Satellite Imagery Segmentation: a region growing approach
LEONARDO SANT’ANNA BINS
LEILA M. GARCIA FONSECA
GUARACI JOSÉ ERTHAL
FERNANDO MITSUO II
INPE--National Institute for Space Research
DPI -- Image Processing Division
C.P. 515, 12201 São José dos Campos, SP, Brasil
leila@dpi.inpe.br
Abstract. This work presents a segmentation method based on a region growing approach. It has been
implemented in the geographic information and image processing system (SPRING) which has been
developed at INPE. The technique is applied to segment images which are being used to assess land use
changes in the Amazon region. Segmented Landsat-TM images are shown to illustrate the technique.
Keywords: Image Segmentation, Region Growing, Euclidean Distance.
1 Introduction
Image segmentation is a basic task in image analysis
whereby the image is partitioned into meaningful
regions whose points have nearly the same properties,
e.g., grey levels, mean values or textural properties.
The segmentation process is one of the first steps
in the remote sensing image analysis: the image is
partitioned into regions which best represent the
relevant objects in the scene. Region attributes such as
area, shape, statistical parameters and texture can be
extracted and used for further analysis of the data.
The segmentation task can be accomplished in two
ways: 1) dividing up the images into a number of
homogeneous regions, each having a unique label, 2)
determining boundaries between homogeneous regions
of different properties. These segmentation techniques
are known as region-based segmentation and edge
detection, respectively.
Each approach is affected differently by various
factors. For some applications edge detection approach
has not been successful. The prime cause is the presence
of small gaps in edge boundaries which allow merging
of dissimilar regions. Other disadvantages are that these
techniques are also often very sensitive to local
variations intensity and the contours obtained are
usually not closed. Therefore, in order to yield closed
boundaries the edges must be linked up.
On the other hand, region-based segmentation
always provides closed contour regions which is a
requirement in many applications. Besides, it is very
simple and effective in many applications. Errors in the
regions boundaries are the main drawback of this
approach: edge pixels might be joined to any of the
neighboring regions. Among region-based segmentation
approaches are region growing methods which will be
discussed in the next section in more detail.
Other segmentation techniques combine the edge-
based and region-based information taking into account
their complementary nature (Le Moigne and Tilton,
1995) or use ancillary information to guide the
segmentation procedure (Mason et al, 1988). The
selection of any of these segmentation approaches
greatly depends on the type of data being analyzed and
on the application area.
Within this framework, this paper describes a
region growing segmentation method which has
demonstrated technical feasibility for images of forest
and agricultural regions. This algorithm has been
developed at INPE and implemented in the Geographic
Information and Image Processing System- SPRING
(DPI et al, 1995 ). It has been intensively used in the
segmentation of Amazon region images to assess land
use changes. Some resulting segmentation of Landsat-
TM images are also shown.
2 Region growing approach
Let X denote the grid of sample points of the image,
and P let be a logical predicate which measures the
homogeneity of a region. The segmentation can be
defined (Zucker, 1976; Schoenmakers et al , 1991) as a
partition of X into disjoint non-empty regions
R R Rn1 2, , . . . , so that the following conditions hold:
i. Ri i n, ,2, . . . ,= 1 is digitally connected, i.e., the
regions must be composed of contiguous lattice
points (see Rosenfeld (1970) for a discussion of
connectivity in digital pictures).
ii. ∪
=
=
i
n
Ri X
1
.
iii. ( )P Ri TRUE= for i n= 1,2, . . . , .
iv. ( )P Ri R j FALSE∪ = for i j≠ , where Ri and R j
are adjacent (4-connected at some point).
Zucker (1976) has pointed out that these properties
suggest many important aspects of the segmentation
algorithms but do not lead to a unique algorithm for
performing the segmentation. Many segmentation
schemes have incorporated these conditions but
Schoenmakers et al (1991) have proposed some changes
in these constraints in order to adapt the algorithms with
heuristics appropriate to each application, given the
needs of the end-user.
The region growing technique is an iterative
process by which regions are merged starting from
individual pixels, or another initial segmentation, and
growing iteratively until every pixel is processed.
Roughly speaking, it can be described by the following
steps:
1. Segment the entire image into pattern cells ( 1 or
more pixels).
2. Each pattern cell is compared with its neighboring
cells to determine if they are similar, using a
similarity measure. If they are similar, merge the
cells to form a fragment and update the property
used in the comparison.
3. Continue growing the fragment by examing all of its
neighbors until no joinable regions remain. Label
the fragment as a completed region.
4. Move to the next uncompleted cell, and repeat these
steps until all cells are labeled.
The drawback of this traditional scheme is that, at
each iteration, one or several merges occur so that the
resulting segmentation is dependent on the order of the
merges. In order to solve this problem Tilton (1989) has
proposed an iterative parallel algorithm in which only
the best merges are authorized. At each iteration, a set
of subimages is defined, and the most similar pair of
spatially adjacent regions is merged in each subimage.
The segmentation here proposed is simpler than Tilton’s
algorithm but it has been provided satisfactory results in
the segmentation of some forest and agricultural
images.
3 Segmentation algorithm
Our approach is based on the traditional region
growing technique, with some modifications which
partially solve the problem of the dependence on the
order of the merges.
Before describing the segmentation scheme some
notation and definitions used in this section is provided.
ℜ is used to denote the set of regions of the image, and
R ∈ℜ is an element of this set. Let ( )T t denote the
threshold value below which two regions are considered
similar at instant t, and let Mi be the mean value vector
of the region Ri . Let ( )D Ri Rk Mi Mk, = − be the
Euclidean distance between the spectral mean values of
the regions Ri and Rk , and let N R( ) be the set of
neighboring regions of R (not including R itself). The
region Rk is the most similar neighboring region of Ri
if ( ) ( )D Ri Rk D Ri R, ,≤ l for every ( )R N Ril ∈ .
The stages of the procedure can be outlined in the
following way:
1. In the beginning of the segmentation process, a list
of regions { }Ri i n, , . . . ,= 1 is created ( n is the
number of pixels in the image). Initially each region
is composed of only one pixel, so-called “seed”. For
each region Ri , its mean value vector and
neighboring regions are stored.
2. For each region Ri its neighboring regions N Ri( )
are examined and:
• the most similar neighboring region
Rk N Ri∈ ( ) is chosen. If ( ) ( )D Ri Rk T t, <
then Rk is called “the best neighbor” of Ri .
• If the best neighbor of Rk exists and is Ri ,
then both regions are merged.
3. Every time one region is aggregated to another one
it is taken out of the list.
4. The fragment mean value is updated every time one
region is aggregated to it.
5. The same procedure is repeated until no joinable
regions remain.
6. In the last step, small regions are merged with larger
adjacent regions, in accordance with an area
threshold value defined by the user.
In our implementation ( ) ( )T t t T= α 0 , with
( )T 0 0> , t = 0 1, ,2, . . . and α < 1. This specification
imposes that the initial merges are harder to accomplish
than those at the end of the merging process, i.e., only
very similar regions are merged first.
The similarity threshold value must be manually
provided by the user and, therefore, a tradeoff is
inevitable: if it is set too low the growing process will
generate over-segmented regions, otherwise segments
representing different land cover will be incorrectly
merged together. The choice of this threshold value as
well as the area threshold value will greatly depend on
the specific application and data.
3 Results
The region-based segmentation algorithm has been
tested mainly on forest and agricultural images with
satisfactory results. In order to show here some results,
the Landsat TM images of bands 3, 4, and 5 (path/rows
231.67, July 12, 1994 ) of two test sites were processed.
These images cover an area in the Amazon region
(Rondônia, Brazil).
Figures 1(a) and 2(a) show the original images
(band 5) of the two different regions, namely site 1 and
site 2, respectively. For visualization of the resulting
segmentation, the original image overlapped with the
regions boundaries are shown in the Figures 1(b) and
2(b) corresponding to sites 1 and 2, respectively. For
the site 1 the bands 3, 4 and 5 were used in the
segmentation process while only the band 5 was
processed in the case 2. As we can observe the results
obtained so far have been encouraging. The region
boundaries have good correspondence with the contours
of the landcover in the test images.
The technique here presented has been intensively
used by remote sensing specialists to assess land use
changes in the Amazon region (Alves et al, 1996;
Batista et al, 1994; Santos et al, 1995 ). These works
show more segmentation results.
4 Conclusions
There is still much work to be done on the problem of
segmentation of satellite images. This is a critical
problem since satellite images are very complex. Here
we have presented a practical and simple segmentation
scheme based on a region growing approach. In spite of
its simplicity a preliminary assessment of the technique
using images of Amazon region has shown satisfactory
results. In the future we plan to apply the algorithm in
other kinds of images and to perform a quantitative
evaluation in more details. Improvements on the scheme
proposed, such as the integration of the region-based
and edge-based information, which has been proposed
in many papers, are also planned.
Acknowledgments
The authors would like to thank E. K. Mello and J. C.
Moreira for providing the images and the segmentation
results shown in the Figure 2.
References
D. S. Alves, J. C. Moreira, E. K. Mello, J. V. Soares, O.
Fernandez, S. Almeida, J. D. Ortiz, “Mapeamento do
uso da terra em Rondônia utilizando técnicas de
segmentação e classificação de imagens Landsat-TM”,
Submetido ao VIII Simposio Brasileiro de
Sensoriamento Remoto, Salvador, 1996.
G. T. Batista, J. S. Medeiros, E. M. K. Mello, J. C.
Moreira, L. S. Bins, “A new approach for deforestation
assessment”, In: ISPRS Symposium on Resource and
Environmental Monitoring, Rio de Janeiro, Brazil, V.
30, Part 7a, pp. 170-174, 1994.
DPI et al, “SPRING: Integrating Remote Sensing and
GIS by Object Oriented Data Modeling”, Computer
Graphics, 20(3), 1996. (To be published).
J. Le Moigne, J. C. Tilton, “Refining Image
segmentation by integration of edge and region data”,
IEEE Transactions on Geoscience and remote Sensing
33(3), 605-615, May 1995.
D. C. Mason, D. G. Corr, A. Cross, D. C. Hogg, D. H.
Lawrence, M. Petrou, A. M. Tailor, “The use of digital
map data in the segmentation and classification of
remotely-sensed images”, International Journal of
Geographical Information Systems, 2(3), 195-215,
1988.
A. Rosenfeld, “Connectivity in digital pictures”, J.
ACM, 17, 146-156, 1970.
J. R. Santos, A. Venturieri, R. J. Machado, “Monitoring
land use in Amazonia based on image segmentation and
neural networks”, IGARSS’95, V.I, pp.108-111, 1995.
R. P. H. M. Schoenmakers, G. G. Wilkinson, Th. E.
Schouten, “Segmentation of remotely-sensed images: a
re-definition for operational applications”, IGARSS’91
Espoo, Finland, V.II, 1087-1090, 1991.
J. C. Tilton, “Image segmentation by iterative parallel
region growing and splitting”, in Proc. 1989 Int.
Remote Sensing Symp., Vancouver, BC, Canada, July
10-14, pp. 2420-2423, 1989.
S. W. Zucker, “Region growing: childhood and
adolescence”, Computer Graphics and Image
Processing 5, 382-399, 1976.
(a)
(b)
Figure 1. (a) original image (band 5) of the Amazon
region (site 1) (b) edge boundaries superimposed with
the original image.
(a)
(b)
Figure 2. (a) original image (band 5) of the Amazon
region (site 2) (b) edge boundaries superimposed with
the original image.
sumario: Sumário
botao_sumario:
cb: Anais VIII Simpósio Brasileiro de Sensoriamento Remoto, Salvador, Brasil, 14-19 abril 1996, INPE, p. 677-680.
677: 677
678: 678
679: 679
680: 680
本文档为【卫星图像分割】,请使用软件OFFICE或WPS软件打开。作品中的文字与图均可以修改和编辑,
图片更改请在作品中右键图片并更换,文字修改请直接点击文字进行修改,也可以新增和删除文档中的内容。
该文档来自用户分享,如有侵权行为请发邮件ishare@vip.sina.com联系网站客服,我们会及时删除。
[版权声明] 本站所有资料为用户分享产生,若发现您的权利被侵害,请联系客服邮件isharekefu@iask.cn,我们尽快处理。
本作品所展示的图片、画像、字体、音乐的版权可能需版权方额外授权,请谨慎使用。
网站提供的党政主题相关内容(国旗、国徽、党徽..)目的在于配合国家政策宣传,仅限个人学习分享使用,禁止用于任何广告和商用目的。