British Columbia, Canada) using a TRSI Mark II
discrete return sensor attached to a fixed wing platform. The sensor
was configured to record first and last returns with a pulse repetition
frequency of 50 kHz, platform altitude of 800 m, maximum off-nadir
view angle of 15°, wavelength of 1064 nm, and a fixed beam
divergence angle of 0.5 mrad. The average pulse spacing equalled
one laser pulse return per 0.7 m2
. Ground and non-ground returns
were classified using TerraScan software (Terrasolid, Finland). The
area surveyed includes a 1 km wide and 9 km long transect from
Stanley Park, through downtown Vancouver which contains a large
number of high rise buildings, to a research observation tower located
in a residential area (Fig. 1).
2.3. Field data
In our analysis detailed vegetation characteristics including
condition and species were recorded to provide the necessary training
Fig. 4. Location of the field plots and shadow plots within the spatial extent of the Quickbird imagery.
401T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
data for decision tree classifications of urban vegetation. Fig. 2 displays
a list of the vegetation characteristics recorded in the field to
investigate the location of these features within the mixing space of
high resolution satellite imagery particular to the City of Vancouver.
Vegetation is separated into tree and vegetated ground cover
categories, and then each category is further classified based on the
Fig. 5. True colour a) multispectral Quickbird image and b) aerial photograph compared with image fractions for the three endmembers: c) vegetation, d) high albedo substrate, and e)
dark.
402 T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
species and condition of the vegetation. The vegetated ground cover
class is separated into wild (unmanicured and long grasses),
herbaceous plants, dry grasses (senesced and dead), and manicured
grasses. The tree class is separated according to whether the
vegetation is a deciduous broadleaf or evergreen needleleaf species.
2.4. Field plots
Field locations were selected to capture a wide range of vegetation
patterns, condition and species. Due the significant areas of private
land uses in the urban setting sites were chosen in areas with minimal
limitations to access and as a result most of the field plots were located
in public spaces. Accessibility was evaluated using orthophotographs,
street maps, and a priori knowledge of the region. Navigation to the
field plots was done using a differential GPS unit applying wide area
augmentation system (WAAS) technology and orthophotographs. The
extreme heterogeneity of urban environments compared to natural
landscapes provides new challenges when attempting to validate the
results of high resolution imagery to field based observations. Because
numerous contrasting surfaces are being analyzed according to their
influence on the spectral response of a pixel or group of pixels, it is
critical that the field plots representing those portions of the image
align with sub-pixel accuracy. To address this issue imagery was
referenced using spatial subsets from the orthophotographs and
rectified to a pansharpened Quickbird image with RMS errors less
than 1 m.
2.5. Shaded area estimation using LiDAR data
To identify image pixels representing shadow a hillshade model
was developed using a LiDAR transect overlapping the center of the
Quickbird scene. First return heights were gridded with 1×1 m pixels
to represent the maximum height/urban vertical profile across the
study area. Using collection date specifications from the Quickbird
imagery and a hillshade model the incoming solar radiation was
estimated. Input parameters consisted of a sun zenith angle of 43.5
and azimuth orientation of 167.1°. The resulting model and multispectral
Quickbird image are shown in Fig. 3.
2.6. Linear spectral mixture analysis
SMA divides each pixel of an image into the representative fraction
of selected endmember spectra. Each endmember consists of spectra
that represent materials on the ground (Adams & Gillespie, 2006).
Linear mixing assumes that the spectral reflectance profile of each
pixel is a linear combination of the selected endmembers (Goodwin et
al., 2005). To find the best combination of endmembers to explain the
mixed reflectance signal of a pixel, matrix inversion is performed by
Eq. (1).
Ri = ∑
n
j = 1
fjREij + ei and 0V ∑
n
j = 1
fj V 1 ð1Þ
where Ri is the total pixel reflectance; fj, the endmember image
fraction; REij, the reflectance of image endmember j at band i; n, the
number of endmembers; and εi is the residual error for band i. The
number of possible endmembers equals the number of bands minus 1.
Our spectral mixture analysis procedure is performed following
the methods presented by Small and Lu (2006). The first step involves
a principle components (PC) analysis on the four bands of the
Quickbird image. Performing PC transformations on broadband
imagery enables the topology of the mixing space to be constructed
as a three dimensional model which encompasses all the combinations
of theoretically pure physical elements within each pixel of the
scene (Small & Lu, 2006). These pure features, referred to as
endmembers, compose the apexes of the three dimensional mixing
space (modeled as a convex hull) and must be carefully selected to
ensure that accurate mixture models are produced. To maintain the
integrity of the analysis pixels are manually selected from the apexes
of the mixing space and verified against georeferenced orthophotos.
Similar to previous research (Rashed et al., 2001; Small, 2001; Small &
Lu, 2006) the endmembers were defined as vegetation, high albedo
substrate, and dark features. Dark endmember pixels include shadows
typically cast by tall buildings in the central business district of the
city, in addition to significant areas of shadowed forest canopies in
evergreen needleleaf forests. Vegetation endmember pixels are
characterized as highly manicured grasses typically located in golf
courses and public parks. The final step in the analysis uses Eq. (1) to
perform SMA and generate three abundance images representing each
of the individual endmembers. To produce abundance images with
meaningful values that can be coupled with field observations sum to
unity and positivity constraints are applied to the analysis and
fractions represented as a percentage.
2.7. Decision tree classification
Decision trees (DTs) have emerged recently as an alternative land
cover classification method and may provide improved accuracies
over maximum likelihood and neural network classifications when
applied to multispectral imagery (Mahesh & Mather, 2003). DTs offer
advantages over these other types of classification methods in that
they can process data measured at different scales and no assumptions
are made concerning the frequency distributions of the data. In
addition DTs are relatively quick; requiring minimal computational
time compared to neural networks (Mahesh & Mather, 2003). The
basic process of DT construction involves the repeated division of a set
of training data into increasingly distinct subsets based on tests to one
or more of the feature values. Once a set of hierarchically structured
rules, or branches, are produced based on the provided training data,
then these rules can be applied to an entire image in order to produce
accurate land cover maps and inventories for further spatial analysis.
In our analysis, the training data are derived from the field based
observations of vegetation species and condition in addition to the
LiDAR derived shadow plots (Fig. 4). Single trees for each hierarchical
level of vegetation class are developed in DTREG using a 10 V-fold
cross validation technique which has been demonstrated to produce
highly accurate results without requiring an independent dataset for
assessing the accuracy of the model (Sherrod, 2008). A more detailed
explanation of decision tree validation and pruning techniques can be
found in Sherrod (2008). The V-fold cross validation technique used in
this paper involves first developing an initial large tree using all the
available data, known as the reference tree. Secondly the total dataset
is partitioned into 10 groups (or folds) and 10 new subsets of the total
are created using 9 out of 10 of the folds. Ten test trees are then built
Fig. 6. Statistically developed decision tree classification for the extraction of broad level
vegetation characteristics including trees and vegetated ground cover.
403T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
using the reduced datasets with the unused 10% in each case then run
through each test tree and the classification error for that tree
computed. Once the 10 test trees have been built, their classification
error rate as a function of tree size is averaged and the reference tree is
pruned to the number of nodes matching the size that produces the
minimum cross validation cost (Breiman et al., 1984).
3. Results
The principal component analysis on the 4 band Quickbird image
showed over 99% of the image variance contained within the first
three primary principal components, which is in agreement with
earlier research (Small, 2003; Small & Lu, 2006). The resulting
distribution of pixel values within the mixing space produced
distinctive linear dispersions between the vegetation-dark and darkhigh
albedo apexes, while a concave dispersion was observed between
the vegetation-high albedo apexes indicating that few pure binary
mixtures of these features exist within the city (Small & Lu, 2006).
The abundance images produced from the SMA for the vegetation,
high albedo substrate and dark features of the imagery in addition to
the true colour multispectral Quickbird and orthophoto images is
shown in Fig. 5. The dark abundance image shows the dominance of
dark features throughout the scene resulting from significant
shadowing which comprises this endmember. Comparing the areas
of vegetation in the orthophotograph (Fig. 5) with the corresponding
dark and vegetation abundance images also indicates that certain
forms of vegetation are underestimated in a three endmember mixing
model, specifically trees which have moderate fractions of both dark
and vegetation features. High fractions displayed in the vegetation
abundance image are localized in grassy areas, while surfaces
including roads and parking lots are highlighted in the representative
high albedo substrate image.
Fig. 7. Mapped decision tree results comparing a) the true colour multispectral Quickbird image with extracted b) grass and vegetated ground cover classes, c) manicured and mixed
grass classes, and d) deciduous and evergreen tree species classes.
404 T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
To determine the location and separability of various vegetation
features within the image mixing space, field observations were
collected with classes representing hierarchical levels of vegetation
detail. The first order represents vegetated ground cover and trees and
the second order represents species and condition attributes related to
the vegetation. Data collected for each hierarchical level were used as
input into decision tree classification models and determined which
categories or combination of categories could be separated and
successfully classified within the decision tree models. A total of three
decision trees were successfully produced, the first representing broad
divisionsofvegetation(vegetated ground coverand tree)and the second
two representing more detailed vegetation classes as subsets of the
broader categories (manicured, mixed, evergreen, and deciduous).
The first model representing the first level of vegetation detail
explained 94% of the variance within the vegetated ground cover class
and 80% of the variance within the tree class. The primary factor for
dividing these two classes is the dark abundance image (Fig. 6). The
first split involved separating the dark values at 52%. Pixels with less
than 52% dark features were then classified into vegetated ground
cover pixels with a further rule applying high albedo abundance
values of greater than 2%. Dark values greater than 52% include both
shadow and tree classes and these image features were separated
using the vegetation abundance with a threshold of 41% where pixels
less than this value were classified as shadow and values greater than
41% were classified as tree. Results from this classification are mapped
in Fig. 7b.
Successful extraction of the first order vegetation classes enabled
the application of our decision tree technique for extracting the more
detailed second order classes. Pixels classified as vegetated ground
cover from the previous step were input into a new DT to extract the
classes of manicured, dry, herbaceous and wild. The best result
involved combining dry, wild, and herbaceous categories into a new
class labelled ‘mixed’, while the manicured class remained unchanged
from the original data. This classification used the dark abundance
values as the primary feature for classification (Fig. 8). The first rule
establishes a threshold of 40% dark and all pixels less than this value
were classified as manicured grass. An additional rule was required to
separate the remaining classes and used the high albedo abundance
value of 2% to derived mixed grasses (less than 2%) and the remaining
manicured grass class (greater than 2%). Variance explained for this DT
classification was 100% for the manicured class and 73% for the mixed
grass class (mapped in Fig. 7c).
The final classification extracted evergreen and deciduous classes
from the broader tree category. This classification used the dark
abundance values as the first feature for classification (Fig. 9). The only
branch in this DT establishes a threshold of 57% dark with tree pixels
less than this value classified as evergreen and pixel greater than 57%
dark classified as deciduous. Variance explained for this DT classification
was 80% for the evergreen class and 67% for the deciduous class
(mapped in Fig. 7d).
4. Discussion
4.1. Vegetation separability
Mapping the location and spatial extent of trees, vegetated ground
cover, and high level vegetation detail provides a valuable addition to
urban land cover mapping using high spatial resolution imagery.
Image classification techniques developed to date extract a basic
vegetation class which encompasses a broad range of features whose
structural and spectral diversity have a variety of impacts on urban
processes (Mueller & Day, 2005; Voogt & Oke, 1997). Spectral mixture
analysis provides a successful technique for extracting the fractional
abundance of general land cover features well suited to the
heterogenic composition of urban environments. Small and Lu
(2006) explain that high spatial resolution image vegetation fractions
provide more informative vegetation estimates than moderate
resolution imagery due to the reduction of possible distinct mixtures
and add that Quickbird pixels can resolve many of the individual
components representing urban vegetation. Nonetheless, the dimensionality
of the imagery still produces a three endmember mixture
model encompassing various vegetation conditions and species.
Fractional abundance values (Fig. 5) from our analysis show that
vegetation is generally represented by manicured grasses and as a
result underestimates the spatial extent of trees within the mixing
space.
To improve the classification of vegetation features and explain the
endmember variability within high spatial resolution mixing space,
we applied a decision tree classification using field observations and
the SMA derived abundance images. Of significance, this enabled
accurate estimates of the separability and location of vegetation
features including tree species and vegetated ground cover condition
within the mixing space.
Several hierarchical orders of classes were established with
increasing levels of detail regarding vegetation condition and species.
The first order classification involved separating trees from vegetated
ground cover. As the results indicate, this procedure explains 94% of
the variance for the vegetated ground cover class and 80% of the
variance for the tree class. The spectral distinction between these two
vegetation categories is largely a result of structural differences.
Vegetated ground cover tends to be close to the ground with closelyspaced
foliage resulting in little shadowing, enabling a strong
reflectance of photosynthetically active vegetation back to the sensor.
Alternatively, the horizontal variation in vertical structure of trees
causes significant shadowing interspersed throughout the foliage.
These structural differences result in varying amounts of shadow
Fig. 8. Statistically developed decision tree classification for the extraction of detailed
second order vegetation characteristics related to vegetated ground cover condition
including: manicured and mixed.
Fig. 9. Statistically developed decision tree classification for the extraction of detailed
second order vegetation characteristics related to tree species including: deciduous
broadleaf and evergreen needleleaf.
405T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
captured within a Quickbird pixel which are extracted in the ‘dark’
branch of the DT model in Fig. 6. As a result of greater shadowing, trees
are separable from vegetated ground cover in the mixing space and
are localized closer towards the apex of the dark endmember along
the vegetation-dark mixing line.
After the successful extraction of the general vegetation classes of
vegetated ground cover and tree, the same technique was applied to
provide estimates of more detailed characteristics including species
and condition. Several important discussion points are raised as a
result of the analysis. Separating evergreen needleleaf and deciduous
broadleaf trees using our technique explained 80% and 67% of the
variance respectively. Deciduous trees have significant variations in
leaf characteristics in terms of shape, size, and pigment compared
with many evergreen species. In addition, the Quickbird image was
captured in spring when the spectral response of deciduous broadleaf
trees is strongly affected by the seasonal variability associated with
tree phenology (many deciduous species are beginning to bud in
Vancouver). We suggest that the leaf-off condition of deciduous trees
results in the counterintuitive decision tree shown in Fig. 8 where
pixels with higher dark fractions are classified as ‘deciduous’ resulting
from branches and soils dominating the reflectance values associated
with deciduous tree species.
Fig. 10 displays a model based on the DT classifications indicating
how various levels of vegetation features localize within the mixing
space. This model is image specific and, in this case, represents the
location of vegetation features during the spring season. Future work
may be undertaken to study the seasonal variation of feature location
and separability within the mixing space, and might also benefit from
quantifying the location of higher level mixing spaces within the
original three endmember model. Seasonal selection of high resolution
imagery is an important consideration which will vary the
location and separability of vegetation features. Selecting a spring
image for our analysis provided good separability of vegetated ground
cover and trees and associated condition and species details.
4.2. Applications
Mapping and modelling tree and vegetated ground cover
characteristics from high spatial resolution satellite imagery in
urban areas enables significant advancements in our understanding
of urban systems. Boundary-layer climates are significantly influenced
by the distribution, abundance, condition and characteristics
of urban vegetated surfaces at local and micro-scales (Voogt & Oke,
1997). Imagery representing vegetation dynamics across urban
areas can provide urban planners with vital information required to
mitigate heat island effects and reduce building energy requirements
associated with heating and cooling (Grimmond, 2007). At
the same time, vegetation characteristics across a city can inform
epidemiologists of the spatial distribution of health risks related to
urban air quality (Corburn, 2007). Separating vegetation into more
detailed classes also has strong potential to inform urban ecologists
of species occurrence and vulnerable plant, animal, and bird
habitats associated with urban vegetation cover and its spatial
pattern across the landscape (Zipperer et al., 1997).
Fig. 10. Model depicting where shadow and vegetation characteristics including trees and vegetated ground cover and associated condition and species locate within the three
endmember mixing space of a high spatial resolution spring image over Vancouver.
406 T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407
5. Conclusion
Refining our understanding of urban systems through accurate
vegetation mapping is critical to the wellbeing of urban residents and
the sustainability of our cities. This paper examined the abundance of
urban endmembers in a Quickbird image over the City of Vancouver
and applied decision tree classifications to quantify and separate
various orders of vegetation detail. Results demonstrate successful
extraction of trees and vegetated ground cover using our technique.
This technique also proved successful in extracting vegetation species
and condition including evergreen needleleaf, deciduous broadleaf,
manicured and mixed grass, and highlights the phenological impacts
of more detailed vegetation features on the separability of species and
condition classes within the mixing space. Our analysis provides an
operational technique to enable vegetation extraction and related
studies across a variety or urban areas and will help parameterize
models related to urban biophysical processes. Further research will
involve the collection of summer imagery and a comparison of the
accuracy at which the vegetation classes can be extracted compared to
this current analysis.
Acknowledgments
This research was supported by the Environmental Prediction in
Canadian Cities (EPiCC) project funded by the Canadian Foundation for
Climate and Atmospheric Sciences (CFCAS). The authors would like to
thank field assistants Brad Lehrbass, Colin Ferster, Sam Coggins, and
Trevor Jones as well as Sue Grimmond, Jinfei Wang, and Timothy Oke
for their insight into remote sensing considerations for urban
micrometeorological applications. We would also like to thank
Christopher Small and two anonymous reviewers for their helpful
critiques and comments.
References
Adams, J. B., & Gillespie, A. R. (2006). Remote sensing of landscapes with spectral images: A
physical modeling approach. New York: Columbia University Press.
Avissar, R. (1996). Potential effects of vegetation on the urban thermal environment.
Atmospheric Environment, 30(3), 437−448.
Ben Dor, E. (2006). Imaging spectrometry for urban applications. In F. D. van Meer, &
S. M. de Jong (Eds.), Imaging Spectrometry (pp. 243−281). Dordrecht: The
Netherlands: Springer.
Breiman, L., Friedman, J., Olshen, R., & Stone, C. (1984). Classification and Regression
Trees. Belmont, CA: Wadsworth International Group.
Chavez, P. S., Jr. (1988). An improved dark-object subtraction technique for atmospheric
scattering correction of multispectral data. Remote Sensing of Environment, 24,
459−479.
Coen, S. E., & Ross, N. A. (2006). Exploring the material basis for health: Characteristics
of parks in Montreal neighbourhoods with contrasting health outcomes. Health and
Place, 12, 361−371.
Collinge, S. K. (1998). Spatial arrangement of habitat patches and corridors: Clues from
ecological field experiments. Landscape and Urban Planning, 42, 157−168.
Corburn, J. (2007). Urban land use, air toxics and public health: Assessing hazardous
exposures at the neighbourhood scale. Environmental Impact Assessment Review, 27
(2), 145−160.
Goodwin, N., Coops, N., & Stone, C. (2005). Assessing plantation canopy condition from
airborne imagery using spectral mixture analysis and fractional abundances. International
Journal of Applied Earth Observation and Geoinformation, 7, 11−28.
Grimmond, S. (2007). Urbanization and global environmental change: Local effects of
urban warming. Geographical Journal, 173, 83−88.
Grimmond, C. S. B., Souch, C., & Hubble, M. D. (1996). The influence of tree cover on
summertime energy balance fluxes, San Gabriel Valley, Los Angeles. Climate
Research, 6, 45−57.
Herold, M., Roberts, D. A., Gardner, M. E., & Dennison, P. E. (2004). Spectrometry for
urban area remote sensing — development and analysis of a spectral library from
350 to 2400 nm. Remote Sensing of Environment, 91, 304−319.
Kresller, F., & Steinnocher, K. (1996). Change detection in urban areas using satellite
images and spectral mixture analysis. International Archives of Photogrammetry and
Remote Sensing, 31, 379−383.
Kuo, F. E., & Sullivan, W. C. (2001). Environment and crime in the inner city. Environment
and Behavior, 33(3), 343−365.
Mahesh, P., & Mather, P. M. (2003). An assessment of the effectiveness of decision tree
methods for land cover classification. Remote Sensing of Environment, 86, 554−565.
Mueller, E. C., & Day, T. A. (2005). The effect of urban ground cover on microclimate,
growth andleaf gas exchange of oleander in Phoenix, Arizona. International Journal
of Biometeorology, 49(4), 244−255.
Noilhan, J., & Mahfouf, J. F. (1996). The ISBA land surface parameterisation scheme.
Global and Planetary Change, 13, 145−149.
Nowak, D. J., & Dwyer, J. F. (2000). The Urban Forest Effects (UFORE) Model: Quantifying
urban forest structure and functions. In M. Hansen, & T. Burk (Eds.), Proceedings:
Integrated tools for natural resources inventories in the 21st century. IUFRO Conference,
16– 20 August 1998, Boise, ID. General Technical Report NC-212, U.S. Department of
Agriculture, Forest Service, North Central Research Station, St. Paul, MN (pp. 714−720).
Oke, T. R. (1982). The energetic basis of the urban heat island. Quarterly Journal of the
Royal Meteorological Society, 108, 1−24.
Phinn, S., Stanford, M., Scarth, P., Murray, A. T., & Shyy, P. T. (2002). Monitoring the
composition of urban environments based on the vegetation–impervious surface–
soil (VIS) model by subpixel analysis techniques. International Journal of Remote
Sensing, 23(20), 4131−4153.
Powell, R. L., Roberts, D. A., Dennison, P. E., & Hess, L. L. (2007). Sub-pixel mapping of
urban land cover using multiple endmember spectral mixture analysis: Manaus,
Brazil. Remote Sensing of Environment, 106, 253−267.
Rashed, T., Weeks, J. R., Gadalla, M. S., & Hill, A. G. (2001). Revealing the anatomy of cities
through spectral mixture analysis of multispectral satellite imagery: A case study of
the greater Cairo region, Egypt. Geocarto International, 16(4), 7−17.
Ridd, M. K. (1995). Exploring a V–I–S (vegetation–impervious surface–soil) model for
urban ecosystem analysis through remote sensing: Comparative anatomy for cities.
International Journal of Remote Sensing, 16, 2165−2185.
Sawaya, K. E., Olmanson, L. G., Heinert, N. J., Brezonik, P. L., & Bauer, M. E. (2003).
Extending satellite remote sensing to local scales: Land and water resource
monitoring using high-resolution imagery. Remote Sensing of Environment, 88,
144−156.
Sherrod, P. H. (2008). DTREG Predictive Modeling Software. Users Manual. www.dtreg.
com/DTREG.pdf
Small, C. (2001). Estimation or urban vegetation abundance by spectral mixture
analysis. International Journal of Remote Sensing, 22(7), 1305−1334.
Small, C. (2003). High spatial resolution spectral mixture analysis of urban reflectance.
Remote Sensing of Environment, 88, 170−186.
Small, C., & Lu, J. W. T. (2006). Estimation and vicarious validation of urban vegetation
abundance by spectral mixture analysis. Remote Sensing of Environment, 100,
441−456.
Straley, G. B. (1992). Trees of Vancouver. Vancouver: UBC Press.
Taylor, A. F., Wiley, A., Kuo, F. E., & Sullivan, W. C. (1998). Green spaces as places to grow.
Environment and Behavior, 30(1), 3−28.
Thomas, N., Hendrix, C., & Congalton, R. G. (2003). A comparison of urban mapping
methods using high-resolution digital imagery. Photogrammetric Engineering and
Remote Sensing, 69(9), 963−972.
Voogt, J. A., & Oke, T. R. (1997). Complete urban surface temperatures. Journal of Applied
Meteorology, 36, 1117−1132.
Wang, J., Endreny, T. A., & Nowak, D. J. (2008). Mechanistic simulation of tree effects in
an urban water balance model. Journal of the American Water Resources Association,
44(1), 75−85.
Zipperer, W. C., Sisinni, S. M., & Pouyat, R. V. (1997). Urban tree cover: An ecological
perspective. Urban Ecosystems, 1, 229−246.
407T.R. Tooke et al. / Remote Sensing of Environment 113 (2009) 398–407