Sie sind auf Seite 1von 4

Water Body Extraction from Landsat ETM+ Imagery

Using Adaboost Algorithm


Li Shen*, Changchun Li
College of Resources Science & Technology
Beijing Normal University
Beijing, China
*Corresponding author: shenli@ires.cn

Abstract—Extracting water body information accurately from for identifying the water boundary. Many water body
remotely sensed imagery is significant for surveying, planning extraction strategies from ETM+ imagery have focused on
and protecting water resources, and particularly flood disaster analyzing spectral discrepancy in different bands between
management such as monitoring, evaluation, emergency response water and other ground objects, and making use of a single
and so on. Landsat ETM+ imagery has such features as high band or multiband combination to recognize water body object.
spatial resolution and multi-spectral resolution, which provides a Employing the 5th channel of ETM+ imagery, namely short
rich, reliable and accurate source of fundamental data for wave infrared band, by setting appropriate threshold value K,
research on water body extraction. Now, researchers have water and land can be identified roughly, but is prone to
investigated lots of techniques and methods for deriving water
misclassify mountain, cloud and shadow as water [1]. Zhou
information automatically. Previous work was often devoted to
finding a sophisticated classifier to identify water body, and it is
found that the water body exits special spectral characteristic:
not only difficult to make full use of available features, but also TM band 2 + TM band 3 > TM band 4 + TM band 5 (the
easy to get into predicament of training the complex classifier. In reflectivity value of the nth channel of ETM+ imagery denoted
this paper, a novel scheme for water body extraction from by TM band n) [2]. This method is able to partly eliminate the
Landsat ETM+ imagery using the adaboost algorithm is influence of shadow, but still hard to distinguish residential
proposed. It is based on such consideration that finding many region and water. Dividing the pixels in green band by the
rough rules of thumb can be a lot easier and more effective than corresponding pixels in near-infrared band, Mcfeeters proposed
finding a single, highly prediction rule. Adaboost is a general the normalized water index (NDWI) [3], which largely
strategy for learning classifiers by combining simple ones. The restrains vegetation information, and highlight the water body
idea of adaboost is to take a “weak classifier” - that is, any information. Based the improvement on NDWI, Xu developed
classifier that will do at least slightly better than chance – and use a new water index, MNDWI (Modified NDWI), which works
it to build a much better classifier, thereby boosting the especially well on extracting water body in urban areas [4].
performance of the weak classification algorithm. The excellent
property of adaboost is the ability to integrate disparate Obviously, previous work on water body extraction was
classifiers that concentrate on different aspects of the problem, often devoted to finding a sophisticated classifier to identify
and place more weight on features that could train more accurate water body, and it is not only difficult to make full use of
base classifiers. This paper, with Landsat ETM+ imagery as available features, but also easy to get into predicament of
study object data, combines weak classifiers which are training the complex classifier.
constructed by spectral information of each band, water index
and relationship between spectrums to form a strong water body In this paper, a novel scheme for water body extraction
extraction classifier. The experimental results show that this from Landsat ETM+ imagery using the adaboost algorithm is
method can achieve perfect performance, and is more effective presented. Adaboost is an ensemble strategy, which is used to
than traditional algorithms. boost the classification performance of a weak leaner. It is
based on such consideration that finding many rough rules of
Keywords-water body extraction; remote sensing imagery; thumb can be a lot easier and more effective than finding a
AdaBoost; ETM+ single, highly prediction rule. By combining weak classifiers
which are constructed by spectral information of each band,
I. INTRODUCTION water index and relationship between spectrums, a strong water
body extraction classifier is formed.
Extracting water body information accurately from
remotely sensed imagery has been a hot issue in the object The paper is organized as follows. In Section II, a brief
recognition area. It is significant for surveying, planning and description of adaboost algorithm is provided. Section III
protecting water resources, and particularly flood disaster represents the core of the paper. The realization procedure and
management such as monitoring, evaluation, emergency experiment results of water body extraction from Landsat
response and so on. ETM+ imagery using adaboost algorithm is introduced. Finally,
conclusions are drawn and several issues for future work are
Landsat ETM+ imagery has features of high spatial indicated.
resolution, abundant spectral information and relatively fewer
mixed pixels within a ground resolution cell, which is helpful

Supported by the Key Technologies R&D Program under Grant


No.2008BAC34B03, the Hi-tech Research and Development Program of
China under Grant No.2007AA120205.
II. DESCRIPTION OF ADABOOST ALGORITHM εj = ∑
i: yi ≠ h j
Dt ( xi , yi ) (2)
Adaboost was proposed by Freund and Schapire [5] as an
efficient algorithm of the ensemble learning filed. It combines
iteratively the weak classifiers by taking into account a weight b) Choose the weak classifier ht , with the lowest error
distribution on the training samples such that more weight is εt .
attributed to samples misclassified by the previous iterations. In
this way, adaboost “focuses in” on the informative or c) If ε t ≥ 1/ 2 , then stop.
“difficult” patterns. The final strong classifier takes the form of
a perceptron, a weighted combination of weak classifiers d) If ε t < 1/ 2 , set
followed by a threshold. Fig. 1 shows the algorithm flow chart
for adaboost. 1 1− εt
αt = In( ) (3)
2 εt
Input a training set
e) Update the weights:

Initialize weights
Dt ( xi , yi )e −αt yt ht ( xi )
Dt +1 ( xi , yi ) = (4)
Zt
Train weak
classifiers Z t is simply a normalizing constant computed to insure that
N
Dt ( xi , yi ) represents a true distribution, ∑ D ( x ,y ) = 1 .
i =1
t i

Compute error 4) Output the final classifier:


rate
Update
weights T
H ( x) = sign(∑ α t ht ( x )) (5)
t =1

Termination
condition III. REALIZATION OF WATER BODY EXTRACTION USING
ADABOOST ALGORITHM

A. The Form of Weak Classifiers


Output the
One of the greatest strengths of adaboost algorithm is
final classifier embodied in it avoiding the hard work to find an optimal
sophisicated classifier, but turning to simply combining
Figure 1. The algorithm flow for adaboost different “weak” classifiers, accuracy only slightly better than
chance. As a matter of fact, in real cases we have only limited
The detailed description of adaboost algorithm for the data sets and sometime only few examples are available, thus
binary classification task is as follows. the optimal classifier is difficult to obtain. In this paper, we
adopt the simple linear classifier as the form of weak
1) Establish a training sets: Let the samples and their classifiers, which is expressed as (6):
labels be denoted xi and yi , respectively and let Dt ( xi , yi ) be
the tth distribution over all these traning samples .Given a
training set {( x1 , y1 ), ( x2 , y2 )," , ( xn , yn )} , where yi ∈ {−1,1} . ⎧⎪1 if p j xi , j < p jθ i , j
h ( xi ) = ⎨ (6)
⎪⎩ −1 otherwise
2) Initialize weights:
xi , j is the value of the jth feature of sample xi , p j ∈ {−1,1}
D0 ( xi , yi ) = 1/ N i = 1, 2,..N (1) determines the direction of the inequality sign, θi , j denotes the
threshold value of the jth feature of sample xi .
3) Training Loop: For t = 1, 2,..., T ( T is the maximum
training number):
B. Feature Selection for Constructing Weak Classifiers
a) For each feature j , train a classifier h j which is Extracting water body from ETM+ imagery, spectral
restricted to using a single feature.The error is evaluated with reflectance of seven bands (exclusive of the panchromatic
channel), normalized vegetation index (NDVI), NDWI,
respect to Dt ( xi , yi ) :
MNDWI and water ratio index (WRI) are selected as features 4) WRI
to train the weak classifiers.
Since water body exits special spectral characteristic: TM
1) Spectral reflectance of seven bands band 2 + TM band 3 > TM band 4 + TM band 5, water ratio
index can be expressed as follows:
The seven bands of ETM+ imagery correspond to the
spectral regions of the blue, the green, the red, the near
infrared, the middle infrared, the thermal infrared and another Green + R
middle infrared, respectively. Considering the spectral WRI = (10)
NRI + MRI
reflectance of water, probably the most distinctive
characteristic is the energy absorption at near infrared It is easy to get that the WRI value of Water body is greater
wavelengths and beyond [6]. than 1.
2) NDVI
C. Experimental Results
NDVI is defined as follows [7]:
In order to assess the effectiveness of the proposed
approach, we considered a real data set. Huairou reservoirs in
NRI − R
(7) Beijing City and a part of Bengcuo Lake in Nagchu county of
NDVI =
NRI + R Tibet are selected from the ETM+ imagery for the test, which
were captured in May, 2007 and in December, 2000,
For ETM+ imagery, NRI is TM band 4, and R is TM band 3. respectively. For the convenience of narration, the Huairou
reservoirs imagery is denoted by IMG1 and the Bengcuo Lake
Obviously, NDVI values range from -1 to +1. Because of imagery is IMG2. In addition, it is acknowledged that NDWI is
high reflectance in the near infrared portion of the a perfect characteristic to indicate water, thus, the comparison
electromagnetic spectrum, healthy vegetation is represented by of the extracted water body results by NDWI-based
high NDVI values between 0.1 and 1. Conversely, non- thresholding method and adboost algorithm is conducted.
vegetated surfaces such as water body yield negative values of
NDVI owing to the electromagnetic absorption quality of water. Fig.2 shows some qualitative results for the test imageries.
The background of IMG1 is mainly built-up land, mountain
3) NDWI and MNDWI and vegetation (Fig.2 (a)). It is clear, as shown in Fig.2 (b)- (c),
NDWI is expressed as follows [3]: that the extracted water patches by the NDWI-based
thresholding method are mixed with many built-up land
patches, but the built-up land noise is notably suppressed or
Green − NIR even removed by adaboost algorithm. This is because some
NDWI = (8) built-up land reflects more green light than the near infrared
Green + NIR
light, just as water does. Therefore, its mean value in TM band
For ETM+ imagery, Green is TM band 2. 2 is probably greater than that in TM band 4, resulting in the
positive mean value of the NDWI and low contrast with water.
This index is designed to maximize reflectance of water by Instead of only considering one NDWI feature, the adaboost
using green wavelengths; minimize the low reflectance of the algorithm utilizes the other features such as MNDWI to
near infrared band by water features; and take advantage of eliminate the influence of the building land from the water
the high reflectance of the near infrared band by vegetation body.
and soil features. As a result water features have positive
NDWI values and thus are evenhanded, while vegetation and The background of IMG2 is intentionally selected to be
soil usually get zero or negative values and therefore are dominated by bare mountains, some locations of which are
suppressed. covered by accumulated snow (Fig.2 (d)). Moreover, at the
lower right corner of IMG2, there is obvious mountain shadow.
MNDWI is expressed as follows [4]: Visual inspection from false color composite ETM+ imagery
(RGB: 432) can find that there are great tone discrepancies
between the water body and the accumulated snow. However,
Green − MIR the NDWI value of water body is very close to that of snow.
MNDWI = (9)
Green + MIR Therefore, it is difficult to determine an appropriate NDWI
threshold value to separate them. The result shown by Fig.2 (e)
For ETM+ imagery, MIR is TM band 5. proves this point. One feature is not able to avoid the
interference from other ground objects, but the multiple
The research by Xu has showed that MNDWI is more
features could solve this problem easily. It is clear, as shown
suitable for enhancing and extracting water information for a
in Fig.2 (f), that accumulated snow is accurately classified as
water region with a background dominated by built-up land
non-water body. Especially by adaboost algorithm the
areas.
mountain shadow noise, which is easily confused with water
but usually difficult to be removed if without using
sophisticated procedures, is also reduced remarkably.
(a) (b) (c)

(d) (e) (f)

Figure2. Water body extraction results(the white regions are water in the binary imageries) for NDWI-based thresholding method and adaboost
algorithm.(a)Huairou reservors ETM+ imagery(RGB:432);(d)A part of Bengcuo Lake ETM+imagery(RGB:432); (b) and(e) extracted water body by NDWI-
based thresholding;(c) and (f) extracted water body by adaboost algorithm.

IV. CONCLUSIONS REFERENCES


A water body extraction technique from ETM+ imagery is [1] L. Qian, Remote Sensing Digital Image Processing and Geographical
developed by using adaboost algorithm. The method is Feature Extraction. Beiijng:Science Press, 2004..
essentially an approach of ensemble learning. Various spectral [2] C. Zhou, J. Luo, X. Yang, C.Yang, and Q. Liu, Remote Sensing
Geoscience Understanding and Analysis. Beijing: Science Press, 2001.
features are used to train different weak classifiers, which
[3] S.K. Mcfeeters, “The use of the Normalized Difference Water Index
concentrate on different aspects of the water body extraction (NDWI) in the delineation of open water features,” International Journal
problem but is only required to have accuracy slightly better of Remote Sensing, vol.17, pp. 1425-1432, 1996.
than random chance. The strong classifier, combined by [4] H. Xu, “Modification of normalised difference water index (NDWI) to
weighted sum of weak classifiers, works as the final classifier enhance open water features in remotely sensed imagery,” International
to extract water body. Our results show that adaboost algorithm Joural of Remote Sensing, vol.27, pp. 3025-3033.
compares favorably with the NDWI-based thresholding [5] Y. Freund, “Boosting a weak learning algorithm by majority, ”
method, particularly in the aspect of eliminating the influence Information and Computation, vol.121, pp. 256-285, 1995.
of built-up land, accumulated snow and mountain shadow. [6] T.M. Lillesand, R.W. Kiefer and J. Chipman, Remote Sensing and
Image Interpretation, 4rd ed. New York: John Wiley & Sons, 2001
Furthermore, in order to enhance the performance of the [7] S.N. Goward, B. Markham, D.G. Dye, W. Dulaney, and J. Yang,
proposed method further, taking spatial neighborhood “Normalized difference vegetation index measurements from the
relationship features into account will be worth exploring. adavanced very high resolution radiometer, ” Remote Sensing of
Environment, vol.35, p.257-277, 1991.

Das könnte Ihnen auch gefallen