TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan.

2013 (ISSN 0974-3375)

Building Extraction from High Resolution

Satellite Images
1
Susheela Dahiya, 2P. K. Garg, 3Mahesh K Jat
1, 2
Geomatics Engineering Group, Civil Engineering Department
Indian Institute of Technology Roorkee,
3
National Institute of Technology, Jaipur
1
[email protected], [email protected], [email protected], [email protected]

Abstract- A method for automatic building extraction from are almost similar in satellite images which results in error
high resolution satellite image is given. First, the image is in digital classification. In that case, differentiation
segmented by using the split and merge segmentation. Then between buildings and road becomes very difficult.
the segmented image is filtered by applying different filters. Because of this reason, some additional features (like
After filtering the output raster image is converted into
area, shape etc.) are also required for increasing the
vector image and then the buildings are extracted on the
basis of area from the vector image. Finally, the cleanup accuracy of extracted buildings from satellite images.
methods are applied to increase the accuracy of the results.
Imagine Objective tool of ERDAS 2011 is used. IMAGINE Objective in ERDAS 2011 version [3] is one
of the software solutions to object-oriented classification
I. INTRODUCTION and feature extraction available in the market. Compared
to other tools, it is relatively new and less explored by
Identification of different types of objects on the earth users. It is less complex than some of the other tools,
surface is an important aspect in remote sensing which limits its possibilities while making it easier to use
application. During the past decade a considerable [4]. IMAGINE Objective tool employs feature models
research has been done for automatic building which work on objects produced by image segmentation
identification from satellite images. While detecting and various other pixel-based algorithms which, after
buildings from satellite images, spatial resolution (i.e. being vectorised, can be processed by geometric and
smaller pixel size) should be considered more important textural parameters [5]. With object-oriented analysis it is
than the spectral resolution (i.e. greater number of spectral possible to get better results from remote sensing
bands or narrower interval of wavelength) [1]. The recent information.
availability of commercial high-resolution satellite
imaging sensors such as IKONOS and QuickBird provide A lot of work has been done on building extraction from
a new data source for building extraction. The objective high resolution satellite images. The high spatial
of launching and deploying the above commercial remote resolution of the IKONOS imagery reveals very fine
sensing satellite was to increase the visibility of terrestrial details in urban areas and greatly facilitates the
features, especially urban objects, by reducing per-pixel classification and extraction of buildings [6-13]. Liu et. al.
spectral heterogeneity and thereby improving land cover [14] proposed a method for urban building extraction
identification. from pan images. This method is able to extract large
buildings having shadow evidences. Also, this method is
These finer resolution or larger scale image data exhibit not able to extract small buildings. The segmentation
higher levels of detailed features than those from methods play an important role in building extraction
preceding sensors (e.g., Landsat Thematic Mapper and from satellite images.
SPOT), but this greater level of detail may lead to
complicated urban features in the spectral domain [2]. The building detection method is highly sensitive to
This is because many small objects are concentrated in a segmentation performance as well as the selected
small area when dealing with an urban space, and they thresholds. Building extraction by using region growing
become more and more visible as the spatial resolution and mean shift segmentation is proposed in [15]. But
gets finer and finer. This situation potentially leads to these segmentation methods fail in the case of an
lower accuracy in urban image classification. This may extremely complex scene, and the subsequent steps will
not be the case for other environments, especially when not extract the exact shape of the building. Split-and-
dealing with other natural land covers and land uses merge segmentation technique [16] for automated
(rangeland, evergreen forests, broad leaved forests, pine reconstruction of planar building rooftops is
forests, mangroves, wetland, desert landscape, and computationally efficient and the results are of acceptable
agriculture). accuracy.

In India, building extraction is a difficult task, because the II. METHODOLOGY

building doesn’t follow a specific pattern and the
individual building covers a very small area on the The building extraction methodology includes the
ground. In addition, the reflectance of buildings and roads following steps:

829
TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan. 2013 (ISSN 0974-3375)

1. Raster Pixel Processor are submitted to compute pixel cue metrics to train the
pixel classifier.
 Single Feature Probability (SFP)
During the automated extraction phase candidate pixels
2. Raster Objects Creators from imagery are submitted to the pixels classifier for
query to measure how closely they resemble to the
 Split and Merge Segmentation training pixels. The output of this step is a pixel
probability layer in which each pixel value represents the
3. Raster Object Operators
probability that it is the object of interest.
 Probability Filter
Twelve training samples are selected from the satellite
 Size Filter image. No background pixels are selected. The value of
training sample extension is taken 10 pixels and the
 Reclump probability threshold value is taken as 0.100. The
IKONOS image used in this study is shown in Figure 1.
 Dilate

 Erode

 Clump Size Filter

4. Raster to Vector Conversion

 Polygon Trace

5. Vector Object Operators

 Generalize
Figure 1: IKONOS Satellite Image
6. Vector Object Processor
B. Raster Objects Creators
 Area
Here split and merge segmentation along with edge
7. Vector Cleanup Operators detection is used to identify segments with similar
characteristics. Segmentation is a way of partitioning
 Probability Filter raster images into segments based on pixel values and
locations and also to simplify and/or change the
 Island Filter representation of an image into something that is more
meaningful and easier to analyze. The parameters used for
 Smooth image segmentation are Euclidean distance, minimum
value difference, variation factor, threshold and minimum
 Orthogonality
no. of pixels representing an edge length.
A. Raster Pixel Processor Pixels that are spatially connected and have similar values
SFP (Single Feature Probability) is used to perform pixel are grouped in a single segment. This operator performs
based classification. SFP is a pixel cue that computes the segmentation on the raster image specified by the input
probability metric (a number between 0 and 1) to each variable parameter. The result is a thematic image where
pixel of the input image based on its pixel value and the pixels value represent class IDs of contiguous raster
training samples. objects. The output of raster Pixel Processor is taken as
input to compute the pixel probability zonal mean of each
The buildings to be extracted are selected by the user as segment and that zonal mean is used as the value of the
training samples and based on the pixel values of the segments Pixel Probability attribute.
training samples, the probability values are assigned.
Higher probability values are assigned to those pixels This operator then converts the Pixel Probability Layer
whose values are similar to ones of pixels in the training into a Raster Object Layer. The output raster object layer
samples. Lower probability values are assigned to pixels contains pixels that are grouped as raster objects. A raster
whose values are significantly different from the values of object is a collection of pixels that touch each other and
pixels in the training samples. are discontinuous to the other raster objects. The pixels in
these raster objects share a common class ID and also
The training samples must be chosen very carefully and have an attribute table. For each raster object, the zonal
should not include any background pixel. Pixels other mean probability from the Pixel Probability Layer is
than the pixels that are similar to the training pixels are computed and stored as an attribute. This is the starting
considered as the background pixels. During the training probability value for the object. Output from the
phase, pixels that are representing the individual objects Probability Pixels to Raster Objects Operator contains

830
TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan. 2013 (ISSN 0974-3375)

pixels that are grouped as raster objects which have pixels in a raster object that have any neighboring pixels
associated probability attributes. that are not included in the raster object. Kernel size and
kernel shape both are required to perform this function.
Minimum Value Difference = 30 Kernel Size value should be an odd number. The value of
Kernel size is 3 and kernel shape is square. The output
Variation Factor = 3.50
image after applying probability filter is shown in Figure
Threshold for Edge detection = 10 2(e).

Minimum length = 3 Pixels 6) Clump Size Filter

C. Raster Object Operators This operator filters out raster objects that have been
combined (clumped) with all connected neighboring raster
This class of operators takes the Raster Object Layer from objects. These objects are smaller than a specified
the previous operation and performs mathematical minimum size of 20 pixels. Using Clump Size Filter, the
morphology on the objects in the raster domain. The original objects are retained. Filtering out objects can
output is a new Raster Object Layer. improve efficiency of the model, since fewer objects are
processed in later stages of the model. The output image
1) Probability Filter after applying probability filter is shown in Figure 2(f).
Output from the Raster Object to Raster Object Operator
This operator removes all raster objects whose zonal contains pixels that are grouped as raster objects which
probability mean is less than the specified Minimum have associated probability metrics.
Probability. The higher the minimum probability, the
more objects are removed. The value taken to perform
probability filtering on the image is 0.50. The output
image after applying probability filter is shown in Figure
2(a).

2) Size Filter

This operator filters out raster objects that are too small,
too large, or both. Here, filtering out the objects improves
the efficiency of the model, since fewer objects are
processed in later stages of the model. For processing this
filter both or any one from minimum and maximum pixel
size values must be entered. The values taken as input are (a) (b)
20 pixels. The output image after applying probability
filter is shown in Figure 2(b).

3) Reclump

This operator performs a clump operation on the input

raster object layer. The input layer is usually a layer of
raster objects which have been split, joined or filtered by
other Raster Object Operators, and no longer have correct
clump values. This operator will renumber the raster
objects so that each raster object has a unique value. The
output image after applying probability filter is shown in
(c) (d)
Figure 2(c).

4) Dilate

This operator expands raster objects. It considers all

pixels near the raster object. If any pixel of the raster
object is adjacent (horizontally, vertically or diagonally)
to the pixel being considered, it is added to the raster
object. Kernel size and kernel shape both are required to
perform this function. Kernel Size value should be an odd
number. The value of Kernel size is 3 and kernel shape
was taken as square. The output image after applying
probability filter is shown in Figure 2(d). (e) (f)

5) Erode Figure 2: Image after applying (a) Probability Filter (b) Size Filter (c)
Reclump (d) Dilate (e) Erode (f) Clump Size Filter
This operator shrinks raster objects. It will remove all

831
TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan. 2013 (ISSN 0974-3375)

D. Raster to Vector Conversion G. Vector Cleanup Operators

Here the raster objects are converted to the vector domain. These classes of operators are performed on the Vector
The Raster Object Layer from the previous step is taken Object Layer output from the Object Classifier query and
as input and converts each raster object into a vector are intended for probability thresholding and for cleanup
object, normally a polygon or polyline, and then produces of the object polygon or polyline footprint.
a Vector Object Layer. The rest of the process flow will
deal with these vector objects. 1) Probability Filter

Polygon Trace is used as a raster to vector converter that This operator removes all vector objects whose zonal
traces the outlines of raster objects and converts them to probability mean is less than the specified Minimum
polygonal vector objects. Probability. The minimum probability of 0.50 is taken to
filter out all vector objects
E. Vector Object Operators
2) Island Filter
These operators perform operations on the vector objects,
and produce a new Vector Object Layer. These operations This operator removes all islands or holes of an object.
may be necessary to change the shape of the vector The value taken for Maximum Object size is 450 sq.
objects by generalizing, smoothing, etc. This operator meters.
simplifies polylines and polygons by removing
unnecessary vertices. It helps remove noise and improves 3) Smooth
the performance of downstream processing. This operator smoothens the polygon objects in a vector
layer. The smoothing factor is specified to control how
For removing noise and pixel artifacts in a polygon or
polyline a tolerance value must be entered. This algorithm much smoother the objects become. Enter any value
creates trend lines from each arcs starting point, and to between 0 and 1. 0 indicates no smoothing is applied to
other arcs. Then the distance is measured for each vertex the objects, while 1 applies the greatest smoothing. For
less smoothing, choose less than 0.5. For more smoothing
to its trend line.
(for example, to cut sharp turns more), choose more than
If the distance of the vertex to the trend line is greater than 0.5. The value taken here is 0.5. The bigger the factor, the
the tolerance, the vertex is then removed. Thus, a smoother is the polygon objects.
tolerance of 0 will not remove any vertices and a larger
tolerance will remove many vertices. The value of 4) Orthogonality
tolerance entered in this model is 0.50 meters. The operator first smoothens and generalizes the vector
object, then segments of the object that fall below either
F. Vector Object Processor
the angle threshold or segment threshold are removed
For this reason object training can be specified with from the object. The remaining segments are rotated about
distribution parameters such as a mean and standard the segment’s center point to the nearest orthogonal angle
deviation. During the automated extraction phase and the segments are then re-intersected.
candidate objects from Vector Object Layer are submitted
Output from the Vector Object to Vector Object Operator
to the Object Classifier for query to measure how closely
and contains the vector objects in a layer. This is the final
they resemble the training objects (or fit the parametric
distributions). result of the process. Along with the vector objects are the
different probability metrics and cue algorithm metrics
Of course, objects processed during the extraction phase used in the process. The value taken here is 0.5.
must also pass through the Object Cue Metrics
calculation. The query result for an object will be III. RESULTS
recorded as an object attribute to be used in downstream The method described above automatically extracts
processes. buildings from satellite images. The original IKONOS
image and the extracted buildings from that image are
Area: This metric computes the area of each polygon
shape in the input shapefile. The area of islands within the shown in Figure 3(a) and 3(b). Also, the buildings are
other polygon is subtracted from the outer polygon’s area. manually digitized so that the extracted buildings can be
The output shapefile will have an area attribute containing compared with the original buildings.
the result of this metric. The values taken are: The buildings of the IKONOS image are manually
digitized as shown in Figure 4(a). The buildings extracted
Minimum value = 10 sq. meters
from the proposed method are shown in Figure 4(b). A
Maximum Value = 450 sq. meters comparison is shown in Figure 5.

Mean = 188 sq. meters There are 122 buildings present in the original IKONOS
image. After applying the proposed method, 120 patches
Std. Dev. = 162 sq. meters are extracted as buildings. After comparing the manually
digitized buildings and the extracted buildings, we found

832
TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan. 2013 (ISSN 0974-3375)

that 24 buildings are not extracted at all and 18 patches IV. CONCLUSION
are wrongly identified as buildings.
These results can be improved by adding some more
parameters like shape, shadow (if shadow is present in the
image), texture etc. to the given approach. There is still
much scope for improvement with respect to increase the
extraction rate and the refine delineation of building.
There is also lots of additional information that can be
incorporated into the analysis. In future work, to increase
the extraction rate, some more parameters like shape,
shadow (if shadow is present in the image), texture etc.
can be used in extraction process.

REFERENCES
[1] S. W. Myint, P. Gober, A. Brazel, S. Grossman-Clarke and
Q. Weng, 2011, "Per-pixel vs. object-based classification of
urban land cover extraction using high spatial resolution
imagery," Remote Sensing of Environment, vol. 115, pp.
1145-1161, 2011.

(a) (b) [2] S. W. Myint, V. Mesev and N.S.N. Lam, 2006 , "Texture
analysis and classification through a modified lacunarity
Figure 3: (a) Satellite Image (b) Automated Extracted Buildings analysis based on differential box counting method.,"
Geographical Analysis,, vol. 38, p. 371−390, 2006.

[3] "Imagine Objective, User's Guide 2011," 2008.

[4] L. C. Chepkochei, 2011, "Object-oriented image

classification of individual trees using Erdas Imagine
Objective: Case study of Wanjohi area, Lake Naivasha
Basin, Kenya," proceedings, Kenya Geothermal
Conference, November 21-22, 2011.

[5] N. Lack and S. Bleisch, 2010,"Object-based change

detection for a cultural-historical survey of the Landscape -
From cow trails to walking paths," The International
Archives of the Photogrammetry, Remote Sensing and
(a) (b) Spatial Information Sciences,, vol. XXXVIII-4/C7, 2010.

Figure 4: (a) Manually Digitized Buildings (b) Automated Extracted [6] J. A. Benediktsson, M. Pesaresi and K. Amason, 2003,
Buildings "Classification and feature extraction for remote sensing
images from urban areas based on morphological
transformations," IEEE Transactions on Geoscience and
Remote Sensing, vol. 41, pp. 1940-1949, 2003.

[7] S. K. W. Par, and T. Kim, 2000, "Line-rolling algorithm for

automated building extraction from 1-meter resolution
satellite images," in Proc. International Symposium on
Remote Sensing, Kyung-Ju, Korea, , pp. 31–36, November
2000.

[8] G. Sohn and I. J. Dowman, 2001,"Extraction of buildings

from high resolution satellite data," Automated Extraction
of Man-Made Objects from Aerial and Space Images (III).
Balkema Publishers, Lisse, pp. 345-355, 2001.
Figure 5: Comparison of Manually Digitized Buildings and Automated
Extracted Buildings [9] D. S. Lee, J. Shan, and J. S. Bethel, 2003, "Class-guided
building extraction from Ikonos imagery,"
TABLE 1 Photogrammetric engineering and remote sensing, vol. 69,
pp. 143-150, 2003.
Total No. of Extracted Wrongly Buildings [10] K. Segl and H. Kaufmann, 2001,"Detection of small objects
Buildings in Buildings Identified not from high-resolution panchromatic satellite imagery based
Original Image Buildings Identified on supervised image segmentation," Geoscience and
Remote Sensing, IEEE Transactions on, vol. 39, pp. 2080-
2083, 2001.
122 120 18 24

833
TECHNIA – International Journal of Computing Science and Communication Technologies, VOL.5 NO. 2, Jan. 2013 (ISSN 0974-3375)

[11] E. P. Baltsavias, M. Pateraki, L. Zhang, 2001 , [14] Y. Wei, Z. Zhao and J. Song, 2004, "Urban building
"Radiometric and geometric evaluation of Ikonos GEO extraction from high-resolution satellite panchromatic
images and their use for 3D building modelling," Institute image using clustering and edge detection," pp. 2008-2010
of Geodesy and Photogrammetry, ETH Zürich, 2001. vol. 3, 2004.

[12] J. P. Muller, C. Ourzik, T. Kim and I. J. Dowman, [15] S. Cui, Q. Yan and P. Reinartz, 2012, "Complex building
1997,"Assessment of the effects of resolution on automated description and extraction based on Hough transformation
DEM and building extraction," Automatic Extraction of and cycle detection," Remote Sensing Letters, vol. 3, pp.
Man-Made Objects from Aerial and Space Images (II), pp. 151-159, 2012.
245-256, 1997.
[16] K. Khoshelham, Z. Li and B. King, 2005, "A split-and-
[13] A. K. Shackelford and C. H. Davis, 2003,"A combined merge technique for automated reconstruction of roof
fuzzy pixel-based and object-based approach for planes," Photogrammetric engineering and remote sensing,
classification of high-resolution multispectral data over vol. 71, pp. 855-862, 2005.
urban areas," Geoscience and Remote Sensing, IEEE
Transactions on, vol. 41, pp. 2354-2363, 2003.

834