Building Extraction from Very High Resolution

Multispectral Images using NDVI based

Segmentation and Morphological Operators
Durgesh Singh, Rohit Maurya, Ajay Shankar Shukla, Manish Kumar Sharma and P.R Gupta

Abstract— In this paper, we have proposed an approach for

building extraction from very high resolution (VHR)
multispectral images using NDVI (Normalized Difference
Vegetation Index) based segmentation and morphological
operations. This approach uses both spatial and spectral
properties of an image scene for building detection. Spectral
properties are related to NDVI based segmentation and spatial
properties are related to the morphological operations. Normally
an image scene consists of natural region (vegetation and soil)
and manmade regions (buildings and roads). Use of NDVI
(spectral properties) eliminates the chance of shadow being a
building region and other similar regions that are not road like
soil, vegetation etc because shadow is a spatial property and
NDVI is based on spectral property irrespective of brightness in
the image. By using NDVI we can eliminate the natural regions
from the manmade regions because NDVI values of these two
regions differ a lot, so using NDVI as a threshold we can segment
image into two parts one is natural and other is our desired parts
that consists of manmade regions (buildings and roads). After
segmentation here comes the use of spatial property, use of
morphological operation to separate the roads from building Figure 1 High Resolution Multispectral Image of Buildings
regions on the basis of their spatial property that roads have
elongated and larger area than buildings and mostly building has Lots of previous works have been done on building extraction
the rectangular rooftops. This approach provides very but their performance differs on the basis of quality of result,
satisfactory results with very less overhead and time. speed and the overhead of data availability. Some of which are
semiautomatic and needs user interaction [10] so based on
Keywords- Building extraction, NDVI, Segmentation,
previous work our aim is to develop some method that will
Morphology.
overcome the problem and obtain better results. Aytekin et al.
[1] have developed an approach that is based on remotely
I. INTRODUCTION sensed image composed of low resolution multi spectral image
The building extraction has always been a keen interest area and high resolution bands. They have first identified
for image processing researchers. Due to recent advances in vegetation area using NDVI then using YIQ color space
high resolution image, multispectral image and aerial images it shadow area is extracted. By masking this area only manmade
has gained more attention. Building extraction is very area are left then mean shift algorithm is performed followed
important due to its application in disaster management, by morphological operation and principal component analysis
military application, urban planning, traffic planning. So the to finally extract the building segments. Lie hu et al. [2] has
various semiautomatic and automatic building extraction used the shadow to extract the building on the basis of the
methods have been developed trying to automate the whole shadow angle with the building and structure of shadow. They
building extraction process and to reduce the effort and time
have deduced the fact that the structure of shadow is
required. But it is very difficult and challenging task due to
complexity of building shapes and the similarity between the somewhat will be similar to building rooftops. Aaron K.
building and other similar area like road that are not building. Shackelford et al. [3] has used DMP (Differential
Building extraction also depend on the quality of image, quality Morphological Profile) to extract the building and shadows
of brightness, weather conditions in which image has taken, form the satellite image. Where DMP is a multi-scale image
that’s why it is not possible to obtain 100% results. So by analysis technique where an image profile is constructed
minimizing the dependency of building extraction methods on through the use of morphological opening and closing
these factors we can achieve better results. operation with varying size structuring element. Kannika
Komwong et al. [4] has used supervised classification to
extract the building extraction. Sebastien Lefevre et al. [5] has
978-1-4673-0455-9/12/$31.00 ©2012 IEEE
extracted the building segments based on the assumption that A. Segmentation using NDVI
building either appear as a darker or brighter in the image,
Normalized difference vegetation indices (NDVI) have been
using cluster fusion and morphological operation. Z. J. Liu et extensively used for monitoring and detecting vegetation and
al. [7] has used the multi-scale object oriented classification land cover change. Plants absorb light energy with the help of
and probabilistic hough transforms to extract the building. In chlorophyll, a green pigment present in leaves in order to
the first step multispectral and panchromatic high resolution perform photosynthesis. Light is absorbed more in the red and
image are fused for spatial resolution improvement and color reflected in near-infrared regions. Vegetation Index (NDVI)
information enhancement then multi resolution image which can be defined as follows:
segmentation is applied and in the second step probabilistic
Hough transform is perform to extract the buildings. Our aim NDVI = ρNIR - ρred / ρNIR + ρred (1)
is to minimize the overhead [1] and dependency [2] to get Where ρ is the spectral reflectance.
better results. Our approach consists of basically two main
The equation produces values ranging from –1 to 1.
steps first is the segmentation of high resolution multispectral
image using NDVI and the second step is the removal of false Negative values are indicative of clouds, snow, water and
region using morphological operator. Use of NDVI in other non-vegetated, non-reflective surfaces, while positive
multispectral image makes the segmentation independent of values denote vegetated or reflective surfaces. Soils which
the shadow region because shadow is the spatial property and generally exhibit a near-infrared spectral reflectance somewhat
NDVI is based on the spectral property so by calculating the larger than the red, and thus tend to also generate rather small
NDVI for vegetation, soil and other region we can eliminate positive NDVI values (around 0.3). So the NDVI value for
the area that are not buildings. We don’t have to worry about vegetation is 0.3. [8]
whether shadow is there or not because if there is shadow on So the basis of NDVI values for soil and vegetation we can
the ground it will be detected as false region by masking the segment the image into the natural region and the manmade
soil area. In section II we discuss about the method of building region or we can binarize the image because we are segmenting
extraction. After extracting the buildings we evaluate the the image into two segments.
result based on various quality factors like correctness and
completeness in the section III.
Input Image
II. BUILDING EXTRACTION METHOD

Read NIR and Red Bands

Image

Calculate NDVI
Segmentation Using NDVI = (NIR – Red) / (NIR +
NDVI Red)

Filtering Band Threshold

B(x) = 1 if I(x) <= T (0.30)
0 otherwise
Morphological Opening

Output segmented
Extracted or Binary Image
Buildings

Figure 3 Process of Segmentation Using NDVI

Figure 2 Process of Building Extraction

In the first step the near infra red (NIR) and red bands are
The process of building extraction consists of three steps as read then NDVI value is calculated. In the last step band
shown in figure 2. The first step will segment the image into threshold is done to remove vegetation and soil area from the
two parts, manmade region and natural region. In the second building and similar manmade area.
step manmade region is filtered to remove the larger segments
like road that are not buildings. In the third step the B(x) = 1 if I(x) <= T (0.30) (2)
morphological operation is applied on the image that is
0 otherwise
obtained from the second step to extract the building based on
the assumption that building has rectangular shape.
 A. B. C. D.
Figure 4 (A) NDVI classification (B) Band Threshold (C) Extracted Building (D) Building boundaries

Where B is output binary image. I is the input image and We have used the erosion to find out the rectangular
T is the NDVI threshold which value is 0.30 (Manmade area shape, but also shrink the buildings. So we have used dilation
false below this value). in next step to restore the shape of buildings.
B. Filtering Final output is shown in Figure 4(C).
In a particular scale image, the small isolated region
should be affirmed as non-building. Therefore, after the III. RESULT EVALUATION
image segmentation, use mathematical morphology filter
with the right parameter after image to remove these isolated The method for evaluating the results of building extraction
points. The area of the building should have a reasonable can be done by completeness, correctness and quality
scale. The region whose area is too big is not building measure evaluation. The quality of output depends on how
certainly.
much part of our building region we have extracted.
Filtering is used to separate the building area from other
area. Completeness, Correctness and Quality
MinThresh < Buildings < MaxThresh (3)
The accuracy of building extraction technique can be
Here we used the concept that building cannot be determined by measuring the completeness and correctness.
smaller than a minimum threshold because building has
In order to evaluate the results the quality measures
some minimum predictable size, below which it may be
some vehicle or noise. completeness, correctness and quality (Heipke et al., 1997)
[9] are examined. The classification of building extraction is
MaxThresh is used to remove the area like river or road done as true positive (TP), false negative (FN), or false
because building has some maximum predictable size and positive (FP). Where true positive (TP) means the number of
cannot be larger than roads or river.
buildings among the extracted building that are actually
So in filtering process the isolated region which has area buildings. False positives (FP) means that the number of
greater than minimum threshold and lesser than the buildings among the extracted results that are not buildings
maximum threshold is extracted and other false area is and falsely extracted as building. False negative (FN) means
filtered. no of building that are actually building but are not extracted
and left as non-building area.
C. Morphological Opening
Completeness = Number of extracted buildings that are
Based on the assumption that building has usually actually building
rectangular shape the morphological opening is performed Total number of building in a scene
on the filtered image in order to extract the buildings. [6]
Morphological opening is defined as a combination of ≈ TP (5)
erosion and dilatation: TP + FN
O (I) = (I ⊖ S) ⊕ S (4)
Completeness ϵ [0; 1]
Where I = input image
S = Structuring element of rectangle type Completeness tells us how complete our approach is. It
means that how many buildings out of the actual no of the
O = Output Image buildings have been extracted.
 Correctness = Number of extracted buildings that are IV. CONCLUSION
actually building As we see in the table 1 the building extraction method has
Total number of extracted buildings very good results in terms of completeness, correctness and
quality. This method provides very satisfactory results but
≈ TP (6) building region with very complex shape needs to research
TP + FP further.

Correctness ϵ [0; 1]
Correctness tells us how correct our approach is or what is
the chance of being a building of each building that is being
extracted.
Quality = Number of extracted buildings that are actually
building
No of building extracted in output

≈ TP (7)
TP +FP+ FN

Quality ϵ [0; 1]

Quality tells us how good approach is. It is the measure of

both completeness and correctness the more complete and
correct approach the more the quality. Figure 6 Output 2

Building extraction was performed on landsat data and REFERENCES

results are given in the following table. [1] Aytekin, O.; Ulusoy, I.; Erener, A.; Duzgun, H.S.B.; , "Automatic and
unsupervised building extraction in complex urban environments
from multi spectral satellite imagery," Recent Advances in Space
TABLE 1: RESULTS Technologies, 2009. RAST '09. 4th International Conference on , vol.,
no., pp.287-291, 11-13 June 2009.
TP FP FN Completeness Correctness Quality [2] Hu, Lei; Zheng, Jin; Gao, Feng; , "A building extraction method
106 9 18 0.92 0.85 0.79 using shadow in high resolution multispectral images," Geoscience
and Remote Sensing Symposium (IGARSS), 2011 IEEE International
206 5 25 0.89 0.97 0.87 , vol., no., pp.1862-1865, 24-29 July 2011.
[3] Shackelford, A.K.; Davis, C.H.; Xiangyun Wang; , "Automated 2-D
The output the results are shown in the figure 5 and figure 6 building footprint extraction from high-resolution satellite
respectively. multispectral imagery," Geoscience and Remote Sensing Symposium,
2004. IGARSS '04. Proceedings. 2004 IEEE International , vol.3, no.,
pp. 1996- 1999 vol.3, 20-24 Sept. 2004.
[4] Komwong, K.; Simking, R.; Nuangjumnong, P.; , "Automatic 2D
Building Extraction Using High Resolution Image in Bangpli District,
Samut Prakan Province, Thailand," Computational Science and Its
Applications (ICCSA), 2010 International Conference on , vol., no.,
pp.277-280, 23-26 March 2010.
[5] Lefevre, S.; Weber, J.; , "Automatic Building Extraction in VHR
Images Using Advanced Morphological Operators," Urban Remote
Sensing Joint Event, 2007 , vol., no., pp.1-5, 11-13 April 2007.
[6] Acar, U.; Bayram, B.; , "Building extraction with morphology,"
Recent Advances in Space Technologies, 2009. RAST '09. 4th
International Conference on , vol., no., pp.33-38, 11-13 June 2009.
[7] Liu, Z.J.; Wang, J.; Liu, W.P.; , "Building extraction from high
resolution imagery based on multi-scale object oriented classification
and probabilistic Hough transform," Geoscience and Remote Sensing
Symposium, 2005. IGARSS '05. Proceedings. 2005 IEEE
International , vol.4, no., pp. 2250- 2253, 25-29 July 2005.
[8] Richard Tran Mills, Forrest M. Hoffman, Jitendra Kumar, William W.
Hargrove,Cluster Analysis-Based Approaches for Geospatiotemporal
DataMining of Massive Data Sets for Identification of Forest
Figure 5 Output 1 Threats,International Conference on Computational Science, ICCS
2011,Procedia Computer Science 4 (2011) 1612–1621.
[9] C. Heipke, H. Mayer, C. Wiedemann, and O. Jamet. Evaluation of
automatic road extraction. In International Archives of
Photogrammetry and Remote Sensing, volume 323-4W2, pages 151-
160, 1997.
[10] Jiang, N.; Zhang, J.X.; Li, H.T.; Lin, X.G.; , "Semi-automatic
building extraction from high resolution imagery based on
segmentation," Earth Observation and Remote Sensing Applications,
2008. EORSA 2008. International Workshop on , vol., no., pp.1-5,
June 30 2008-July 2 2008.