Professional Documents
Culture Documents
ORIENTED METHOD
1.Photogrammetric Engineering & Remote Sensing, School of Civil Engineering, Beijing Jiaotong University, Beijing,
100044, China. E-mail: qltanbjtu@163.com; 2. Institute of Photogrammetry and Remote Sensing, Chinese Academy of
Surveying and Mapping, Beijing, China; E-mail: zjliu@casm.ac.cn
Authorized licensed use limited to: GOVERNMENT COLLEGE OF TECHNOLOGY. Downloaded on June 26,2010 at 22:00:32 UTC from IEEE Xplore. Restrictions apply.
the user guide of eCognition software. Grassland and Rural Woodland. Structure Group 3 just
contained only Water body. In the following classification-
2. METHODOLOGY based segmentation (fusion), all neighboring objects that
belong to the same structure group were combined to larger
The approach mainly involves establishing an image object objects. In this study, lengthy objects can be expected for
hierarchy consisting of three levels of different resolution. water bodies, very large objects for rural areas, and medium
In our study, different classifications have been applied to sized objects for the urban areas.
different image object levels. Information from different
image object levels has to be used in the process of
mapping the degree of imperviousness.
Firstly, we created a classified super-scale image object
level by fusing pre-classified thematic image objects. On
the highest object level only three classes would be
distinguished: possibly impervious (urban areas), which are
the areas of interest, rural areas that are not impervious, and
water. To achieve this, three abstract classes were
constructed for the highest object level. They are Water
body, Not Impervious, and Possibly Impervious. Our main
interest is the urban areas, which is where urban
imperviousness has to be expected.
The pre-classified thematic image objects were acquired
by segmentation using default scale parameter 10, and then
performing classifications using Nearest Neighbor classifier
and Membership Function. Nearest neighbor is a classifier
used to classify image objects based on given sample
objects within a defined feature space. After a number of Fig.1 Landsat TM image (R: band3;G:band2;B:band1)
sample objects have been declared as initial information for
a nearest neighbor classification, we can classify the image
objects in the whole scene using the NN classifier. After the
first NN classification, four classes: Woodland General,
Grassland General, Impervious General and Water-body
were classified. With the class-related features derived from
the first classification, we performed further classification
for the above general classes. For example, using the class-
related feature of the relative area of objects of the class
Impervious General within a perimeter of 10 pixels around
the object in question, the child classes (Urban Grassland
and Rural Grassland) of Grassland General can be further
classified. Using the same class-related feature, the child
classes (Urban Woodland and Rural Woodland) of
Woodland General can be also further classified. With the
Impervious General, the child classes of it (Urban
Impervious and Rural Impervious) were classified by
combining the features of area and density with the class- Fig.2 Classified super-scale image object level
related feature of relative border to neighbor objects
classified as Urban Impervious. The latest class-related Now this highest image object level already existed due
classification provided the knowledge base for the fusion of to the classification-based fusion. The objects in this level
the image objects to larger, meaningful objects. will become super-objects of all the objects in the two
After the above two classifications, three structure following generated lower levels. We had to further classify
groups of classes were constructed. Structure Group 1 the highest image object level, because super-object’s
contained all urban classes: Urban Impervious, Urban classification information was needed to use for the
Grassland and Urban Woodland. Structure Group 2 classification of the sub-objects by determining whether
contained all rural classes: Rural Impervious, Rural
V - 165
Authorized licensed use limited to: GOVERNMENT COLLEGE OF TECHNOLOGY. Downloaded on June 26,2010 at 22:00:32 UTC from IEEE Xplore. Restrictions apply.
they belong to a possibly impervious or not impervious classified as Completely Impervious.
super object. Now the objective is to determine the objects in the
On this image object level, water objects can easily be lowest level that represent completely impervious or dense
discerned using a spectral feature (the near infrared mean spots. It has already been ascertained that the near infrared
value). To fully assign an object to Water body class, its layer is best for analyzing urban structures. It is advisable
near infrared mean value must be 50 or smaller. All objects to extract features from it for the classification of dense
that urban imperviousness can occur in were assigned to areas as well. So, the following class rules can be
Possibly Impervious. It was known from the former formulated: objects of the class Completely Impervious
classification and fusion that these are objects that have have a near infrared mean value less than or equal to 59,
been created out of the former urban objects. With that and also have a mean value in blue band layer of 70 or
knowledge, the following class rules can be built. Objects higher. After classification on this level completion, urban
of the class Possibly Impervious: (1) must have a near structures are well represented. Fig. 3 shows the classified
infrared mean value larger than 90 and less than 95, (2) small-scale image object level.
must be of 25,000 pixels or smaller in size, and (3) are not
objects of the class Water body. Objects of the class Not
Impervious are neither objects of the class Possibly
Impervious nor objects of the class Water body. Fig. 2
shows the final classified super-scale image object level.
This classification provides super-scale information for
the classification of the lower levels. That is, for each sub-
object in a lower level, it will know if it belongs to a
possibly impervious urban area, a non-impervious rural
area or water body. In this way, the super-scale
classification information can aid us in further
classifications.
V - 166
Authorized licensed use limited to: GOVERNMENT COLLEGE OF TECHNOLOGY. Downloaded on June 26,2010 at 22:00:32 UTC from IEEE Xplore. Restrictions apply.
assigned to Urban Area, if it has a super-object classified as combination of buildings, roads, grass, trees, soil, water,
Possibly Impervious. By analogy, an object is assigned to and so on. The object-based approach is still fairly new in
the class Pervious Rural Area, if it is not Urban Area and various remote sensing application fields, but it is attracting
not Water body. more and more attention on the part of this community. In
Finally, we performed a final class-related classification our study, although we also just tested the Landsat TM
in the middle level, using the above sub-scale and super- dataset, but we have proved that for estimating and
scale information. Urban imperviousness only occurs in mapping of the degree of urban impervious surface, the
areas represented by objects classified as Urban Area. object-oriented remote sensing image analysis method is an
Other than Water and Not Impervious, for each object effective, simple and rapid way to acquire the urban
impervious surface information. It is not only suitable for
classified as Urban Area the part of its sub-objects’ area
the medium spatial resolution images, but also especially
classified as Completely Impervious will be calculated.
suitable for high-resolution images for its huge data volume,
Thus, image objects in urban areas have been divided into
because segmentation drastically reduces the sheer number
six classes (0%, 1-25%, 26-50%, 51-75%, 76-99%, and of units to be handled for classification. Even if a lot of
100% impervious) according to the actual degree of intelligence is applied to the analysis of each single image
imperviousness they represent. After the classification object, the classification works relatively fast. Furthermore,
completed, a thematic map of showing the degree of homogeneous image objects provide a significantly
imperviousness in urban areas has been created. Fig.4 increased signal-to-noise ratio compared to single pixels as
shows the final classification of the degree of to the attributes to be used for classification. Thus,
imperviousness in urban areas on middle-scale image object independent of the multitude of additional information, the
level. classification is more robust and more accurate comparing
From the final classification of imperviousness degree, with pixel-based methods when using the same resolution
we can examine what share of the urban area the single remote sensing data.
impervious degrees cover by dividing the total areas of the
impervious degree classes by the total areas (urban area). 4. REFERENCES
Tab.1 shows the statistical result. Also, analysis of their
spatial distributions can give us important information [1] Arnold, C.L. Jr., and Gibbons, C.J., “Impervious surface
coverage: the emergence of a key environmental indicator”.
about the economic and environmental implications relating Journal of the American Planning Association, 1996, 62, 243–258.
with the city development status. [2] Qihao Weng, Remote sensing of impervious surfaces (edited
by Qihao Weng), Taylor & Francis series in remote sensing
Tab. 1 Statistical Result of the Degree of Urban applications, Boca Raton, CRC Press, 2008.
Imperviousness [3] S. Phinn, M. Stanford, P. Scarth, A. T. Murray, and P. T. Shyy,
“Monitoring the composition of urban environments based on the
Share of urban area
Class vegetation impervious surface-soil (VIS) model by subpixel
(%) analysis techniques,” Int. J. Remote Sens., vol. 23, no. 20, pp.
0% impervious 4.1 4131-4153, 1999.
1-25% impervious 11.0 [4] M. Ji and J. Jensen, “Effectiveness of subpixel analysis in
26-50% impervious 24.6 detecting and quantifying urban imperviousness from Landsat
51-75% impervious 27.7 Thematic Mapper imagery,” Geocarto Int., vol. 14, no. 4, pp. 33-
41, 1999.
76-99% impervious 31.8 [5] C. Wu and A. T. Murray, “Estimating impervious surface
100% impervious 0.8 distribution by spectral mixture analysis,” Remote Sens. Environ.,
vol. 84, pp. 493-505, 2003.
3. DISCUSSIONS [6] Xiaojun Yang, “Estimating Landscape Imperviousness Index
From Satellite Imagery”, IEEE Geoscience and Remote Sensing
Most previous researches for extraction of impervious Letters, vol. 3, no. 1, January 2006, pp. 6-9.
surfaces in urban areas adopted pixel-based methods on the [7] Baatz, M., Heyen, M., and Hofman, P. (2004). eCognition ̢
medium spatial resolution images such as Landsat Object oriented image analysis. User Guide. Definiens AG.
Abrufbar unter: http://www.definiens-imaging.com.
TM/ETM+ and Terra’s ASTER images. However, their
[8] Baatz, M. and Sch¨ape, A. (2000). Multiresolution
spatial resolutions are regarded as too coarse, due to the
segmentation – an optimization approach for high quality multi-
heterogeneity of urban landscapes and the complexity of
scale image segmentation. In: J. Strobl, and T. Blaschke (Eds.):
impervious surface materials. Urban landscapes are Angewandte Geographische Informationsverarbeitung XII.
typically composed of features that are smaller than the Beitr¨age zum AGIT-Symposium Salzburg. Heidelberg, New
spatial resolution of such sensors and are a complex York.
V - 167
Authorized licensed use limited to: GOVERNMENT COLLEGE OF TECHNOLOGY. Downloaded on June 26,2010 at 22:00:32 UTC from IEEE Xplore. Restrictions apply.