GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

Water Area Extraction from RGB Aerophotograph Based on Chromatic and

Textural Analysis
Meng Zhao Huazhe Shang Wenchao Huang Lizhi Zou Yongjun Zhang
School of Remote Sensing and Information Engineering,
Wuhan University
Wuhan, China
[email protected] [email protected] [email protected] [email protected] [email protected]

Abstract—When triangulating RGB aerophotograph, if effective methodology has been proposed to extract water
automatically and randomly selected matching pass points area from RGB aerophotograph. In order to supply such a
unfortunately locates into water areas, these points, limited by gap, we put forth a methodology in this paper to extract
their inaccuracy, will decrease the precision of triangulation. water area from RGB aerophotograph.
Therefore, extraction of water area beforehand is conducive to With the enhancement of image resolution,
eliminate water falling pass points and guarantee the quality of aerophotographs contain more abundant space information
aerotriangulation. A new methodology to extract water area as well as geometric and textual features, which create a
from RGB aerophotograph is put forth in this paper. favorable condition for extracting objects via chromatic and
Procedure initiates by segment ing the whole aerophotograph
segmenting
textual analysis. W. Ma and B. S. Manjunath had used
segments.. Subsequently, compute
into homogenous and united segments
chromatic and textural features of every segment and compare
textural analysis to create a texture thesaurus for browsing
segment’’s features to sampled water segments
each segment segments’’ features. large aerophotographs and achieved relatively good retrieval
Finally extract those segments whose chromatic and textural effect [3]. Other trials include W. Niblack’s querying image
segmentss’ as water areas. This
features are similar to sampled segment using color, texture and shape [4]. All these researches
methodology has a relatively obvious merit in effectiveness and manifest a fact that chromatic and textural analyses are
generality. effective in extracting and retrieving objects.
Enlightened by chromatic and textural analysis, we
Keywords
Keywords——CIELAB
CIELAB;; chromatic analysis
analysis;; textual analysis
analysis;; propose a new series of procedures to extract water areas
segmentation;; ISODAT
watershed segmentation ISODATAA from RGB aerophotographs. Procedure initiates by
segmenting the whole aerophotograph into homogenous and
I. INTRODUCTION united regions. Subsequently, compute chromatic and
Aerotriangulation is a key step in aerophotogrammetry. textural features of every region and then compare each
The basic goal of aerotriangulation is to compute all region’s feature to sampled water regions’ features. Finally
elements of exterior orientation and pass points of a region extract those segments whose chromatic and textural features
by block adjustment. The accuracy of matching points on are similar to sampled segments’ as water areas. This
aerophotograph directly relates to the precision of block methodology has a relatively obvious merit in effectiveness
adjustments and the final production of aerotriangulation. and generality.
However, water areas, influenced by wind force and gravity, Water areas on aerophotographs are usually homogenous
are usually in irregular motion. If automatically and and united; therefore water regions after segmentation
randomly selected matching pass points (a few pass points always are homogenous and united. We have no need to care
selected beforehand to compute the exterior orientation about the accuracy in segmentation of other objects because
elements) unfortunately falls into water areas, these points, they contribute little to water extraction. We choose
limited by their inaccuracy, will decrease the precision of watershed segmentation algorithm (Section Ⅲ ) because it
triangulation. Therefore, extraction of water areas can help attain satisfactory segmentation result. The
beforehand is conducive to eliminate water falling pass homogenous and united characteristic of water region also
points and guarantee the quality of aerotriangulation. In this provides pleasing condition to conduct chromatic analysis
paper, we concentrate on extracting water areas from RGB (Section Ⅱ ) and textural analysis (Section Ⅳ ). Chromatic
aerophotograph and mark them. and textural features of every region will be added to its own
Multispectral aerophotograph and remote sensing image feature vector (Section Ⅴ ). Water appears differently on
can synthesize information of different spectrums to extract photographs. In order to automatically recognize water areas,
water region. J. Deng operated different bands of SPOT-5 we need to sample water regions of all appearances
images to extract water area [1]. H. Xu modified MNDWI beforehand and classify them. ISODATA algorithm (Section
value to extract water area from multispectral remotely Ⅴ ) is employed to classify those samples into several
sensed data [2]. In contrast to multispectral data, RGB categories. If the distance is within a threshold between a
aerophotograph (every pixel of the image is consisted of red, certain region’s feature vector and any one of those
green and blue values) has only three bands and little extra
spectrum information. Consequently, until recently, no

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 46

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

categories’, then this segment will be identified and extracted a stage when only the tops of the dams are visible above the
as water area. water lines. These dam boundaries correspond to the divide
RGB aerophotograph consists of three channels (Red, lines of the watersheds. Therefore, they are boundaries
Green, and Blue) and they are closely pertinent to each other. extracted by a water shed segmentation algorithm [7].
We need to integrate them into one channel in order to
conduct image segmentation and textural analysis. In this II. CIELAB COLOR SPACE AND CHROMATIC ANALYSIS
paper, we utilize CIELAB color space to integrate those There are two steps to transform RGB color space to
three channels, which decreases their pertinence as well as CIELAB color space [5]. The first step is to apply the
provides prerequisites for image segmentation and textural following matrix to convert RGB color space to CIEXYZ
analysis later on. color space.
Any color is unable to be both green and red or both blue
and yellow. This principle prompts the CIELAB color space.
CIELAB color space inherits the XYZ standard color system. ⎡ X ⎤ ⎡ 0.412453 0.357580 0.180423⎤ ⎡ R ⎤
CIELAB indicates distinctions between light and dark, red ⎢ Y ⎥ = ⎢ 0.212671 0.715160 0.072169⎥ ⋅ ⎢ G⎥ (1)
and green, and blue and yellow by using three axes: L*, a*, ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
b*. The central vertical axis represents lightness (L*) whose ⎣⎢ Z ⎦⎥ ⎣⎢ 0.019334 0.119193 0.950227⎥⎦ ⎢⎣ B ⎦⎥
value ranges from 0 (black) to 100 (white). On each axis the
Subsequently, (X*, Y*, Z*) need to be calculated.
value runs from positive to negative. On a-a' axis, positive
values indicate amounts of red while negative indicate When X / X n , Y / Yn , Z / Z n > 0.008856 ,
amounts of green. On the b-b' axis, yellow is positive and
blue is negative. For both axes, zero is neutral [5]. X= 3
X / Xn (2)
“Texture” refers to the arrangement or characteristic of
the constituent elements of anything. A texture feature is a Y= 3
Y / Yn (3)
value, computed from the image of a region, which
quantifies gray-level variation within the region. According Z= 3
Z / Zn (4)
to Kenneth [6], a texture feature value is irrelevant to When X / X n , Y / Yn , Z / Z n ≤ 0.008856 ,
region’s position, orientation, size, shape and brightness
(average gray level). Therefore, the shape and size of X* = 7.787 ⋅ (X / X n ) + 0.138 (5)
segmented area do not interfere with the result of chromatic Y* = 7.787 ⋅ (Y / Yn ) + 0.138 (6)
and textural analysis.
According to Julesz [7] and his deduction, human texture Z* = 7.787 ⋅ (Z / Zn ) + 0.138 (7)
discrimination is based on the second order statistic of image
intensities, which gave rise to the emergence of a popular (X n , Yn , Z n ) = (0.312779, 0.329184, 0.358037) represents
textural descriptor: co-occurrence matrix. A co-occurrence the white reference point, which indicates a completely matte
matrix counts the exact times of different grey level pairs of white body.
pixels, separated by a certain distance (one pixel, two pixels, Then L*, a*, b* can be calculated through following
etc.). The (i, j) element of the co-occurrence matrix P for a equations:
certain region is the number of times, divided by M, that L = 116 ⋅ Y * −16 (8)
gray levels i and j occur in two pixels separated by that a* = 500 ⋅ (X*-Y*) (9)
distance and lying along that direction in the region, where
b* = 200 ⋅ (Y * − Z*) (10)
M is the number of pixel pairs contributing to P. The P
The application of CIELAB color space in this paper
matrix is N by N, where the gray scale has N shades of gray
[6]. can be briefly illustrated using diagramⅠ.
Watershed segmentation is an approach based on the DIAGRAM Ⅰ APPLICATION OF CIELAB COLOR SPACE
concept of morphological watersheds. In such a
“topographic” interpretation, every pixel’s gray level
denotes its “height”. For a particular regional minimum, the
set of points at which a drop of water, if placed at the
location of any of those points, would fall with certainty to a
single minimum is called the catchment basin. And the
points at which a water drop would be equally likely to fall
to more than one such minimum constitute crest lines on the
topographic surface and termed watershed lines. Assume
that a hole is punched in each regional minimum and that
the entire topography is flooded from below by letting water The L* component of CIELAB color space can be used
rise through the holes at a same rate. When the rising water in watershed segmentation (Section Ⅲ). L* component can
from different catchment basins is about to merge, a dam is also be used in textual analysis (Section Ⅳ ). We use
built to prevent the merging. The flooding will finally reach equations mentioned above to convert every pixel on an

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 47

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

aerophotograph from RGB color space to CIELAB color In Fig. 1, (b) shows a correct and clear boundary between
space. Then form a gray image by using the L* component water area and non-water area, while result (a) displays an
of every pixel’s CIELAB color space. Although the gray inaccurate boundary. This manifests effectiveness and
photograph is no longer quantified 256 gray levels, L* correctness of the L* component in segmenting RGB
component brings unexpected results in chromatic analysis aerophotographs.
and textural analysis later on.
Segmented areas usually have similar color space and IV. TEXTURAL ANALYSIS
we can use the average a*, b* of all pixels in a region to In our methodology, we replace traditional 256 scale gray
represent its chromatic feature. Experiments have with the L* component of CIELAB color space. We attain
demonstrated that chromatic feature between non-water two advantages by this substitute. First of all, the RGB
areas and water areas is obvious, which means a*, b* can be components of aerophotograph has been integrated into one
employed to chromatically describe a region. We utilize a* L* component, which ranges from 0 to 100; in addition, the
and b* as parameters to describe a segmented region dimension of co-occurrence has been decreased from
chromatically in feature vector (Section Ⅴ). If regions have 256*256 to 101*101. Traditionally, we reduce quantization
similar chromatic feature with sampled water areas, they are level of the input data (e.g., from 8-bit data with values from
possibly water areas we want to extract. 0-255 to 5-bit data with values from 0 to 31) when creating
co-occurrence matrices so that the matrices will not become
III. IMAGE SEGMENTATION
too large. But this method usually causes texture information
In our methodology, we replace the 256 gray level with loss, especially detailed information. Taking advantage of
L* component of CIELAB color space. Using this the L* component of CIELAB color space, the
replacement, we could form a gray picture by synthesizing co-occurrence matrices will decrease its dimension with
three channels of RGB aerophotographs. Compared to relatively little textural information loss.
simply using average gray level of three channels to form a Haralick proposed 14 features based on co-occurrence
gray photograph, the usage of L* component can decrease matrices [8]. The angular second-moment feature (ASM) is a
fragments after segmentation and attain a relatively clear measure of homogeneity of the image.
and accurate edge. 2
ASM = ∑∑ p ( i, j) (11)
i j

p(i, j) is the joint probability of gray pair (i, j) in

co-occurrence P.
Texture is an innate property of virtually all surfaces [8].
Although water seems homogenous and with little obvious
texture, texture feature can still be utilized to discriminate
water area from other non-water area because water area has
a bigger homogeneity than other objects.
In a homogeneous image, there are very few dominant
gray-tone transitions. Hence the P matrix will have some
entries of large magnitude and the ASM feature will be
relatively bigger. In contrast, an image with irregular textures
(a) will have a large number of small entries and hence the ASM
feature will be smaller [8]. Because water areas are usually
homogeneous and united, so the ASM feature for water areas
is relatively bigger than other objects. So we employ ASM as
a textural descriptor.
In order to avoid texture rotating, we compute the
average value of ASM in four directions (0°, 45° ,90°
and 135°) to texturally describe a region. ASM is added to
the feature vector (SectionⅤ).
V. CLASSIFICATION AND WATER EXTRACTION
A. Feature Vector
A vector is a quantity that has magnitude and direction
(b)
and that is commonly represented by a directed line segment
Figure 1. Comparison of Segmentation: (a) is attained by using the
average gray level of RGB to form a gray picture. (b) is attained by using
whose length represents the magnitude and whose
L* component of CIELAB color space to form a gray picture orientation in space represents the direction. Feature vector
in this paper denotes that the quantity consists of a region's
feature values. We define feature vector as following:

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 48

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

FV = ( ASM, a*, b *) (12) SDV

vector (SDV FV. The jth
SDV). Both have the same dimensions as FV
class's CCVj and SDVj can be computed out using following
Different regions have different chromatic and textural equations.
features; therefore every segmented region possesses its
exclusive FV
FV, which represents its uniqueness among other W
FVj
regions. However, slight difference in FV exists among CCVj = ∑ W
(13)
i =0
regions with similar textural and chromatic features, which
means we could use classification algorithm to categorize
2
regions with similar feature vectors. Every region's FV has W
(FVj − CCVj )
to be computed for later extraction. SDVj = ∑ W
2
(14)
i =0
B. Sampling and Unsupervised Classification Using
ISODATA Classification (Where W is the number of regions in jth class and FV
denotes the ith region's FV
FV.)
Practical experience tells us that variances, such as According to statistical principles, when the amount of
contamination, fluctuation, ship transportation, etc, render samples in each class has reached a certain level, the
water display different color and texture feature on RGB distribution of segments’ vectors in every class is subject to
aerophotographs. Therefore, the segmented image has more Gauss distribution. Nearly all members of the class fall
than one kind of water. From the segmented image we need within CCVj ±3·SDV SDVj. Ideally, we can sample numerous
to sample a certain number of water regions via our practical regions within a class and work out its CCV and SDV SDV. Then
experience. Such experience can simply tell water regions compare remaining regions’ feature vectors to the class
apart from other regions like forest and residence, however, center. If a region’s feature vector falls within CCV SDV
CCV±3·SDV
SDV,
it is less reliable when applied to discriminate differences we identify it belonging to the class.
within waters. To settle down this problem, we propose to Unfortunately it is impossible to sample enough water
use ISODATA algorithm to automatically classify sampled regions to form Gauss distribution within a class, but we can
water regions into several categories. still eliminate those non-water regions by checking whether
The ISODATA algorithm [9], abbreviated for Iterative their FV
FVs locate within a threshold of a certain class’s CCV
CCV.
Self-Organizing Data Analysis Technique, is a modification We define a parameter "ratio" to mark the dynamic range of
of the k-means clustering algorithm. ISODATA is each CCV
CCV, that is to say regions with CCV falling into
self-organizing because it requires relatively little human CCV
CCV±ratio·SDVSDV will be extracted as water area.
input. Therefore it is a good choice to substitute human
perception to classify waters. VI. TRIALS AND RESULTS
ISODATA algorithm begins by setting a certain amount
In order to test the feasibility of our methodology, we
of cluster centers for all the samples. Then classify all
samples by shortest distance algorithm. Modify the centers conduct following five trials. We define three indexes to
of every cluster. If two clusters' separation distance in feature evaluate the result of extraction. First index is “w-w”, which
space is below a user-specified threshold, then the two means pixels extracted as water using methodology in this
cluster centers should be merged into one and reclassify all paper and perceived as water by human eyes. Second index
the samples. If the current iterative time is odd or the amount is “w-n”, which means pixels extracted as water using
of clusters is half fewer than expected, then split the cluster methodology in this paper but perceived as non-water by
that has the maximum subparameter of standard deviation human eyes. Third index is “n-w”, which means pixels
vector and reclassify all samples. If the current iterative time extracted as non-water using methodology in this paper but
is even or the amount of clusters is twice bigger than perceived as water area by human eyes.
expected, then merge two clusters whose separation distance Trial 1: The following Fig. 2 is the water extraction effect
is the closest and reclassify all samples. When the iterative of Fig. 1 (b). We have 9 samples of water segments.
time has reached maximum time, then the whole algorithm ISODATA algorithm classifies those 9 samples into 3
stops. We have conducted quite a number of experiments different types. The final extraction is attained when ratio is
and find that the result of classification is satisfactory if only set at 1.1.
the maximum standard deviation and minimum distance
between cluster means are properly set (details, [9]). In our
methodology, every segment's FV is the sample in
ISODATA algorithm.
C. Water Extraction
Each sampled water region has its unique FV FV. As
discussed above, similar water regions have slight difference
in FV
FV; therefore, ISODATA algorithm will automatically
classify regions with analogous FV into one same class.
After the completion of ISODATA algorithm, each class has
CCV
a cluster center vector (CCV
CCV) and a standard deviation

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 49

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

Figure 2. Extraction Result of Trial 1 by setting the ratio at 1.1 (b)

TABLE I. EVALUATION OF TRIAL 1

Fig. 2 w-w w-n n-w Total Pixels

Pixels 260018 2108 18 262144
Percentage 99.2% 0.79% 0.01% 100.0%

Trial 2: The following Fig. 3 (a) is taken of somewhere in

Shenzhen, China. We notice that the segmentation of water
area is homogenous and united in Fig. 3 (b). We have 8
samples of water segments. ISODATA algorithm classifies
those 8 samples into 2 different types. The final extraction is
attained when ratio is set at 1.2 as Fig. 3 (c) shows. (c)

Figure 3. Extraction Result of Trial 2: (a) is the original aerophotograph

of Shenzhen, China. (b) is the result of watershed segmentation. (c) is the
final extraction of water by setting the ratio at 1.2

TABLE II. EVALUATION OF TRIAL 2

Fig. 3 (c) w-w w-n n-w Total Pixels

Pixels 3941135 70597 43726 4055458
Percentage 97.2% 1.7% 1.1% 100.0%

Trial 3: To test the effectiveness in the combination of

chromatic and textural analysis, we conduct another
(a) extraction using the aerophotograph in Trial 1. When
conducting chromatic analysis only, the feature vector will
be shortened to have only a* and b* parameters. When
conducting textural analysis, the feature vector will be
shortened to have only ASM parameter. We also selected 8
same water areas as samples in trial 1. In chromatic
extraction only, ISODATA classifies those 8 samples into 3
classes and the extraction result is Fig. 4 (a). In textural
extraction, ISODATA classifies those 8 samples into 4
classes and the extraction result is Fig. 4 (b). For both trials,
the ratio is set 1.2 as the ratio set in trial 1. From the result
we can see that the chromatic extraction or textural
extraction individually can not attain an effect as good as
the combination of them. The percentage of w-n and n-w is

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 50

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

a nightmare for correctively extract water areas compared to

the combination of both chromatic and textural analysis.

(a)

(b)
(a) Figure 5. Extraction Result of Trial 4:(a) is the orginal RGB
aerophotograph. (b) is attained using the CCV generated in trial 1 when
ratio is set 1.2.

TABLE IV. EVALUATION OF TRIAL 4

Fig. 5 (b) w-w w-n n-w Total Pixels

Pixels 1437772 26133 10216 1474121
Percentage 97.5% 1.8% 0.7% 100.0%

Trial 5: In order to further test the generality of this

methodology, we apply it to extract water areas from
aerophotograph with many fluctuations. Following
aerophotograph (Fig. 6 (a)) is taken of somewhere in
Qingdao. The biggest characteristic of this aerophotograph
(b) is its rippled surface. As a result of this irregular fluctuation,
the segmentation result (Fig. 6 (b)) is not very satisfactory.
Figure 4. Extraction Result of Trial 3: (a) is attained using ony
There are many fragments in water area, but we can
chromatic analysis when ratio is set 1.2. (b) is attained using only textural manually merge those fragments into united one in order to
analysis when ratio is set 1.2 improve the result of segmentation. We choose 12 water
samples from Fig. 6 (b). ISODATA algorithm automatically
TABLE III. EVALUATION OF TRIAL 3
classifies those samples into 4 categories. When ratio is set
Fig. 4 (a) w-w w-n n-w Total Pixels 1.5, we can attain a pleasing result (Fig. 6 (c)).
Pixels 2601923 391245 425432 3418600
Percentage 76.1% 11.4% 12.5% 100%
Fig. 4 (b) w-w w-n n-w Total Pixels
Pixels 2615619 398715 743726 3758060
Percentage 69.6% 10.6% 19.8% 100.0%

Trial 4: In order to test the generality of this methodology,

we use the CCV generated in trial 1 to extract water area
from another aerophotograph (Fig. 5 (a)) taken of the same
area. Because they are the same region, therefore their
chromatic and textural features are similar. The following
(a)
result (Fig. 5 (b)) is attained when ratio is set 1.2.

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 51

GEOProcessing 2011 : The Third International Conference on Advanced Geographic Information Systems, Applications, and Services

ACKNOWLEDGMENT
Special thanks to Professor Yongjun Zhang’s guide in
this research and the cooperation of our team members.

REFERENCES

[1] J. Deng, “An Effective Way for Automatically Extracting Water

Body Information from SPOT-5 Images”, Journal of Shanghai
Jiaotong University (Agricultural Science), vol. 23, no. 2, pp.
198-201, 2005, June.
[2] H. Xu, “A Study on Information Extraction of Water Body with the
(b) Modified Normalized Difference Water Index(MNDWI)”, Journal of
Remote Sensing, vol. 9, no. 5, pp. 589-595, 2005, September.
[3] W. Ma and B. S. Manjunath, “A Texture Thesaurus for Browsing
Large Aerial Photographs”, Journal of the American Society for
Information Science, CA, vol. 49, no. 7, pp. 633-648, 1999.
[4] Niblack W, Barber R et al., “The QBIC project: Querying images by
content using color, texture and shape”, Proc of SPIE: Storage and
Retrieval for Image and Video Database, San Jose, CA, pp. 58-70,
1994.
[5] A. R. Robertson, “Historical development of CIE recommended color
difference equations”, Color Research and Application, 15(3), pp.
458-461, 1990.
[6] K. R. Castleman, Digital Image Processing, New Jersey: Prentice
Hall, Inc., a Simon&Schuster Company, pp. 499, 1996.
[7] B. Julesz, “Experiments in the visual perception of texture”, Scientific
Americian, vol. 232, pp. 34-43, 1992, May.
[8] R. M. Haralick and K. S. Dinstein, “Textural Features for Image
Classification”, IEEE Transactions on Systems Man and Cybernetics,
(c) pp. 610-621, 1973, November.
Figure 6. Extraction Result of Trial 5: (a) is the orginal RGB [9] J. R. Jensen, Introductory Digital Image Processing-A Remote
aerophotograph. (b) is the result of segmentation. (c) is the final result of Sensing Perspective, 3rd ed., New Jersey: Prentice Hall, pp. 383-389,
extraction. 1995.

TABLE V. EVALUATION OF TRIAL 5

Fig. 6 (c) w-w w-n n-w Total Pixels

Pixels 316823 10192 716 327731
Percentage 96.7% 3.1% 0.2% 100.0%

VII. CONCLUSION
It is innovative to combine watershed segmentation,
CIELAB color space and chromatic and textural analysis as
well as ISODATA classification together to extract water
from RGB aerophotograph. Through numerous
experimental comparisons, we discover that ratio set
between 1.1 and 1.8 is optimal for water extraction.
However, it is undeniable that ripples or ship transportation
will cause a few fragments after segmentation, which put
sand in the wheel of chromatic analysis and textural analysis
because feature values of those fragments are not accurate
and not representative. Although merging those fragments
into united and homogenous areas is a good solution to this
problem, it demands much more time and manual
involvements. This is what we could do to improve the
result of water extraction later on.

Copyright (c) IARIA, 2011. ISBN: 978-1-61208-118-2 52