Sie sind auf Seite 1von 9

Proceedings of the 13th International Conference on Environmental Science and Technology

Athens, Greece, 5-7 September 2013

URBAN VEGETATION COVER EXTRACTION FROM HYPERSPECTRAL


REMOTE SENSING IMAGERY AND GIS-BASED SPATIAL ANALYSIS
TECHNIQUES: THE CASE OF ATHENS, GREECE

IRO A. GEORGOPOULOUa D. P. KALIVASa and GEORGE P. PETROPOULOSb


a
Agricultural University of Athens, Department of Soil Science and Agricultural
Chemistry, Iera Odos 75, Botanikos 11855, Athens, Greece.
b
Institute of Geography and Earth Sciences, University of Aberystwyth, King Street,
Aberystwyth, SY23 2DB, United Kingdom.
kalivas@aua.gr

EXTENDED ABSTRACT

Information on the urban vegetation cover spatial coverage is important in sustainable


urban planning and resourceful environmental management, whereas the same time it
plays a very important role in human-environment interactions. The present study aims at
evaluating the combined use of Hyperion hyperspectral imagery with the Support Vector
Machines (SVMs) and Spectral Angle Mapper (SAM) pixel-based classifiers for
discriminating different land-cover classes at a typical urban setting focusing particular in
urban vegetation cover. As a case study, the city of Athens Greece was used. Evaluation
of the derived land cover maps was performed on the basis of the error matrix statistics
which was assisted by co-orbital higher resolution imagery available and field visits
conducted in our study region. To ensure consistency and comparability of our results,
the same set of training and validation points were used. Our analyses showed that
SVMs outperformed SAM in terms of both overall classification and urban vegetation
cover mapping accuracy. In particular, an overall accuracy of 86.53% and Kappa 0.823
was reported for the SVMs’ results, whereas for SAM were 75.13% and 0.673
respectively. The SVMs’ ability to identify an optimal separating hyperplane for classes’
separation exemplified the algorithm’s ability to perform better, in comparison to SAM, at
least this appears to be the case in our study. However, both techniques were influenced
by the relatively coarse spatial resolution of the sensor, which resulted to misclassification
cases due to spectral mixing effects. Yet, the potential of hyperspectral remote sensing
for efficient and up-to-date derivation of the spatial representation of urban vegetation
presence was evidenced, providing supportive results to efforts currently ongoing globally
towards the development of accurate and robust techniques in mapping the
spatiotemporal distribution of urban vegetation cover and dynamics from space.

KEYWORDS: remote sensing, Geographical Information Systems, Hyperion, Support


Vector Machines, Spectral Angle Mapper, Athens, Greece.

1. INTRODUCTION

Urban vegetation in particular is one of the central infrastructural components of any


urban ecosystem and it plays significant role within cities. The physical impact of
vegetation in the urban environment is to affect the thermal environment, air quality and
noise levels. In fact, urban vegetation provides ecological, social, health and economic
benefits to a city’s inhabitants. Urban areas are continuously increasing today as more
than half of the Earth’s population now lives in urban areas (Martine, 2007) and the
estimated annual urban population growth rate of 1.78% is almost twice as fast as that of
the global population (Van de Voorde etal, 2008). An important consequence of this is

CEST2013_0209
evidently the change of land cover types from natural to anthropogenic impervious
surfaces consisting of roofs, roads, parking lots, driveways, and sidewalks (Xian et al,
2005). Thus, the identification of urban areas especially in order to monitor urban
vegetation, becomes an important issue in modern world.

Earth Observation (EO) technology provides today a promising avenue in mapping and
monitoring urban vegetation cover structures and of their changes on a local, regional
and global scale. Use of remote sensing, often combined with Geographical Information
Systems (GIS), has shown great potential in this direction thanks to its advantages over
traditional procedures in terms of cost effectiveness and timeliness in the availability of
information over larger areas. Urban areas in particular are characterised by a wide range
of spectrally complicated properties across the electromagnetic region range because of
the presence of numerous spectrally unique and ambiguous materials such as dark-
shingles and asphalt roads (Herold et al., 2003). Other factors that further complicate the
analysis of urban areas leading to high within-class spectral variability include the 3-
dimensional heterogeneity of urban areas and urban vegetation cover material aging
(Herold et al., 2003; 2004; Herold & Roberts, 2005). In addition, some buildings and
open spaces are covered by spectrally similar urban surface materials, which further
hamper a clear discrimination between them. The recent advances in remote sensing
radiometers technology have led to the launch of hyperspectral EO systems. These are
sophisticated sensors which are able to record reflected light from land surface objects,
ranging from visible to shortwave infrared parts of the electromagnetic spectrum,
acquiring a vast amount of spectral information (Xu et al., 2008). Use of hyperspectral
imagery has generally shown a promising potential in terms of different land surface
targets identification and land cover mapping, including urban vegetation cover, have
been proposed for this purpose (Walsh et al., 2008; Petropoulos et al., 2012a;b).

Different algorithms have been proposed and image classification is evidently perhaps
the most widely used approach in urban vegetation mapping (Thoreau et al., 2009;
Franke et al., 2009). Image classification includes among others the following groups of
approaches: pixel, sub-pixel and object based classification techniques (Lu & Weng,
2007). Unsupervised classifiers group pixels with similar spectral values into unique
clusters according to statistically predefined criteria and re-assign the generated spectral
clusters into information classes. On the other hand, supervised classifiers use samples
of given identity for each land cover class, known as ‘’training sites’’, to classify image
pixels of unknown identity. Supervised classifiers are also commonly divided into
parametric and non-parametric. Pixel-based classifiers typically develop a signature by
combining the spectra of all training set pixels for a given feature. Pixel-based
classification algorithm may be parametric or non-parametric. In the case of urban areas,
the assumption of normal spectral distribution is not valid due to the complexity of such
landscape (Lu & Weng, 2007). In the above mentioned category of parametric classifiers
the Maximum Likelihood algorithm is included (Paola & Schowengerdt, 1995). Non-
parametric classifiers are suitable for the incorporation of non-spectral data into
classification procedure. Among the most commonly used non parametric classifiers are
artificial neural networks, decision tree classifiers, support vector machines and expert
systems (Foody & Arora, 1997).

Yet, studies performing comparative analysis of the performance of different classification


approaches with satellite hyperspectral imagery such as that from Hyperion sensor for
urban vegetation mapping in particular, are evidently scarce in scientific literature.
Hyperion is the first spaceborne imaging spectrometer having the same orbital
characteristics as the LandSat ETM+ multispectral sensor, acquiring spectral information
in 242 spectral bands at the resolution of 30 meters. Hyperion, which is onboard the
Earth Observer-1 (EO-1) satellite platform was launched in 2000 under NASA’s New

CEST2013_0209
Millennium Program. The sensor has two spectrometers, one in the visible and near-
infrared (VNIR) (bands 8-57, region 427-925nm) and the other in the shortwave infrared
(SWIR) region (bands 77-224, region 912-2395nm). In particular, an investigation of the
potential use of Hyperion hyperspectral imagery combined with Support Vector Machines
(SVMs) (Vapnik, 1995) and Spectral Angle Mapper (SAM) for deriving urban vegetation
mapping, has so far been very limited, if not existent. The comparison of these two
supervised algorithms’ performance may demonstrate their potential use in urban
vegetation extraction applications. Vegetation is of particular interest as it presents a
versatile resource of effectively managing and moderating a variety of problems
associated with urbanization (Thoreau et al., 2009).

In this context, the present study aims to appraise the use of SVMs and SAM in deriving
information on the regional distribution of urban land cover types when combined
specifically with Hyperion imagery. As a case study, the city of Athens has been selected
due to its high heterogeneity in terms of urban vegetation features and structures as well
as the importance of this area to urban studies as evidenced by various previous studies
(e.g. Grimmond, 2010; Thermopolis ESA project).

2. STUDY SITE AND DATASETS

Our study area includes the wider area of Athens, capital of Greece, situated in the
prefecture of Attiki, extending from 23o39’ to 23o42’ Easting and from 38o4’ to 37o55’
Northing. The area represents an extensive urban area with very high level of
construction and high population density. It is mainly occupied by structures like buildings
and roads, although vegetation presence is low. The Hyperion imagery used in our study
was acquired on August 27th, 2009. The Hyperion imagery was obtained from a previous
study conducted in the area (see Petropoulos et al., 2012b), in which authors had
originally acquired the imagery from United States Geological Survey (USGS) archive as
a full long scene in geotiff image format and already radiometrically corrected,
geometrically resampled and registered to WGS84 coordinate system, with elevation
correction applied. Additionally, co-orbital Google Earth imagery was used in our study to
support different aspects of methodology implementation and results interpretation.

3. METHODS

3.1. Data pre-processing

Image classification was conducted to the Hyperion imagery by applying the Support
Vector Machine (SVM) and the Spectral Angle Mapper (SAM) classification approaches
in ENVI 4.7 software platform. No pre-processing was applied to our Hyperion imagery as
it was supplied already pre-processed (further details in Petropoulos et al., 2012b).
Briefly, at first the Hyperion imagery was converted into ENVI format files that contain
wavelength, full width half maximum and bad band information. Then the water
absorption bands were eliminated to minimize the influence of atmospheric scatter and of
water vapor absorption, caused by well mixed gasses. Subsequently, Minimum Noise
Fraction (MNF) transformation was performed on all Hyperion bands that had not been
masked out (136 in total) in ENVI as a linear transformation in order to separate the noise
from data and to minimize the influence of systematic sensor noise during image
analysis. Hyperion final dataset, after the implementation of an inverse MNF, consisted of
136 bands, 46 in the VNIR and 90 in the SWIR. After this step, the resulting image was
reduced to a subset of the studied region. These final 136 bands were used in the
present study to perform our classification using the selected classifiers.

CEST2013_0209
3.2. Training data selection

In this study, both SVMs and SAM classifiers were applied to the Hyperion imagery using
training data representative of the different land cover types included in our classification
key (Table 1). Selection of the training data was assisted by the co-orbital Google earth
imagery and selected field visits conducted. The training sites were selected from the
Hyperion imagery and were carefully delimited. As training sites were selected pixels
representative of the most homogeneous areas. In total, five classes were created:
asphalt and buildings which represented the impervious surface of study area and trees,
low vegetation and bare soil which represented the according pervious surface.
Approximately 200-300 pixels per class (in total 1007 pixels) were identified as training
data, representing the classes defined in the classification scheme. Subsequently, the
two algorithms were implemented, using the same training sites collected.

Table 3.1. The classification key used in the study area.

Class Name ID Class description


Asphalt 1 Surfaces covered principally with asphalt, roads
Buildings 2 Urban fabric, urban area
Bare soil 3 Open areas with no vegetation, rocks or previously burnt
Trees 4 Surfaces principally covered with trees
Low vegetation 5 Open areas with little or vegetation of low height

3.3. Hyperion classification

3.3.1. Support Vector Machines (SVM)

SVMs is a supervised non-parametric statistical learning technique. The subset of points


that lie on the margin (called support vectors) are the only ones that define the
hyperplane of maximum margin (Vapnik, 1995). The most important characteristic of
SVMs is its ability to generalize a limited amount and/or quality of training data. This is in
line with the ‘’support vector’’ concept that relies only on a few data points to define the
classifier’s hyperplane. SVM can yield comparable accuracy using a much smaller
training sample size, compared to other alternative methods. Furthermore, as non
parametric, SVMs do not assume a known statistical distribution of the data to be
classified. The binary classification scheme in SVMs can be extended to a larger number
of classes N (where N>2). SVMs was implemented using the radial basis function (RBF)
kernel for performing the pair-wise classification, as its use has generally shown
satisfactory results (e.g. Petropoulos et al., 2012), while it requires defining a small
number of parameters and it produces generally good results in most classification cases.
RBF kernel performs the pair-wise classification (one against one approach) which is a
technique that N(N-1)/2 SVMs are produced following a binary tree-like fashion. The input
parameters required for running SVMs in ENVI software include the gamma (γ) in the
kernel function, the penalty parameter, the number of pyramid levels to use and the
classification probability threshold value. Regarding the parameterization RBF kernel
function, the γ parameter was set to a value equal to the inverse of the number of the
spectral bands of Hyperion imagery (0.007) and the penalty parameter was set to its
maximum value (100) focusing at no misclassification during the training process. The
pyramid parameter was set to a value of zero forcing the Hyperion imagery to be
processed at full resolution, whereas a classification probability threshold of zero was
used meaning that all pixels had to be classified into one class.

CEST2013_0209
3.3.2. Spectral Angle Mapper (SAM)

SAM is a supervised pixel based classification method that permits rapid mapping by
calculating the spectral similarity between the image spectrums to reference reflectance
spectra (Kruse et al., 1993). SAM measures the spectral similarity by calculating the
angle between the two spectra, treating them as vectors in n-dimensional space (Rowan
& Mars, 2003). Small angles between the two spectrums indicate high similarity and high
angles indicate low similarity. This method is not affected by solar illumination factors,
because the angle between the two vectors is independent of the vectors length. In the
present study, the maximum angle (radians) selected was 0.3. The input parameters
required for evaluating SAM in ENVI software include the set of maximum angle
(radians). The option of single value was set, which indicates that a single threshold was
used for all classes. The field of maximum angle (radians) indicates the maximum
acceptable angle between the endmember spectrum vector and the pixel vector. Pixels
with an angle larger than this value are not classified by the application. The default value
was 0.1 but as proven the value of 0.3 attributed better results.

3.3.4. Classification accuracy assessment

Accuracy assessment of the thematic maps produced from the implementation of the
SVMs and SAM classification techniques to the Hyperion imagery was performed in ENVI
based on the confusion matrix analysis (Congalton, 1991; Congalton & Green, 1999). It
was based on the computation of the overall accuracy (OA), the user’s accuracy (UA),
the producer’s accuracy (PA) and the Kappa statistic coefficient (Kc). OA is the ratio of
the number of validation pixels that have been correctly classified to the total number of
validation pixels used for all classes and is expressed as a percentage (%). Kc is the
proportion of correctly classified validation points after random agreements are removed
and it expresses the extent to which the matrix results are not obtained by chance or
random. In comparison to OA, Kc indicates a more conservative estimation than a simple
percentage value. PA expresses the probability that the classifier has correctly labeled an
image pixel, whereas UA expresses the probability that a pixel belongs to a given class
and the classifier has labeled the pixel correctly into the same given class. In performing
the accuracy assessment herein, a total of 193 sampling points for the different classes
were selected (approximately 40-60 pixels per class) directly from the Hyperion imagery
following a random sampling strategy, and these points formed the validation dataset.
The selection of these validation points was performed following exactly the same criteria
used for the selection of training points. They were selected from Hyperion imagery and
the interpretation was completed with the support of an image of the same date in Google
Earth application.

4. RESULTS AND DISCUSSION

The urban vegetation cover maps produced after the implementation of the SVMs and
SAM classifiers to the Hyperion image of the study region are illustrated below, whereas
the statistical results obtained after the classification accuracy assessment conducted
using the same set of validation points are presented in Table 4.1 for both classifiers
respectively. Both classification methods produced comparable results in terms of
describing not only the spatial distribution but also the cover density of each land cover
category in the test site. it appears that SVMs generally outperformed the SAM classifier
in both overall accuracy and individual classes’ accuracies. SVMs’ overall accuracy and
Kappa coefficient were 86.53% and 0.823 respectively, while SAM’s classification overall
accuracy and Kappa coefficient were 75.13% and 0.672 respectively. Among the two
classifiers, SVMs was more accurate than SAM in describing the spatial distribution of

CEST2013_0209
classes, which was also indicated according to the statistics of the individual classes’
results. In terms of the individual classes’ accuracy, PA varied from 57.6% to 96.9% and
UA varied from 73.3% to 90.5% for SVMs, whereas PA varied from 27.3% to 98.5% and
UA ranged from 52.9% to 86.1% for SAM. On the basis of PA statistical measure, it can
be observed that the class of low vegetation produced the lowest percentage in both
techniques. In addition, as it can be observed the class of low vegetation performed the
lowest percentage in SAM technique and the class of bare soil performed the lowest
percentage in SVM technique. This can be attributed in part to the similar spectral
characteristics between the two classes, which was most probably affected by the mixed
pixels combined with the low spatial resolution of the Hyperion sensor. SVMs have been
designed to identify an optimal separating hyperplane for classes’ separation, which other
machine learning classifiers may not be able to locate. SVMs technique is also able to
generalize this optimal separating hyperplane unseen samples with least errors among all
separating hyperplanes, thus producing the best classes’ separation at the end of
classification (Huang et al., 2002). SVMs are also successful in addressing ill-posted
problems providing high classification accuracy results in comparison to other classifiers,
even in cases when small training sets are used and has important advantages. As
regards SAM, its main advantages are that it is an easy and rapid method for mapping
the spectral similarity of image spectra to reference spectra. It is also a powerful
classification method because it represses the influence of shading effects to accentuate
the target reflectance characteristics (De Carvalho & Meneses, 2000).

Table 4.1. Summary of the results from the classification accuracy assessment
conducted.

Ground truth (percent)


Class SVM PA SVM UA SAM PA SAM UA
(%) (%) (%) (%)
Asphalt 96.92 88.73 98.46 79.01
Buildings 95.00 86.36 77.50 86.11
Bare Soil 64.71 73.33 76.47 56.52
Trees 94.74 85.71 73.68 80.00
Low vegetation 57.58 90.48 27.27 52.94
Producer’s accuracy (%) 86.53 75.13
Kappa Coefficient 0.823 0.672

CEST2013_0209
Figure 4.1. The Hyperion classification using SVMs RBF classifier (right image) and SAM
classifier (left image).

5. CONCLUSIONS

The combined use of Hyperion hyperspectral imagery with the Support Vector Machines
(SVMs) and Spectral Angle Mapper (SAM) classifiers for discriminating different land-
cover classes at a typical urban setting was the main objective of this study. The research
focused on extracting urban vegetation cover. The comparative performance of two
classifiers showed that SVMs outperformed SAM in terms of both overall and individual
classes’ accuracy, at least for the experimental setting. The higher classification accuracy
reported by SVMs is mainly attributed to the fact that this classifier has been designed to
identify an optimal separating hyperplane for classes’ separation. However the main
disadvantage of SVM classification technique is that it is not able to consider the sub-
pixel heterogeneity which is very high in urban areas like the test site of the study.
Moreover, the spatial resolution of sensors such as Hyperion is very high which makes
spectral mixing problematic as the Earth’s surface heterogeneity increases (Xu et al.,
2008). Consequently, there is a demand for further evaluation of classification algorithms
and further research in techniques which perform better results in sub-pixel level and deal
with the problem of heterogeneity. Spectral confusion in pixels, may potentially lead to
classification errors for a spectral class, which may consist ofan important barrier in both
classification approaches examined, especially when those applied in remote sensing
imagery that is not of very high spatial resolution. Finally, results of this study indicated
that using low cost Hyperion hyperspectral satellite imagery can potentially enhance a
wider use of techniques towards managing and monitoring classification of urban

CEST2013_0209
vegetation. Yet, further studies evaluating the ability of different algorithms applied with
Hyperion imagery should be carried out in different urban environments which will allow
improving our understanding on the capabilities of EO technology in mapping urban
vegetation structures and of their changes from space.

Acknowledgments
The authors would like to thank all reviewers for their useful and constructive comments.

References
1. De Carvalho, O.A. and Meneses, P.R.(2000). Spectral correlation mapper (SCM); An
improvement on the Spectral Angle Mapper (SAM). Summaries of the 19th JPL
Airborne Earth Science Workshop. JPL Publication 00-18, 9 p.
2. Congalton, R. G. (1991). A review of assessing the accuracy of classifications of
remotely sensed data. Remote Sensing of Environment, 37, 35–46.
3. Congalton, R., and Green, K. (1999). Assessing the accuracy of remotely sensed
data: principles and practices. Boca Raton, FL: CRC/Lewis Press. 137 pp.
4. C.S.B. Grimmond, M. Roth, T.R. Oke, Y.C. Au, M. Best, R. Betts, G. Carmichael,H.
Cleug, W. Dabberdt, R. Emmanuel, E. Freitas, K. Fortuniak, S. Hanna, P.
Kleinm,L.S. Kalkstein, C.H. Liu, A. Nickson, D. Pearlmutter, D. Sailor and J. Voogt
(2010). Climate and More Sustainable Cities: Climate Information for Improved
Planning and Management of Cities (Producers/Capabilities Perspective),
ELSEVIER, pp. 247–274
5. Foody, G.M., & Arora, M.K. (1997). An evaluation of some factors affecting the
accuracy of classification by an artificial neural network. International Journal of
Remote Sensing, 18, 799-810.
6. Herold, M., Gardner, M., & Roberts, D. (2003). Spectral resolution requirements for
mapping urban areas. IEEE Transactions on Geoscience and Remote Sensing,
41(9), 1907−1919.
7. 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.
8. Herold, M., & Roberts, D. A. (2005). Spectral characteristics of asphalt road aging
and deterioration: Implications for remote-sensing applications. Applied Optics,
44(20), 4327−4334.
9. Jonas Franke, Dar A. Roberts, Kerry Halligan, Gunter Menz (2009). Hierarchical
multiple endmember spectral mixture analysis (MESMA) of hyperspectral imagery for
urban environments. ELSEVIER. Remote sensing of environment 113, 1712-1723.
10. Kruse, F.A., Boardman, J.W., Lefkoff, A.B., Heidebrecht, K.B., Shapiro, A.T.,
Barloon, P.J., and Goetz, A.F.H. (1993). The spectral image processing system
(SIPS) –Interactive visualization and analysis of imaging spectrometer data. Remote
sensing of environment, Vol. 44, p. 145-163.
11. Lu, D., & Weng, Q. (2007). A survey of image classification methods and techniques
for improving classification performance. International journal of remote sensing, 28
(5), 823-870.
12. Paola, J. D., & Schowengerdt, R. A. (1995). A review and analysis of back
propagation neural networks for classification of remotely sensed multispectral
imagery. International Journal of remote sensing, 16, 3033-3058.
13. Petropoulos P. George, Chariton Kalaitzidis, Krishna Prasad Vadrevu (2012a).
Support vector machines and object-based classification for obtaining land-use/cover
cartography from Hyperion hyperspectral imagery. ELSEVIER, Computers &
Geosciences 99-107.
14. Petropoulos P. George, Kostas Arvanitis, Nick Sigrimis (2012b). Hyperion
hyperspectral imagery analysis combined with machine learning classifiers for land
use/cover mapping. ELSEVIER, Expert systems with applications 39, 3800-3809.

CEST2013_0209
15. Rowan, L.C and Mars, J.C. (2003). Lithologic mapping in the Mountain Pass,
California area using advanced spaceborn thermal emission and reflection
radiometer (ASTER) data. Remote sensing of Environment, Vol. 84, p. 250-266.
16. Thoreau Rory Tooke, Nicholas C.Coops, Nicholas Goodwin, James A.Voogt (2009).
Extracting urban vegetation characteristics using spectral mixture analysis and
decision tree classifications. ELSEVIER, Remote Sensing of Environments 113, 398-
407.
17. Vapnik, V., (1995). The nature of statistical learning theory. Springer-Verlag, New
York, NY.
18. Walsh, S. J., McCleary, A. L., Mena, C. F., Shao, Y., Tuttle, J. P., Gonzαlez, A., et al.
(2008). QuickBird and Hyperion data analysis of an invasive plant species in the
Galapagos Islands of Ecuador: Implications for control and land use management.
Remote Sensing of Environment, 112, 1927–1941.
19. Xu, D.-Q., Ni, G.-Q., Jiang, L.-L., Shen, Y-T., Li, T., Li, T., Ge, S.-L., Shu, X-B., 2008.
Exploring for natural gas using reflectance spectra of surface soils. Advances in
Space Research 411, 188-1817.
20. Martine, G. The State of the World Population 2007; United Nations Population Fund:
New York, 2007; pp. 1.
21. Van de Voorde T., Vlaeminck J., and Canters F., 2008. Comparing Different
Approaches for Mapping Urban Vegetation Cover from Landsat ETM+ Data: A Case
Study on Brussels. Sensors 2008, 8, 3880-3902.
22. Xian, G.Z., Crane, M., McMahon, C., 2005. Assessing urban growth and
environmental change using remotely sensed data. Pecora 16 “Global Priorities in
Land Remote Sensing” October 23 – 27, 2005 Sioux Falls, South Dakota
http://www.asprs.org/publications/proceedings/pecora16/Xian_G.pdf [accessed June
28th, 2010].

CEST2013_0209

Das könnte Ihnen auch gefallen