Multi sensor data fusion for change detection

Scientific Research and Essays Vol. 5 (18), pp. 2823-2831, 18 September, 2010
Available online at http://www.academicjournals.org/SRE
ISSN 1992-2248 ©2010 Academic Journals
Full Length Research Paper
Multi sensor data fusion for change detection
O. Gungor1
* and O. Akar
Department of Geomatics, Karadeniz Technical University, Trabzon, Turkey.
Accepted 6 April, 2010
When performing post-classification comparison using images of different sensors, change detection is
still possible even if images have different resolutions. However, in this case, change pixels are
detected in the pixel size of coarser resolution image. This problem can be solved using higher
resolution aerial photographs or panchromatic images if available. In multi sensor image fusion, images
of higher resolution can be used to increase the resolutions of the multispectral images. In this study
IKONOS panchromatic (1 m) and multispectral bands (4 m), taken in 2003, is fused using à trous image
fusion algorithm to get 1 m multispectral image. Analogously, 1 m resolution aerial photograph
(degraded down from its original 0.5 m resolution) and 30 m resolution ETM+ images, both taken in
2000, are fused to get 1 m resolution ETM+ multispectral image. Once both multispectral image data has
1 m resolution, post-classification comparison method is applied to detect changes occurred on the
coastal zone in city of Trabzon. Preliminary results show that à trous fusion algorithm could keep the
original spectral content of multispectral images in the fused products. Overall classification
accuracies for fused ETM+ and IKONOS images are obtained as 92.40 and 95.2%, respectively. The area
of filled earth on costal zone due to highway construction is detected as 186023 m
2
, in 1 m
2
precision.
Key words: Data fusion, change detection, à trous algorithm, aerial photograph, multispectral image.
INTRODUCTION
The change on coastal areas is very important for
environmental monitoring since the impact of human
activities and natural processes on the coastal
environment is a major concern (Huang and Fu, 2000;
Chen and Rau, 1998). Maps of coastal zones are
produced using different techniques and data such as
ground surveys, aerial images, lidar technology, or
satellite images. Although it offers high accuracy,
mapping coastal resources by ground surveys is costly,
labor intensive and time consuming (Huang and Fu,
2000; Chen and Rau, 1998). Airborne image sensors
provide adequate information about coastal zone, yet
data acquisition and data reduction process also costly
and time consuming at a certain degree (Chen and Rau,
1998). Being state of the art sensor technology, lidar
sensors have some advantages over traditional
photogrammetry and land surveying for topographic
mapping since it offers high accuracy, fast acquisition
and processing time with minimum human dependence.
*Corresponding author. E-mail: ozlemerden@gmail.com
However, especially airborne lidar systems are expensive
and are currently less available.
As a remote sensing sensor, IKONOS satellite is
successfully launched on September 24, 1999 from
Vandenberg Air Force Base, California. It is the world’s
first commercial, high resolution imaging sensor and still
provides 1 m panchromatic images. IKONOS sensor also
has a multispectral sensor which offers four band (blue,
green, red and infrared) multispectral images with 4 m
spatial resolution. This remote sensing sensor changed
the way the public and private industry collect data of
Earth's surface, which is not publicly available before to
farmers, city planners, environmental planners, realtors,
geologists, the media and the others (ITT, 2008).
Subsequently, other recently launched remote sensing
sensors carried by QuickBird, GeoEye and WorldView-2
satellites made possible to map large coastal areas at
regular time intervals with sub-m resolution via their
panchromatic bands.
Many change detection algorithms have been proposed
so far. Multi-date visual composite, image differencing
and post-classification algorithms are used for change
detection problems (Tardie and Congalton, 2004).

2824 Sci. Res. Essays
Aerial Photograph (0.5 m)
ETM+ Multispectral Bands (30 m).
Displayed using bands: 4,3,2
IKONOS Pachromatic Band (1 m)
IKONOS Multispectral Bands (4 m).
Displayed using bands: 4,3,2
Figure 1. Multi Sensor Data Used. Aerial photograph and ETM+
image obtained in 2000, and IKONOS panchromatic and
multispectral images obtained in 2003.
Furthermore, cross-correlation analysis, multi-date
principal components analysis and RGB-NDVI color
composite change detection is added to these methods
(Civco et al., 2002). Additionally, techniques such as
spectral mixture analysis, the Li–Strahler canopy model,
Chi-square transformation, fuzzy sets, artificial neural
networks and integration of multi-source data that have
been used for change detection applications. All these
algorithms give the best result if the remotely sensed
data have the same spatial resolution on each date.
However, it is not always possible to have same spatial
resolution image data of the scene which is taken in
different times. As indicated above, sensors offering high
resolution images become publicly available after end of
1990s. Before this date, remote sensing sensors
provided panchromatic and multispectral images that
have several ms of spatial resolution, e.g. SPOT 4 (10 m
panchromatic and 20 m multispectral bands), ASTER (15
m multispectral bands), IRS-3 (5 m panchromatic and 25
m multispectral bands) and Landsat ETM+ (15 m
panchromatic and 30 m multispectral bands). Hence, if
any possible change on Earth’s surface is intended to be
monitored from about 20 years back till now by post-
classification comparison approach using multispectral
images, then images of coarse resolution sensors need
to be compared with modern higher resolution satellite
images of that scene, and this which will result in
detecting the changes at the spatial resolution of older
images. In other words, if 30 m resolution Landsat ETM+
image and 4 m resolution IKONOS images are used, the
change areas found by comparing the classification
results of these two images can be displayed at 30 m
spatial resolution. This is a drawback which results in
poor resolution display of the change areas and, it also
cause wrong results in terms of the area of change
regions. Aerial photographs can be a solution to the
aforementioned resolution problem. Even more than 20
years before, aerial photographs could have around sub-
m spatial resolution. Hence, an aerial photograph from
past and a recent satellite image that both have sub-m
resolution can be compared to detect any change on the
scene with a sub-m resolution. However, most aerial
photographs, particularly very old ones, offer
monochromatic images, which do not provide spectral
information. However, these aerial photographs can be
considered as panchromatic images and fused with
coarser resolution multispectral ones to transfer the
spatial detail content of the aerial photographs into the
multispectral images. This process will provide
multispectral images having the same spatial resolution
as the aerial photographs.
In this study, the changes occurred on the costal zone
of the city center of Trabzon due to a new major highway
construction initiated along Black Sea cost in 2000 are
evaluated. Two image sets acquired in 2000 and 2003
are used for post-classification comparison (Figure 1).
First image set contains a 30 m resolution ETM+
multispectral image and a 0.5 m resolution aerial
photograph taken in 2000. The second image set
consists of IKONOS panchromatic and multispectral
images (1 and 4 m resolution, respectively) taken in
2003. To tackle the different spatial resolution problem of
both multispectral data, à trous image fusion algorithm is
employed and their spatial resolutions are balanced at 1
m. IKONOS panchromatic band and multispectral bands
are fused to make the spatial resolution of the
multispectral image 1 m. To make the resolution of the
ETM+ data 1 m, the aerial photograph is down sampled
such that it has 1 m spatial resolution. Then ETM+ image
is also degraded down from 30 - 32 m. Then, 1 m
resolution aerial photograph and 32 m resolution ETM+
images are fused to get 1 m resolution ETM+ multispectral
image. Once both multispectral image data has 1 m
resolution, post classification comparison method is
applied to detect changes occurred on the coastal zone
in Trabzon at 1 m spatial resolution.

PRE-PROCESSING METHODS
Four different images of four different sensors, having different
spatial resolutions, are used to compare and detect changes on
coastal zone occurred due to a major highway construction.
Because all images used have different resolutions, they need to be
pre-processed first for a precise post-classification comparison
process. The pre-processing of the input images includes
resampling, registration and image fusion.
Resampling
Post classification comparison approach is applicable if the images
of different dates have the same pixel sizes i.e., same number of
rows and columns. Since the images used in this study have
different spatial resolutions, their pixel size varies. To make the
spatial resolution of all input images 1 m, image fusion technique is
used and most image fusion algorithms need resampling before
fusion process since fusion essentially occurs when the involved
images have the same pixel size. The nearest neighbor method is
the simplest one among resampling methods such as bilinear,
bicubic interpolation, 8-point or 10-point approximation to sinc
function interpolation methods. In the nearest neighbor method, the
new pixel value is assigned as the value of its nearest pixel, i.e., the
original pixel values repeat. However, other aforementioned
methods interpolate new pixel values using the contiguous pixels,
resulting in color content change in the original image (Gungor and
Shan, 2005). Since post-classification comparison method has
been adopted as change detection algorithm, and keeping the
original spectral properties of the multispectral image is crucial for a
robust image classification result, the nearest neighbor method is
used for resampling.
Registration
Change detection studies require that all spatial information be
registered to a common coordinate frame so that each pixel
location on both input images represents the same location on the
ground (Macleod and Congalton, 1998; Kwarteng and Chavez,
1998; Tardie and Congalton, 2004). Image fusion also requires
registration of the input images since different images of the same
area are used together to combine separate images pixel by pixel.
Registration can be done by geo-referencing the images to a map
projection such as UTM (Universal Transverse Mercator). The most
accurate way of doing this is georegistration, in which geometric
relationship between the image and the ground is established using
rigorous photogrammetric methods. See (Lee and Bethel, 2001) for
more information about georegistration. Using polynomial
transforms is another way of registering the input images. In this
case, either ground control points are used to define transformation
between image space and ground or one image (generally, the one
that have better resolution) is taken as the reference and the other
one is registered using this reference image. In this study,
registration is needed for both image fusion and post classification
comparison purposes. Landsat ETM+ image is registered by taking
the aerial photograph, which is degraded down from 0.5 to 1 m
resolution, as the reference for image fusion. Since IKONOS
multispectral and panchromatic images are from the same sensor
and taken at the same time, they are already co-registered and are
directly used for fusion. Then, for accurate post classification
comparison, fused ETM+ is registered by taking the fused IKONOS
multispectral image as the reference. During registration refinement
of fused multispectral images, common points on both images are
selected carefully such that they produce an RMS error smaller
than 0.5 pixels with respect to the pixel size of the fused
multispectral images.
Gungor and Akar 2825
Image fusion
Image fusion methods are used to enhance the spatial resolution of
the multispectral images with the help of higher spatial resolution
panchromatic ones. A successful image fusion algorithm does not
distort the spectral content of the multispectral image while
increasing its spatial resolution. Gungor (2008) states that image
fusion algorithms using wavelet-transform are successful in terms of
keeping the color quality of the original multispectral images;
however, their performances are not appealing as the results
obtained from color-based (e.g., IHS and Brovey method) and
statistical methods (e.g., PCA method) in terms of improving the
spatial detail content. On the other hand, color-based and statistical
methods are not successful as the wavelet-transform based
methods in terms of color quality. This study uses multispectral
images for image classification to detect the changes occurred in
the scene. Therefore, keeping the color quality of the original
multispectral images is crucial for precise change detection. For this
reason, the à trous wavelet transform method, one of the most
successful representative of the wavelet-transform based fusion
methods, is chosen as the fusion method since it is fast and easy to
implement, and reliable in terms of good color quality.
The à trous algorithm is a non-orthogonal, dyadic, undecimated,
DWT algorithm. It uses a low-pass filter associated to a scale
function )
(t
φ to obtain successive approximations of the original
image and a high-pass filter to get three detail images (Aiazzi et al.,
2002). Since decimation is not applied, transformed images have
the same size as the original input images at every decomposition
level.
Starck and Murtagh (1994) proposes a practical implementation
of the à trous wavelet transform algorithm by using the following 2-
D 5 x 5 convolution mask 0
H .
=
1
4
6
4
1
4
16
24
16
4
6
24
36
24
6
4
16
24
16
4
1
4
6
4
1
256
1
0
H
To calculate the approximation image that contains the low
frequency part of the original image, 0
H is up sampled before
each decomposition level j by inserting 2j-1
-1 zeros between filter
coefficients. This methodology keeps the size of the transformed
images same as the original input images at every decomposition
level since decimation is not applied (Wegner et al., 2006). The
detail image w , which contains the high frequency part of the
original image, is obtained as the difference of two successive
approximation images. Image fusion using à trous algorithm is
performed as follows:
(1) Each multispectral band is resampled to the same size of the
panchromatic image.
(2) Panchromatic image is decomposed until its resolution equals to
the resolution of original multispectral bands.
(3) Summation of { j
w } is calculated as
=
=
J
j
j
w
W
1
to detect the
spatial detail difference between panchromatic and multispectral
image, where J is the total number of decomposition levels applied
to the original panchromatic image.
(4) W is added to each original multispectral band to construct
fused bands. J is determined with respect to the resolutions of the
panchromatic and the multispectral images. For example, for

Original ETM+ Multispectral
(Bands: 4,3,2)
Fused ETM+ Multispectral
(Bands: 4,3,2)
Original IKONOS Multispectral
(Bands: 4,3,2)
Fused IKONOS Multispectral
(Bands: 4,3,2)
Figure 2. Fusion results of IKONOS and ETM+ multispectral images.
IKONOS panchromatic and multispectral images, two times wavelet
decomposition to the panchromatic image is enough to make its
resolution the same as the multispectral image since the resolution
of the multispectral image is four times coarser than the
panchromatic image.
First, 4 m resolution IKONOS multispectral image is fused with its
panchromatic band to get 1 m resolution IKONOS multispectral
image. Secondly, the spatial resolution of aerial photograph is
degraded down from 0.5 to 1 m using bi-cubic resampling method.
The reason of using bi-cubic resampling method is the intention of
minimum distortion to the spatial structure of the aerial photograph
while decreasing its resolution. Analogously, ETM+ image is
degraded down from 30 m to 32 m using nearest neighbor
resampling method. The reason of using nearest neighbour method
is the intention of not disturbing the spectral structure of ETM+
image to get good fusion result. Finally, 1 m resolution aerial
photograph is fused with 32 m resolution ETM+ image to get 1 m
resolution ETM+ multispectral image. Hence, spatial resolutions of
two multispectral images, which will be used for image
classification, made identical to be able to detect and display
regions of change in 1 m spatial resolution. The original and fused
IKONOS and ETM+ multispectral images are displayed in Figure 2.
DETECTING REGIONS OF CHANGE
Post-classification comparison method is used as the change
detection algorithm, in which two images from different dates are
separately classified and compared pixel by pixel to create a new
change image (Tardie and Congalton, 2004; Mas, 1999). In this
method, detected areas of change is displayed as a matrix
containing “from X class to Y class” information, where X and Y
indicates land cover classes on the scene. The advantage of post-
classification comparison method is that the atmospheric effects,
sensor and environmental differences between multi-temporal
images are minimized and a complete matrix of change information
is obtained as input images are classified independently (Lu et al.,
2004).
Accurate classification is very important for reliable change
detection results as the final accuracy depends on the quality of the
classified image of each date (Lu et al., 2004).Therefore, choosing
the best classification algorithm, depending on the nature of the
data, is crucial. Since the main motivation in this study is to detect
the extent and the area of filled earth on costal zone, it is vital to
perform precise classification such that sea and land classes are
classified correctly. Especially in IKONOS multispectral image,
wave patterns which causing shadow in the sea and sediments at
the river mouth could result in misclassification of pixels as shadow
and land in Black Sea, and misclassification affects the post
classification comparison results negatively. Thus, input images are
classified using Maximum Likelihood (ML), SAM (Spectral Angular
Mapping), Fisher Linear Likelihood (FLL), and ECHO classifiers and
best result is obtained from ECHO classifier. ECHO is a spatial-
spectral classifier that incorporates not only spectral variations but
spatial ones as well into the decision-making process. It is a
classifier that first segments the scene into spectrally homogenous
objects. Then, it performs a sample classification algorithm based
upon a maximum likelihood object classification scheme in which

Figure 3. Flowchart of entire change detection approach.
each object delineated would be classified based on the statistical
properties of the pixels of the object (Landgrebe, 2003).
Once fused ETM+ and IKONOS images are classified separately,
then each pixel on both classified images are compared to create a
change image at 1 m spatial resolution. The entire process from the
beginning until change image is obtained is given in Figure 3.
RESULTS AND DISCUSSION
Once pre-processing methods are applied to all input
images, preliminary classifications are performed on the
fused ETM+ and IKONOS multispectral images using
MultiSpec software created in Purdue University to carry
out a post-classification analysis procedure.
In IKONOS image, 7 classes (Water, Vegetation,
Shadow, Bare Soil, Urban Fabric Type_1, Urban Fabric
Type_2, and Urban Fabric Type_3) are determined to
classify the scene. Especially in urban area there are so
many different features which have different spectral
reflectance values. As a matter of fact, the same type of
feature such as building also may have different
reflectance values due to the structure of the roof
materials. For this reason, three different classes for
urban fabric are determined, based on not their geometric
structure but their colors. For example asphalt roads,
concrete structures and some type of buildings are
considered as urban Fabric Type_1. Since the purpose of
the study is to distinguish the land from sea, this
methodology of grouping and defining classes within land
is reasonable and does not affect the final goal of the
study. After final classification, all three urban related
classes (Urban Fabric Type_1, 2 and 3) are merged as a
single Urban Fabric class.
As mentioned above, preliminary classification result of
fused IKONOS image shows that some pixels in the
Black Sea are misclassified such that shadows resulting
from wave patterns are classified as Urban Fabric (See
red ellipse in Figure 4c). Same problem also occurred
within land. Some pixels belong to Shadow class are
misclassified as Sea because of the spectral

Figure 4. The preliminary and final classification results of fused ETM+ and IKONOS
multispectral images.
resemblance of shadows in the land and the ones in the
sea. This problem prevents accurate detection of change
since these areas are considered as change (from land to
sea or vise versa), although visual evaluation of input
images contradicts and reveals no change. To reduce the
number of misclassified pixels, majority filter is applied to

Black Sea
Land
Figure 5. Change image showing the differences between 2000 and 2003.
Table 1. Accuracy assessment of fused ETM+ image classification result.
Class name
Reference
totals
Classified
totals
Number
correct
Producers
accuracy (%)
Users
Accuracy (%)
Sea 97 94 94 96.91 100.00
Sea2 15 18 15 100.00 83.33
River 4 5 4 100.00 80.00
Vegetation_1 7 9 7 100.00 77.78
Vegetation_2 26 27 24 92.31 88.89
Shadow 9 6 6 66.67 100.00
Urban Fabric Type_1 76 68 67 88.16 98.53
Bare soil 6 6 5 83.33 83.33
Totals 250 250 231
Overall Classification Accuracy 92.40%
Overall Kappa Statistics 0.8987
the classified images. This approach substantially
corrected the misclassification and produced more
accurate results. Although all misclassified pixels, those
classified as Urban Fabric in the sea, are corrected and
converted to Sea (except some ships, encircled by a
circle in Figure 4d), there are still some misclassified
pixels, those classified as Sea in the land. To solve this
problem, a Matlab code is created which segments the
classified image and detects the polygons for Sea class
in entire image. Then it converts the misclassified pixels
from Sea to Urban Fabric, except the ones actually
belong to the real Sea class for a more accurate post
classification comparison.
Analogously, the same classification procedure is
performed for the fused ETM+ image; however, 9 classes
are created (Sea, Sea_Type 2, River, Vegitation_Type 1,
Vegitation_Type 2, Shadow, Urban Fabric Type_1, Urban
Fabric Type_2, and Bare Soil). Similarly, two different
classes for urban fabric are determined based on not
their geometric structure but their colors. Additionally,
after preliminary classification a new class is detected
right in the intersection of land and the Sea. This class is
also present in the main land but they are very rare. The
pixels belong this class are dark blue and encircled by a
red ellipse in Figure 4a. For this reason, these pixels are
considered as a separate class Sea_Type2. A similar
problem is also encountered for the pixels belong to
River. These pixels, pale blue ones showed with a red
circle in Figure 4a, are also exist in Sea. For this reason,
three different classes for water are determined as Sea,
Sea_Type2, and River. After final classification, all three
classes are merged as one single Sea class (Figure 4b).
Misclassified pixels, those classified as Sea within main
land, are corrected using the same Matlab Code created
for this purpose. Merge process is also done for two
vegetation classes to produce one final Vegetation class.
The preliminary and final classification results of fused
ETM+ are given in Figure 4a and b. The preliminary and

Table 2. Accuracy assessment of fused IKONOS image classification result.
Class name Reference
totals
Classified
totals
Number
correct
Producers
accuracy (%)
Users
Accuracy (%)
Sea 125 125 125 100.00 100.00
Vegetation 19 20 19 100.00 95.00
Shadow 7 7 7 100.00 100.00
Bare soil 31 37 29 93.55 78.38
Totals 250 250 238
Overall Classification Accuracy 95.20%
Overall Kappa Statistics 0.9311
final classification results of fused IKONOS multispectral
images are given in Figure 4c and d. The preliminary and
final classification results of fused ETM+ and IKONOS
multispectral images are given in Figure 4.
In Erdas Imagine software 250 random points are
generated for the accuracy assessment of the
classification results. Hence, overall classification
accuracies for fused ETM+ and fused IKONOS images
are obtained as 92.40 and 95.2%, and Kappa statistics
are 0.8987 and 0.9311, respectively as given in Tables 1
and 2. Accuracy assessment procedure is performed for
preliminary classification results of fused ETM+ and
IKONOS images (images A and C in Figure 4), not for the
final classification results since majority filter and the
created Matlab code are applied to them.
Finally all classes in main land in both classification
results of 2000 and 2003 merged and created one single
Land class. This procedure produced only two classes
(Land and Sea) in the final results. Hence these images
are used for post-classification comparison procedure in
ERDAS IMAGINE software. As the output, the change
image showing the spatial alteration of Land and Sea is
generated. The change image is illustrated in Figure 5.
The pixels classified as Sea in fused ETM+ and Land in
IKONOS image are shown in white in Figure 5 (Large
blue ellipses). Hence, white pixels designate the filled
earth and the area turned from Sea to Land during
construction between 2000 and 2003. The total area of
such region is computed as 186023 m
2
. Correspondingly,
the area turned from Land to Sea between 2000 and
2003 is depicted in black. The total area of such change
region is computed as 30542 m
2
. The linear black and
white objects in Black Sea, which are showed in two
small blue ellipses, are ships approaching to the port.
These ships are detected as change (from Sea to Land)
in the change image since one ship is not exist in the
other image. The results shows that the extent and
amount of earth fill is detected and displayed at 1 m
spatial resolution. Such high level of spatial accuracy
would not be reached if 30 m Landsat ETM+ image is
directly used for change detection.
Conclusion
In this study, the change on costal zone of Trabzon is
detected using image data of four different sensors,
which offer different spatial resolutions. As expected, the
amount and extent of earth fill on costal zone due to new
high way construction is detected successfully at 1 m
spatial resolution. It is also showed that the spatial
information content of a monochromatic aerial
photograph can be used to improve a coarse resolution
of a multispectral satellite image. This could be done with
the aid of à trous image fusion algorithm. This method
successfully enhanced the spatial resolution of 30 m
ETM+ image to 1 m using the aerial photograph, while
keeping the original spectral content of ETM+ image.
92.40% and 95.2% overall classification accuracies of
fused IKONOS and fused ETM+ images are the evidence
of preserving the color content after fusion.
Classifiers such as ML, SAM, and FLL misclassify
wave patterns in IKONOS image as Urban Fabric. This
problem may be misinterpreted by post-classification
comparison and perceived as a change from Sea to
Land. The amount of misclassified pixels is reduced
using ECHO classifier, which combines the spectral
information with spatial information. However, it is
understood that only classification algorithms do not
solve misclassification problem. For this reason, majority
filter and created Matlab code are applied to preliminary
classification results to get final classified images free
from misclassified pixels both in Land and Sea.
REFERENCES
Aiazzi B, Alparone L, Baronti S, Garzelli A (2002). Context-Driven
Fusion of High Spatial and Spectral Resolution Images Based on
Oversampled Multiresolution Analysis. IEEE Trans. on Geosci.
Remote Sens., 40(10): 2300-2312.

Chen LC, Rau JY (1998). Detection of shoreline changes for tideland
areas using multi-temporal satellite images. Int. J. Remote Sens.,
19(17): 3383-3397.
Civco DL, Hurd JD, Wilson EH, Song M, Zhang Z (2002). A Comparison
of Land Use and Land Cover Change Detection Methods. ASPRS-
ACSM Annual Conference and FIG XXII Congress.
Gungor O (2008). Multi Sensor Multi Resolution Image Fusion. PhD
Dissertation, Purdue University, Indiana, USA.
Gungor O, Shan J (2005). A statistical approach for multi-resolution
image fusion. UPPECORA 16UP - Global Priorities in Land Remote
Sensing, held at Sioux Falls Convention Center, Sioux Falls, SD,
USA.
Huang W, Fu B (2000). Remote Sensing for Coastal Area Management
in China. Coastal Manage., 30: 271-276.
ITT (2008). ITT Space System Division.
<http://www.ssd.itt.com/heritage/ikonos.shtml> Last Accessed
September 01, 2008
Kwarteng AY, Chavez PS (1998). Change detection study of Kuwait
City and environs using multitemporal Landsat Thematic Mapper
data. Int. J. Remote Sens., 19(9): 1651-1662.
Landgrebe DA (2003). Signal Theory Methods in Multispectral Remote
Sensing, John Wiley and Sons, Hoboken, New Jersey, USA.
Lee C, Bethel J (2001). Georegistration of Airborne Hyperspectral
Image Data. IEEE Trans. Geosci. Remote Sens., 39(7): 1347-1351.
Lu D, Mausel P, Brondizio E, Moran E (2004). Change detection
techniques. Int. J. Remote Sens., 25(12): 2365-2407.
Macleod RD, Congalton RG (1998). A Quantitative Comparison of
Change-Detection Algorithms for Monitoring Eelgrass from Remotely
Sensed Data. Photogrammetric Eng. Remote Sens., 64(3): 207-216.
Mas JF (1999). Monitoring land-cover changes: a comparison of change
detection techniques. Int. J. Remote Sens., 20(1): 139-152.
Starck JL, Murtagh F (1994). Image Restoration with Noise Suppression
Using the Wavelet Transform. Astron. Astrophys, 288: 342-348.
Tardie PS, Congalton RG (2004). A Change-detection Analysis: Using
Remotely Sensed Data to Assess the Progression of Development in
Essex County, Massachusetts from 1990 to 2001. http://www.aars-
acrs.org/acrs/proceeding/ACRS2008/Papers/PS%203.21.pdf.
Wegner FV, Both M, Fink RHA (2006). Automated Detection of
Elementary Calcium Release Events Using the A Trous Wavelet
Transform. Biophys. J., 90: 2151-2163.

Multi sensor data fusion for change detection

More Related Content

What's hot

Viewers also liked

Similar to Multi sensor data fusion for change detection

Recently uploaded

Multi sensor data fusion for change detection