pan‐sharpening of very high resolution multispectral images using genetic algorithms
TRANSCRIPT
This article was downloaded by: [University of Miami]On: 24 September 2013, At: 01:26Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of RemoteSensingPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tres20
PAN‐sharpening of very high resolutionmultispectral images using geneticalgorithmsA. Garzelli a & F. Nencini aa Department of Information Engineering, University of Siena, ViaRoma 56, 53100 Siena, ItalyPublished online: 22 Feb 2007.
To cite this article: A. Garzelli & F. Nencini (2006) PAN‐sharpening of very high resolutionmultispectral images using genetic algorithms, International Journal of Remote Sensing, 27:15,3273-3292, DOI: 10.1080/01431160600554991
To link to this article: http://dx.doi.org/10.1080/01431160600554991
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.
This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
PAN-sharpening of very high resolution multispectral images usinggenetic algorithms
A. GARZELLI* and F. NENCINI
Department of Information Engineering, University of Siena, Via Roma 56, 53100 Siena,
Italy
(Received 28 July 2005; in final form 22 December 2005 )
A novel image fusion method is presented, suitable for sharpening of
multispectral (MS) images by means of a panchromatic (PAN) observation.
The method is based on redundant multiresolution analysis (MRA); the MS
bands expanded to the finer scale of the PAN band are sharpened by adding the
spatial details from the MRA representation of the PAN data. As a direct,
unconditioned injection of PAN details gives unsatisfactory results, a new
injection model is proposed that provides the optimum injection by maximizing a
global quality index of the fused product. To this aim, a real-valued genetic
algorithm (GA) has been defined and tested on Quickbird data. The optimum
GA injection is driven by an index function capable of measuring different types
of possible distortions in the fused images. Fusion tests are carried out on
spatially degraded data to objectively compare the proposed scheme to the most
promising state-of-the-art image fusion methods, and on full-resolution image
data to visually assess the performance of the proposed genetic image fusion
method.
1. Introduction
Spaceborne imaging sensors allow a global coverage of the Earth surface to be
achieved on a routine basis. Multispectral (MS) observations, however, exhibit
ground resolutions that may be inadequate to specific identification tasks, especially
where urban areas are concerned. Following the successful launch of the new
generation of satellite imagers, Ikonos, QuickBird and SPOT-5, very high-resolutionMS and panchromatic (PAN) images are now available.
Data fusion techniques, originally devised to allow integration of different
information sources, may take advantage of the complementary spatial/spectral
resolution characteristics for producing spatially enhanced MS observations. This
specific aspect of data fusion is often referred to as data merge (Scheunders and
Backer 2001) or band-sharpening (Kumar et al. 2000). More specifically, PAN-
sharpened MS is a fusion product in which the MS bands are sharpened through the
higher-resolution PAN image. The latter is acquired with the maximum resolution
allowed by the imaging sensor, as well as by the datalink throughput, while theformer are acquired with coarser resolutions, typically two to four times lower,
because of signal-to-noise ratio (SNR) constraints and transmission bottlenecks.
After being received at ground stations, the PAN images may be merged with the
MS data to enhance their spatial resolution.
*Corresponding author. Email: [email protected]
International Journal of Remote Sensing
Vol. 27, No. 15, 10 August 2006, 3273–3292
International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online # 2006 Taylor & Francis
http://www.tandf.co.uk/journalsDOI: 10.1080/01431160600554991
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
Data merge methods, based on injecting high-frequency components taken from
the PAN image into resampled versions of the MS data, have demonstrated superior
performances (Wald et al. 1997, Schowengerdt 1980). The pioneering method of
high-pass filtering (HPF; Schowengerdt 1980) consists of an addition of spatial
details, taken from a high-resolution PAN observation, into a bicubically resampled
version of the low-resolution MS image. Such details are obtained by taking the
difference between the PAN image and its lowpass version achieved through a
simple local pixel averaging, that is a box filtering. Later efforts benefit from
multiresolution analysis (MRA), which provides effective tools, such as wavelets
and Laplacian pyramids, to help carry out data fusion/merge tasks (Ranchin and
Wald 2000, Aiazzi et al. 2002, Garzelli and Nencini 2005).
Redundant multiresolution structures, such as the generalized Laplacian pyramid
(GLP; Aiazzi et al. 2002) matching even fractional scale ratios between the images to
be merged, the undecimated discrete wavelet transform (UDWT), and the ‘a-trous’
wavelet transform (ATWT), have been found to be particularly suitable for image
fusion because of their translation-invariance property (not strictly possessed by the
GLP). As all these decompositions are not crucially subsampled, injection artefacts
and canvas-like patterns originated by aliasing are avoided.
Fusion schemes based on the ATWT have been proposed recently (Nunez et al.
1999, Garzelli et al. 2000, Chibani and Houacine 2002, Garzelli and Nencini 2005)
and have been successfully used within the ARSIS (Amelioration de la Resolution
Spatiale par Injection de Structures) concept (Ranchin et al. 2003). Data merge
based on multiresolution analysis, however, requires the definition of a model
establishing how the missing highpass information to be injected into the resampled
MS bands is extracted from the PAN image. Wavelet-based techniques that do not
consider any injection model (Li et al. 2002) may produce unsatisfactory results in
terms of spectral preservation of the fused product. This means that the spatial
information of the PAN data has to be opportunely weighted and equalized before
being injected onto each band of the MS data. This is the case with Spectral
Distortion Minimizing (SDM; Alparone et al. 2003), Context Based Decision (CBD;
Aiazzi et al. 2002) and Ranchin, Wald, Mangolini (RWM; Ranchin et al. 2003)
techniques, which use space-varying models. These fusion algorithms produce good
results but are typically effected by numerical instabilities. In particular, the SDM
algorithm implements an easy and efficient model of injection but suffers the
drawbacks of instability and data-dependent results.
To overcome these problems, in this paper we propose an injection model PAN-
sharpening by genetic algorithm (GA), in which the coefficients that equalize the
PAN image before details are injected into the MS image are derived globally – one
for each band – from coarser scales, similar to previous schemes such as SDM, CBD
and RWM, but not a priori defined on image local statistics (e.g. variance, mean,
correlation coefficient). A GA is applied to determine the gains that maximize an
image quality score index, namely the Q4 quality index (Alparone et al. 2004), which
has proved to be particularly efficient for measuring radiometric and spectral
distortions on four-band multispectral images. The ability to find quickly and
efficiently an optimum solution is verified by testing the GA on Quickbird images,
and both visual and objective experimental comparisons with advanced fusion
methods are also reported.
The remainder of the paper is organized as follows. Section 2 deals with the image
analysis tools used in the proposed PAN-sharpening algorithm with emphasis on the
3274 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
ATWT. In section 3 a general GA scheme is described, based on the floating point
representation of chromosomes and on genetic operators borrowed from the most
promising studies on GAs. How the GA is used to regulate the PAN injection is
explained in section 4. Experimental results and comparisons are presented and
discussed in section 5 on Quickbird data. Conclusions are drawn in section 6.
2. Image analysis
2.1 Wavelet ‘a-trous’ transform
The octave multiresolution analysis introduced by Mallat (1999) for digital images
does not preserve the translation invariance property. In other words, a translation
of the original signal does not necessarily imply a translation of the corresponding
wavelet coefficient. This property is essential in image processing. On the contrary,
wavelet coefficients generated by an image discontinuity could disappear arbitrarily.
This non-stationarity in the representation is a direct consequence of the down-
sampling operation following each filtering stage.
To preserve the translation invariance property, the down-sampling operation is
suppressed, but filters are up-sampled by 2j, that is dilated by inserting 2j21 zeros
between any couple of consecutive coefficients. An interesting property of the
undecimated domain (Aiazzi et al. 2002) is that at the jth decomposition level, the
sequences of approximation, cj(k, m), and detail, dj(k, m), coefficients are
straightforwardly obtained by filtering the original signal through a bank of
equivalent filters, given by the convolution of recursively up-sampled versions of the
lowpass filter h and the highpass filter g of the analysis bank:
h�j ~ 6
j{1
m~0h : 2mð Þ ð1Þ
g�j ~ 6
j{2
m~0h : 2mð Þ
� �6 g : 2j{1� �
~h�j{16 g : 2j{1� �
ð2Þ
The ATWT (Dutilleux 1989) is an undecimated non-orthogonal multiresolution
decomposition defined by a filter bank {hi} and {gi5di2hi}, with the Kronecker
operator di denoting an allpass filter. In the absence of decimation, the lowpass filter
is up-sampled by 2j, before processing the jth level; hence the name ‘a-trous’, which
means ‘with holes’. In two dimensions, the filter bank becomes {hihj} and
{didj2hihj}, which means that the 2-dimensional (2D) detail signal is given by the
pixel difference between two successive approximations, which have all the same
scale 20, i.e. 1. The jth level of the ATWT, j50, …, J21, is obtained by filtering the
original image with a separable 2D version of the jth equivalent filter, as in
equation (1).
For a J-level decomposition, the ‘a-trous’ wavelet accommodates a number of
coefficients J + 1 times greater than the number of pixels. Because of the absence of
decimation, as well as the zerophase and 26 dB amplitude cutoff of the filter, the
synthesis is simply obtained by summing all detail levels to the approximation:
exx k, mð Þ~XJ{1
j~0
dj k, mð ÞzcJ k, mð Þ ð3Þ
in which cJ(k, m) and dj(k, m), j50, …, J21, are obtained through 2D separable
PAN-sharpening by genetic algorithm 3275
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
linear convolution with h�J and g�j , j50, …, J21, equations (1) and (2), respectively.
Equivalently, they can be calculated by means of a tree-split algorithm, that is by
taking pixel differences between convolutions of the original signal with
progressively up-sampled versions of the lowpass filter.
The number of decomposition levels J, which directly defines the equivalent
lowpass filter hJ, given a half-band prototype filter {hi}, rules the smoothness of the
lowpass approximation.
2.2 Structure of the PAN-sharpening model
The definition of a suitable model for the injection of PAN details is important for
the good quality of the data fusion product. The most promising methods have
developed different models, generally related to a common approach that consists of
calculating the parameters that regulate the injection (typically gain and offset) at a
coarser resolution and then adopting those parameters to the finer resolution. This
means that the scale persistence is exploited, assuming that the characteristics of
edges and texture at coarser scales are not too different from those at finer scales.
This hypothesis is verified if the ratio between the spatial resolutions of MS and
PAN data is not too high and if the model is opportunely defined. Some successful
techniques, namely SDM, CBD and RWM, have developed interesting models in
which the PAN injection is regulated by the ratio between the standard deviations of
the MS and PAN data, or by the first moment of inertia (Wald 2002). Furthermore,
the injection model can be local, that is contextualized to neighbourhood pixels, or
global, derived from analysis of the whole acquired area.
The problem that generally occurs when applying those fusion techniques is that
the definition of the model parameters does not correspond to an optimum choice in
terms of geometric, radiometric and spectral distortions of the fused product. In
addition, the definition of a local model often gives rise to numerical instability and
unsatisfactory visual quality throughout the image.
The proposed model is a simple linear model in which the unknown global gain
parameters gl, l51, 2, …, N, with N denoting the number of MS bands, are not a
priori defined or computed from local statistics, but are calculated by maximizing an
appropriate function at coarser scales (i.e. the Q4 index), which will be recalled in
the following section. In synthesis, the proposed model defines how the spatial
details of the fused MS image edd MSl jð Þ (at scales j50, 1 in the case of the 1 : 4 scale ratio
between PAN and MS images) are obtained from the PAN spatial details
d PANl jð Þ m, nð Þ at the same scales:
eddMSl jð Þ m, nð Þ~gl dPAN
l jð Þ m, nð Þ{dPAN
l jð Þ
� �zdMS
l jð Þ m, nð Þ j~0, 1 ð4Þ
where the index l is related to the band being analysed, :ð Þ is the mean operator, and
the dMSl jð Þ m, nð Þ coefficients of the expanded MS image (which are near to zero) are
also added to avoid possible impairments due to the ATW filtering operations and
to preserve accurately the mean value of the original MS image.
The task of the GA is to find the best combination of the real-valued coefficients
gl, l51, 2, …, N, according to an objective criterion that describes the enhancement
of the MS images. The representation of the chromosomes, g, is therefore a string of
3276 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
real numbers as reported in equation (5).
Chromosome g : g1, g2, g3, :, :, :, gNf g ð5Þ
2.3 Quality evaluation criterion
The image quality index Q4 for multispectral images having four spectral bands can
be calculated on PAN-sharpened MS images as described in Alparone et al. (2004).
The index Q4 is derived from the theory of hypercomplex numbers, in particular of
‘quaternions’, which can be represented in the form a5a1 + a2i + a3j + a4k, where a1,
a2, a3, a4 are real numbers, and i25j25k25i j k521. For MS images with four
spectral bands, typical for new generation satellite images, a1, a2, a3, a4 represent the
values assumed by a given image pixel in the four bands, acquired in the blue, green,
red and near-infrared (NIR) wavelengths. The quality index is a generalization of
the Q index defined in Wang and Bovik (2002) for an original image signal x and a
test image signal y, which can be stated as
Q~4 cov x, yð Þxy
var xð Þzvar yð Þð Þ xð Þ2z yð Þ2h i : ð6Þ
and may be equivalently rewritten as
QN|N~cov x, yð Þ
sxsy
2xy
xð Þ2z yð Þ22sxsy
s2xzs2
y
ð7Þ
where sf denotes the standard deviation of f, and cov(x, y) is the cross-covariance of
x and y, all computed over a given N6N block. In practice, the first factor is the
correlation coefficient (CC), the second factor (always,1 and51 iff x5y) accounts
for the mean bias; analogously, the third factor measures the change in contrast.
Eventually, the quality index Q of y is obtained by averaging the values obtained
starting from all the N6N blocks of the images x and y. This quality factor can be
applied only to monochrome images.
The unique score index Q4 for four-band MS images, which assumes a real value
in the interval [0, 1], is 1 iff the MS image is identical to the reference image. Again,
Q4 is made up of different components (factors) to take into account the correlation,
the mean of each spectral band, the intraband local variance, and the spectral angle.
The first three factors are also taken into account by Q for each band while the
spectral angle is introduced by Q4 by properly defining a CC of multivariate data. In
this way, both radiometric and spectral distortions are considered by a single
parameter. Q4 can be computed from
Q4N|N~4 E x:y�½ �{x:y�½ �
E xk k2h i
{ xk k2zE yk k2
h i{ yk k2
: xk k: yk kxk k2
z yk k2ð8Þ
where quaternions are indicated in bold case (e.g. x5{xr + i x1 + j x2 + k x3}), the
product of x by the (hyper)complex conjugate of y, namely y*, has to be intended as
a product of quaternions, E[?] denotes the quaternion obtained by averaging the
pixel quaternions within an N6N block and ||x||5!(x2r + x2
1 + x22 + x2
3) is the
magnitude of quaternion x. Finally Q4 is obtained by averaging the magnitudes of
PAN-sharpening by genetic algorithm 3277
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
all Q4N6N over the whole image, i.e.
Q4~E Q4N|Nk k½ � ð9Þ
The more Q4 approaches unity, the higher is the radiometric and spectral quality of
the fused image. This suggests that this index can be used not only to evaluate the
performances of fusion algorithms but also as a target function to be maximized to
compute optimal fusion parameters.
Determining the best solution analytically is a difficult task, particularly if we are
looking for a global solution. A genetic approach designed for optimal parameter
computation can solve this problem in a powerful and efficient way.
3. Genetic components
GAs (Davis 1991, Michalewicz 1994) are inspired by the evolution of populations.
In a particular environment, individuals who fit the environment better will be able
to survive and hand down chromosomes to their descendants, while less fit
individuals will become extinct. The aim of GAs is to use simple representations to
encode complex structures and simple operations to improve these structures.
Therefore, GAs are characterized by their representations and operators. A fitness
function is defined that measures the fitness of each individual. The populations are
evolved to find good individuals as measured by the fitness function. GAs have been
used to solve linear and nonlinear problems (Michalewicz 1994, Ghoshray and Yen
1995, Chaiyaratana and Zalzala 1997, Vasconcelos et al. 2001) and in several remote
sensing applications, such as classification (Tso and Mather 1999, Pal and
Bandyopadhyay 2001, Pal et al. 2001, Maulik and Bandyopadhyay 2003, Mertens
et al. 2003), image processing (Liu and Tang 1998, Caorsi et al. 2000, Munteanu
and Rosa 2000, Ho and Lee 2001), feature selection (Lin and Sarabandi 1999, Jin
and Wang 2001, Jeon et al. 2002, Chen 2003, Yao and Tian 2003, Zhan et al. 2003)
and electromagnetic modelling (Anyong et al. 2001). The ability to explore all
regions of the state space of chromosomes, which constitute the population, using
genetic operators (mutation, crossover) and reproduction have led to increased
attention in recent studies. A GA flow diagram is shown in figure 1, and each of the
major components is discussed in the following sections. A GA requires the
definition of these fundamental steps: chromosome representation, selection of a
function, also called the fitness function, creation of the initial population,
reproduction function, mutation and crossover operators, termination criteria,
and evaluation of the fitness function. The following subsections describe these
issues.
3.1 Chromosome representation
A chromosome representation is necessary to describe each individual in the GA
population. The representation scheme determines how the problem is structured in
the GA and also determines the genetic operators that are used. Each chromosome
is made up of a sequence of genes from a predefined alphabet. Binary digits (0 and
1), floating point numbers, integers, symbols (i.e. A, B, C, D), matrices, etc., can be
used as the alphabet in the genes’ representation. In the original design, the alphabet
was limited to binary digits, but it has since been shown that natural representations
are more efficient and more accurate (Michalewicz 1994). One useful representation
of a chromosome for function optimization involves genes from an alphabet of
3278 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
floating point numbers with values limited by an upper and a lower bound.
Michalewicz (1994) showed that a real-valued GA is more efficient in terms of CPU
time and more accurate in terms of precisions for replications than binary GA
representations.
3.2 Reproduction
An important role in GAs is the selection of individuals to produce successive
generations, usually called reproduction. A probabilistic selection is performed
based on the individual’s fitness, such that the better individuals have an increased
chance of being selected. An individual in the population can be selected more thanonce, with all individuals in the population having a chance of being selected to
reproduce into the next generation. Several schemes are available for the selection
process: roulette wheel selection, scaling techniques, tournament, elitist models and
ranking methods (Holland 1975, Goldeberg 1989, Michalewicz 1994).
A common selection approach assigns a probability of selection, Pj, to each
individual j based on its fitness value. A series of N random numbers uniformly
Figure 1. Flow diagram of GA.
PAN-sharpening by genetic algorithm 3279
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
distributed between 0 and 1, U(0, 1), is generated and compared against the
cumulative probability, Ci~Pi
j~1 Pj, of the population. If the random number
linked to the appropriate individual i is limited by Ci21 and Ci, the i-individual is
copied into the new population. Various methods exist to assign probabilities to
individuals: roulette wheel, linear ranking and geometric ranking.
Ranking methods, which produce the best performances, require the evaluation
function to map the solutions to a partially ordered set and assign Pi based on the
rank of solution i when all solutions are sorted. Normalized geometric ranking
(Joines and Houck 1994) defines Pi for each individual by:
Pi~q 1{qð Þr{1
1{ 1{qð ÞPopSizeð10Þ
where q is the probability of selecting the best individual, PopSize is the overall
number of chromosomes and r is the rank of the individual, where 1 is the best.
3.3 Genetic operators
Genetic operators provide the basic search mechanism of the GA. The operators are
used to create new solutions based on existing solutions in the population. There are
two basic types of operators: crossover and mutation. Operators for real-valued
representations, that is an alphabet of floats, were developed by Michalewicz (1994).
Crossover takes two individuals and produces two new individuals, while mutation
alters one individual to produce a single new solution. The application of these two
basic types of operators and their derivatives depends on the chromosome
representation used. For real X and Y m-dimensional vectors representing
chromosomes, the following operators are defined: uniform mutation, non-uniform
mutation, multi-non-uniform mutation, boundary mutation, simple crossover,
arithmetic crossover, and heuristic crossover. Let ai and bi be the lower and upper
bounds, respectively, for each variable i.
3.3.1 Mutation. Uniform mutation randomly selects one variable, j, and sets it
equal to a uniform random number bounded by ai and bi terms:
x0i~U ai, bið Þ if i~j
xi otherwise
�ð11Þ
Boundary mutation randomly selects one variable, j, and sets it equal to either its
lower or upper bound, where r5U(0, 1):
x0i~
ai if i~j, rv0:5
bi if i~j, r§0:5
xi otherwise
8><>: ð12Þ
Non-uniform mutation randomly selects one variable, j, and sets it equal to a non-
uniform random number:
x0i~
xiz bi{xið Þ r2 1{ GGmax
� �� �b
if r1v0:5
xi{ xizaið Þ r2 1{ GGmax
� �� �b
if r1§0:5
xi otherwise
8>>>><>>>>:
ð13Þ
3280 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
where r1 and r2 are uniform random numbers between (0, 1), G and Gmax are
respectively the current and the maximum number of generations, and b is a shape
parameter.
The multi-non-uniform mutation operator applies the non-uniform operator to
all of the variables in the parent X.
3.3.2 Crossover. Real-valued simple crossover generates a random number r from
a uniform distribution from 1 to m and creates two new individuals (X9 and Y9)
according to equation (14).
x0i y0i� �
~xi yið Þ if ivr
yi xið Þ otherwise
�ð14Þ
Arithmetic crossover produces two complimentary linear combinations of the
parents, where r5U(0, 1):
X0~rXz 1{rð ÞY ð15Þ
Y0~ 1{rð ÞXzrY ð16Þ
Heuristic crossover produces a linear extrapolation of the two individuals. This is
the only operator that uses fitness information. A new individual, X9, is created
using equation (17), where r is a random number r5U(0, 1) and X is better than Y in
terms of fitness. If X9 is not feasible, that is there is at least a new gene smaller than
ai or larger than bi, then generate a new random number r and create a new solution
using equation (17), otherwise stop. After t failures, the process is not repeated and
the children are set equal to the parents.
X0~Xzr X{Y
� �ð17Þ
Y0~X ð18Þ
3.4 Initialization, termination and fitness function
To start the search of the optimal solution by GA it is necessary to provide an initial
population as indicated in figure 1. The most common method is to randomly
generate solutions for the entire population. However, as GAs can iterativelyimprove existing solutions, the starting population can be seeded with potentially
good solutions, with the remainder of the population being randomly generated.
The GA moves from generation to generation selecting and reproducing parents
until a termination criterion is met. A maximum number of generations is
commonly used to stop the GA search. Another termination strategy involves
population convergence criteria. In general, GAs will force much of the entirepopulation to converge to a single solution. The algorithm can be terminated when
the sum of the deviations among individuals becomes smaller than some specified
threshold or due to a lack of improvement in the best solution over a specified
number of generations. Alternatively, a target value for the evaluation measure can
be established based on some arbitrarily ‘acceptable’ threshold. Several strategies
can be used in conjunction with each other.
Evaluation functions of many forms can be used in a GA, subject to the minimal
requirement that the function can map the population into a partially ordered set.
PAN-sharpening by genetic algorithm 3281
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
As stated in section 2.3, the evaluation function to be optimized is Q4, which is
particularly suited for minimizing radiometric and spectral distortions.
3.5 Summary
The GA parameters selected for the optimization of Q4 are listed in table 1. The
initial random population is set to 50 and the maximum number of generations is
selected equal to 100. In section 5 we show that these two parameters are proper and
adequate to find the optimal solution in the sense of Q4 maximization. The first (or
unique, in some cases) value for the mutation/crossover operators indicates the
number of times that the operator is applied at each generation. The second and
third components of the non-uniform mutation and multi-non-uniform mutation
parameter vectors are the maximum number of generations and the shape factor
(b53), respectively. The second component of the heuristic crossover parameter
vector indicates the number of failures (t) and it is set to 3. In the reproduction
operator, the probability of selecting the best individual is set to 0.05. The interval of
variation for the gain parameters is the same for all bands and spans the interval
[210, 10] to ensure a wide state space. Each gain parameter gl is spatially constant
on the corresponding band l and will assume a positive value, typically lower than 3;
the interval [0, 3] would therefore be sufficient to ensure satisfactory results.
4. Data fusion
Figure 2 outlines a procedure based on ATWT, suitable for fusion of MS and PAN
image data whose scale ratio is 4. Both the higher resolution PAN image and the
lower resolution MS image dataset are decomposed by the two-level ATWT. The
MS images are previously interpolated by 4 along rows and columns to process MS
images having the same spatial scale as the PAN image. The interpolator and
decimator blocks are implemented by applying twice the up-sampling and down-
sampling operators with a 23-tap filter as described by Aiazzi et al. (2002).
To drive the injection of PAN data, the operator T [di(m, n)] reported in
equation (4) is applied, with g representing the best chromosome derived by the GA
search. The Q4 index, which is able to measure distortions but requires the fused and
the reference MS images as inputs, is calculated at coarser resolution, that is at a
resolution degraded by a factor equal to 4. The Q4N6N values are evaluated on
868 pixel blocks instead of 32632 as suggested by Alparone et al. (2004), as the GA
Table 1. GA parameters used for real-valued Q4 function optimization.
Operation Parameters
Initial population 50Normalized geometric selection 0.05Uniform mutation 4Non-uniform mutation [4, 100, 3]Multi-non-uniform mutation [6, 100, 3]Boundary mutation 4Simple crossover 2Arithmetic crossover 2Heuristic crossover [2, 3]Maximum generation 100Chromosomes bounds (for each band) [210, 10]
3282 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
is applied to MS images at a lower scale. The GA is finally applied with the
parameters described in table 1.
5. Experimental results and comparisons
Evaluation of the efficiency and robustness of the proposed fusion algorithm was
carried out on three different dataset acquired by the Quickbird spaceborne MS
scanner on the urban and suburban areas of Pavia (Italy), Rome (Italy) and Ceuta
(Spain). The four MS bands span the visible and NIR wavelengths with blue
(B15450–520 nm), green (B25520–600 nm), red (B35630–690 nm) and NIR
(B45760–900 nm); the PAN band is acquired on the wavelength interval 450–
900 nm. All the data were radiometrically calibrated from digital counts,
orthorectified (i.e. resampled to uniform ground resolutions of 2.8 m and 0.7 m
Ground Spatial Distance (GSD) for MS and PAN, respectively), and packed in 16-
bit words. The full scale of all the bands is 2047 (11 bits) and is reached in the NIR
wavelengths. The original sizes of the PAN and MS images were:
N 204862048 and 5126512 for the PAN and MS area of Pavia;
N 102461024 and 2566256 for the PAN and MS area of Rome;
N 153661536 and 3846384 for the PAN and MS area of Ceuta.
To obtain a good fusion of the injected PAN data, that is with low distortion, the
fused MS images should be as similar as possible to the MS set of images that the
corresponding sensor would observe with the highest spatial resolution (0.7 m), if it
existed. Unfortunately, those images are not available; therefore, PAN and MS data
are first spatially degraded by a factor equal to the ratio between MS and PAN
Figure 2. Flowchart of PAN-sharpening based on GA with 1 : 4 scale ratio.
PAN-sharpening by genetic algorithm 3283
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
resolutions (Wald et al. 1997). Thus, 2.8 m PAN and 11.2 m MS are synthetically
generated and processed to enhance MS images up to 2.8 m. The true MS data at
2.8 m used as reference images are available for objective distortion measurements.
Table 2 lists the values of the chromosomes after 100 iterations, providing the best
values of Q4. For each result, it is shown that equalization gains are between 0 and
1.3, hence validating the conservative choice of limiting the state space between 210
and 10. The gains change considerably depending on the data set and on the MS
band; in particular, they are lower on the blue band (B1) than on the other three
bands, especially on NIR (B3). It is worth noting that the gains determined by the
GA are not derivable from the equivalent gains computed by any statistical injection
model (CBD, RWM, SDM, etc.), because the GA injection process is driven by the
minimization of not only radiometric (through intraband processing) but also
spectral (through interband processing) distortions. Figure 3 reports the trend of Q4
for the best chromosome between a population of 50 individuals vs. the generation
step. The three curves show that 100 iterations are sufficient to reach the asymptotic
value for optimum search. Note that after 20 iterations the variations in Q4 are
extremely low. Thus, if necessary, CPU time can be sensibly reduced from 100
iterations to 20–30 iterations without compromising the fusion quality. In our
simulations the number of iterations have not been reduced as the CPU
requirements were not prohibitive.
The proposed method was compared to some very efficient fusion methods, listed
below:
N Synthetic Variable Ratio (SVR) as proposed by Zhang (1999);
N SDM (Alparone et al. 2003), with a 23-taps filter as reported by Aiazzi et al.
(2002);
N RWM, as described by Ranchin et al. (2003);
N CBD (Aiazzi et al. 2002) with a 23-tap filter;
N Gram Schmidt (GS) as implemented by ENVIr software (Laben and Brower
1998).
The distortion measures of simply resampled MS bands, referred to as EXP, are also
presented for comparison.
Mean bias values were calculated on each band and are listed in table 3. All values
are low and tend to zero except those related to the SVR method, which shows
significant bias values. In the MRA methods and also in the GS method the results
are acceptable; in particular, GA and GS methods are designed to maintain bias
equal to the corresponding values of the resampled MS images (EXP). Table 4
reports, for each band, the CC on the left and the universal quality index on the
right. CC measures how much the texture of the fused MS images reflects that of the
reference MS images, but it does not take into account mean bias and contrast
deviation. The values of CC and Q on the blue band (B1) are lower than those
Table 2. Best chromosomes (g) computed from degraded data (PAVIA, ROME andCEUTA) after 100 iterations.
B1 B2 B3 B4
PAVIA 0.160 0.446 0.688 1.282ROME 0.376 0.788 0.792 1.186CEUTA 0.449 0.930 0.899 0.893
3284 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
calculated on the other bands, especially for the Pavia data set. Note that the GA
method generally provides the best results for all the data sets. A performance
ranking of algorithms indicates that the GA method is followed by RWM, CBD and
GS methods, whose performances depend highly on the particular data set being
considered. As an example, the Q index of CBD is higher than that of RWM and GS
for all bands in Pavia but this is not true for the Rome data set, and it is lower than
that of RWM on all bands of the Ceuta MS image. Table 5 reports other distortion
measures that are not band dependent:
N Relative dimensionless global error in synthesis, ERGAS, to measure radio-
metric distortions (Wald 2000);
N Spectral Angle Mapper (SAM), to measure spectral distortions (Alparone et al.
2003);
N Q4 index, to measure radiometric and spectral distortions jointly (Alparone et
al. 2004).
An excellent fusion quality is obtained when the values of ERGAS and SAM
converge to 0. A unique index can be useful for determining whether one fusion
method is globally better than another. The GA results are the best for all three data
sets analysed. The validity of the optimization process driven by the Q4 index,
jointly minimizing spectral and radiometric distortions, is confirmed by the
reduction in the single-band distortion indexes, i.e. ERGAS and SAM. The GA
method shows the best performance, while the RWM, CBD and GS methods have
similar behaviour but are strongly dependent on the particular data set. This result
proves the robustness of the GA method: when the statistical properties of the MS
Figure 3. Plot of optimal Q4 values at each generation step. The three curves refer toPAVIA data (solid line), ROME data (dotted line) and CEUTA data (dash-dot line). Theiteration steps are visualized in logarithmic scale.
PAN-sharpening by genetic algorithm 3285
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
images change, the fusion quality is still very satisfactory. Note also that the
computational burden of GA is lower than that in the RWM or CBD methods,
which use a local injection model, and is comparable to that of the GS method.
Figure 4 shows, on the left, the fusion results obtained by the GA method at 2.8 m,
and, on the right, the reference images. The sizes of the MS images in figure 4 are set
according to the spatial dimensions of the original data, as listed at the beginning of
this section. Figure 5 reports 2566256 tiles of the GA fused data at 0.7 m on the left,
and the corresponding tiles of the resampled data on the right side. In all figures,
true colour (B3–B2–B1) composites, instead of false colour (B4–B3–B2) composites,
are shown because fusion methods typically fail on the blue band (B1). The fused
images of figure 4 are very similar to the reference MS images and they give a very
accurate result in the colour representation. Figure 5 demonstrates that PAN
accurately sharpens and does not overenhance the MS images. Finally, a
comparison between fused and lower-scale original images shows that evident
chromatic distortions are avoided completely.
6. Conclusion
A novel GA fusion method has been proposed for PAN-sharpening of MS images.
The algorithm is based on undecimated multiresolution analysis, which adopts
injection-model parameters derived at coarser scales by means of a GA.
Table 3. Mean bias between fused and reference 2.8 m MS bands.
BIAS
B1 B2 B3 B4
PAVIAEXP 0.002 0.007 0.011 0.027GA 0.002 0.007 0.012 0.027SVR 0.590 0.624 0.097 20.398SDM 20.255 20.310 20.134 20.097RWM 20.080 20.127 20.118 20.024CBD 20.027 20.058 20.117 0.204GS 0.002 0.007 0.012 0.023
ROMEEXP 0.066 0.005 0.001 0.020GA 0.066 0.005 0.001 0.020SVR 3.208 2.310 20.162 20.973SDM 20.800 21.090 20.765 20.790RWM 20.002 0.025 0.013 0.060CBD 20.067 20.049 20.022 0.220GS 0.006 0.006 0.001 0.020
CEUTAEXP 0.013 0.032 0.026 0.034GA 0.013 0.032 0.026 0.034SVR 2.939 2.468 20.280 21.387SDM 20.501 20.644 20.394 20.377RWM 0.061 0.052 0.019 0.043CBD 20.094 20.147 0.007 0.321GS 0.013 0.032 0.026 0.034
EXP, MS image expanded to PAN scale; GA, genetic algorithm; SVR, Synthetic VariableRatio; SDM, Spectral Distortion Minimizing; RWM, Ranchin, Wald, Mangolini; CBD,Context Based Decision; GS, Gram Schmidt.
3286 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
Table 5. Unique (not band-dependent) quality indexes measuring spectral, radiometric andgeometric distortions between fused and original 2.8 m MS bands.
EXP GA SVR SDM RWM CBD GS
PAVIAERGAS 1.760 1.250 1.589 1.676 1.694 1.429 1.388SAM 2.142 1.661 2.143 2.142 2.075 1.886 1.823Q4 0.750 0.906 0.841 0.864 0.865 0.889 0.874
ROMEERGAS 4.907 3.162 3.744 3.724 3.376 3.366 3.468SAM 4.050 3.597 4.050 4.051 3.747 3.735 3.715Q4 0.791 0.929 0.899 0.916 0.919 0.917 0.911
CEUTAERGAS 3.065 2.021 2.433 2.360 2.109 2.211 2.280SAM 1.765 1.621 1.765 1.765 1.700 1.843 1.981Q4 0.809 0.931 0.916 0.914 0.923 0.915 0.922
EXP, MS image expanded to PAN scale; GA, genetic algorithm; SVR, Synthetic VariableRatio; SDM, Spectral Distortion Minimizing; RWM, Ranchin, Wald, Mangolini; CBD,Context Based Decision; GS, Gram Schmidt.
Table 4. CC and Q between fused and reference MS bands at 2.8 m.
CC Q
B1 B2 B3 B4 B1 B2 B3 B4
PAVIAEXP 0.860 0.843 0.852 0.819 0.749 0.740 0.757 0.684GA 0.898 0.906 0.916 0.925 0.842 0.873 0.886 0.898SVR 0.531 0.735 0.912 0.924 0.413 0.693 0.867 0.848SDM 0.598 0.771 0.912 0.929 0.387 0.659 0.883 0.897RWM 0.794 0.829 0.837 0.898 0.705 0.785 0.801 0.873CBD 0.862 0.878 0.889 0.916 0.802 0.841 0.856 0.889GS 0.884 0.894 0.905 0.908 0.801 0.830 0.848 0.859
ROMEEXP 0.851 0.847 0.855 0.857 0.785 0.782 0.792 0.787GA 0.931 0.941 0.945 0.946 0.899 0.917 0.918 0.932SVR 0.780 0.903 0.945 0.955 0.731 0.867 0.909 0.921SDM 0.843 0.916 0.945 0.951 0.763 0.883 0.924 0.940RWM 0.921 0.932 0.932 0.941 0.882 0.905 0.898 0.927CBD 0.918 0.931 0.934 0.941 0.884 0.905 0.905 0.925GS 0.926 0.936 0.940 0.941 0.868 0.888 0.890 0.918
CEUTAEXP 0.913 0.913 0.928 0.936 0.790 0.784 0.788 0.760GA 0.957 0.963 0.971 0.974 0.902 0.914 0.916 0.907SVR 0.867 0.936 0.970 0.973 0.802 0.882 0.917 0.906SDM 0.907 0.948 0.970 0.973 0.809 0.884 0.916 0.905RWM 0.949 0.958 0.969 0.971 0.888 0.904 0.911 0.896CBD 0.943 0.954 0.967 0.969 0.873 0.893 0.903 0.888GS 0.955 0.959 0.964 0.963 0.891 0.902 0.902 0.885
EXP, MS image expanded to PAN scale; GA, genetic algorithm; SVR, Synthetic VariableRatio; SDM, Spectral Distortion Minimizing; RWM, Ranchin, Wald, Mangolini; CBD,Context Based Decision; GS, Gram Schmidt.
PAN-sharpening by genetic algorithm 3287
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
Experimental results obtained on co-registrated MS and PAN images acquired by
Quickbird satellite sensors confirm that the proposed approach significantly
increases the spectral fidelity of the fused images, with respect to the most
promising fusion techniques in the literature, also maintaining very good properties
in terms of spatial enhancement. A peculiar property of the proposed GA-based
Figure 4. True colour (bands 3–2–1) MS images at 2.8 m resolution: (a), (c), (e) GA fusionresults; (b), (d), (f) reference images, for PAVIA data (first row, 5126512 pixels), ROME data(second row, 2566256 pixels) and CEUTA data (third row, 3846384 pixels).
3288 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
Figure 5. True colour (band 3–2–1) MS details (2566256): (a), (c), (e), GA fusion results at0.7 m; (b), (d), (f) 2.8 m original images resampled to 0.7 m for PAVIA data (first row), ROMEdata (second row) and CEUTA data (third row).
PAN-sharpening by genetic algorithm 3289
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
fusion algorithm is the robustness to different data characteristics, while the
performances of other state-of-the-art fusion methods often depend on global and
local band statistics. The global injection model provides excellent results and even ifthe local-model approach adopted by several algorithms seems more suitable for
context representation, the GA performances are not limited by numerical
instabilities on the model parameters that may produce, in pure statistical fusion
methods, some relevant artefacts in the fused products. The maximization of Q4 at
coarser scales and the application of the derived solution at finer scales are allowed
by the ARSIS concept and are demonstrated by the excellent results.
ReferencesAIAZZI, B., ALPARONE, L., BARONTI, S. and GARZELLI, A., 2002, Context-driven fusion of
high spatial and spectral resolution data based on oversampled multiresolution
analysis. IEEE Transactions on Geoscience and Remote Sensing, 40, pp. 2300–2312.
ALPARONE, L., AIAZZI, B., BARONTI, S. and GARZELLI, A., 2003, Sharpening of very high
resolution images with spectral distortion minimization. In IEEE International
Geoscience Remote Sensing Symposium 1, 21–25 July, pp. 458–460 (Piscatway, NY:
IEEE).
ALPARONE, L., BARONTI, S., GARZELLI, A. and NENCINI, F., 2004, A global quality
measurement of Pan-sharpened multispectral imagery. IEEE Geoscience and Remote
Sensing Letters, 1, pp. 313–317.
ANYONG, Q., CHING, K.L. and LANG, J., 2001, Electromagnetic inverse scattering of two-
dimensional perfectly conducting objects by real-coded genetic algorithm. IEEE
Transactions on Geoscience and Remote Sensing, 39, pp. 665–676.
CAORSI, S., MASSA, A. and PASTORINO, M., 2000, A computational technique based on a real-
coded genetic algorithm for microwave imaging purposes. IEEE Transactions on
Geoscience and Remote Sensing, 38, pp. 1697–1708.
CHAIYARATANA, N. and ZALZALA, A.M.S., 1997, Recent developments in evolutionary and
genetic algorithms: theory and applications. In Second International Conference on
Genetic Algorithms in Engineering Systems: Innovations and Applications, 2–4
September, pp. 270–277 (London: IEEE).
CHEN, L., 2003, A study of applying genetic programming to reservoir trophic state
evaluation using remote sensor data. International Journal of Remote Sensing, 24, pp.
2265–2275.
CHIBANI, Y. and HOUACINE, A., 2002, The joint use of IHS transform and redundant wavelet
decomposition for fusing multispectral and panchromatic images. International
Journal of Remote Sensing, 23, pp. 3821–3833.
DAVIS, L., 1991, The Handbook of Genetic Algorithms (New York: Van Nostrand Reingold).
DUTILLEUX, P., 1989, An implementation of the ‘algorithme a trous’ to compute the wavelet
transform. In Wavelets: Time–Frequency Methods and Phase Space, J.M. Combes, A.
Grossman and Ph. Tchamitchian (Eds), pp. 298–304 (Berlin: Springer).
GARZELLI, A., BENELLI, G., BARNI, M. and MAGINI, M., 2000, Improving wavelet based
merging of panchromatic and multispectral images by contextual information. In
Image and Signal Processing for Remote Sensing VI, Proceedings of SPIE 4170,
EUROPTO Series, pp. 82–91 (Brussels: EUROPTO Series).
GARZELLI, A. and NENCINI, F., 2005, Interband structure modeling for Pan-sharpening of
very high resolution multispectral images. Information Fusion, 6, pp. 213–224.
GHOSHRAY, S. and YEN, K.K., 1995, More efficient genetic algorithm for solving
optimization problems. In IEEE International Conference on Systems, Man and
Cybernetics, 1995. Intelligent Systems for the 21st Century, 22–25 October, pp.
4515–4520 (Piscataway, NJ: IEEE).
GOLDEBERG, D., 1989, Genetic Algorithms in Search, Optimization and Machine Learning
(Wokingam, MA: Addison-Wesley).
3290 A. Garzelli and F. Nencini
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
HO, S.Y. and LEE, K.Z., 2001, An efficient evolutionary image segmentation algorithm. In
Congress on Evolutionary Computation 2, 27–30 May, pp. 1327–1334 (Piscataway, NJ:
IEEE).
HOLLAND, J., 1975, Adaptation in Natural and Artificial Systems (Ann Arbor: The University
of Michigan).
JEON, B.K., JANG, J.H. and HONG, K.S., 2002, Road detection in spaceborne SAR images
using a genetic algorithm. IEEE Transaction on Geoscience and Remote Sensing, 40,
pp. 22–29.
JIN, Y.Q. and WANG, Y., 2001, A genetic algorithm to simultaneously retrieve land surface
roughness and soil wetness. International Journal of Remote Sensing, 22, pp.
3093–3099.
JOINES, J. and HOUCK, C., 1994, On the use of non-stationary penalty functions to solve
constrained optimization problems with genetic algorithms. In IEEE International
Symposium on Evolutionary Computation, 27–30 May, Orlando, pp. 579–584
(Piscataway, NJ: IEEE).
KUMAR, A.S., KARTIKEYAN, B. and MAJUMDAR, K.L., 2000, Band sharpening of IRS-
multispectral imagery by cubic spline wavelets. International Journal of Remote
Sensing, 21, pp. 581–594.
LABEN, C.A. and BROWER, B.V., 1998, Process for enhancing the spatial resolution of
multispectral imagery using Pan-sharpening. US Patent 6,011,875, Eastman Kodak
Company.
LI, S., KWOK, J.T. and WANG, Y., 2002, Using the discrete wavelet frame transform to merge
Landsat TM and SPOT panchromatic images. Information Fusion, 3, pp. 17–23.
LIN, Y.C. and SARABANDI, K., 1999, Retrieval of forest parameters using a fractal-based
coherent scattering model and a genetic algorithm. IEEE Transactions on Geoscience
and Remote Sensing, 37, pp. 1415–1424.
LIU, J. and TANG, Y.Y., 1998, Evolutionary image processing. In IEEE International
Conference on Evolutionary Computation, 4–9 May, pp. 283–288 (Piscataway, NJ:
IEEE).
MALLAT, S.G., 1999, A Wavelet Tour of Signal Processing, 2nd edn (New York: Academic
Press).
MAULIK, U. and BANDYOPADHYAY, S., 2003, Fuzzy partitioning using a real coded variable-
length genetic algorithm for pixel classification. IEEE Transaction on Geoscience and
Remote Sensing, 41, pp. 1075–1081.
MERTENS, K.C., VERBEKE, L.P.C., DUCHEYNE, E.I. and DE WULF, R.R., 2003, Using genetic
algorithms in sub-pixel mapping. International Journal of Remote Sensing, 24, pp.
4241–4247.
MICHALEWICZ, Z., 1994, Genetic Algorithms + Data Structures5Evolution Programs (New
York: Springer-Verlag).
MUNTEANU, C. and ROSA, A., 2000, Towards automatic image enhancement using genetic
algorithms. In Congress on Evolutionary Computation, 2, 16–19 July, pp. 1535–1542
(Piscataway, NJ: IEEE).
NUNEZ, J., OTAZU, X., FORS, O., PRADES, A., PAL, V. and ARBIOL, R., 1999, Multiresolution-
based image fusion with additive wavelet decomposition. IEEE Transactions on
Geoscience and Remote Sensing, 37, pp. 1204–1211.
PAL, S.K. and BANDYOPADHYAY, S., 2001, Pixel classification using variable string genetic
algorithms with chromosome differentiation. IEEE Transaction on Geoscience and
Remote Sensing, 39, pp. 303–308.
PAL, S.K., BANDYOPADHYAY, S. and MURTHY, C.A., 2001, Genetic classifiers for remotely
sensed images: comparison with standard methods. International Journal of Remote
Sensing, 22, pp. 2545–2569.
RANCHIN, T., AIAZZI, B., ALPARONE, L., BARONTI, S. and WALD, L., 2003, Image fusion: the
ARSIS concept and some successful implementation schemes. ISPRS Journal of
Photogrammetry and Remote Sensing, 58, pp. 4–18.
PAN-sharpening by genetic algorithm 3291
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013
RANCHIN, T. and WALD, L., 2000, Fusion of high spatial and spectral resolution images: the
ARSIS concept and its implementation. Photogrammetric Engineering and Remote
Sensing, 66, pp. 49–61.
SCHEUNDERS, P. and BACKER, S.D., 2001, Fusion and merging of multispectral images with
use of multiscale fundamental forms. Journal of the Optical Society of America A, 18,
pp. 2468–2477.
SCHOWENGERDT, R.A., 1980, Reconstruction of multispatial, multispectral image data using
spatial frequency content. Photogrammetric Engineering and Remote Sensing, 46, pp.
1325–1334.
TSO, B.C.K. and MATHER, P.M., 1999, Classification of multisource remote sensing imagery
using a genetic algorithm and Markov random fields. IEEE Transactions on
Geoscience and Remote Sensing, 37, pp. 1255–1260.
VASCONCELOS, J.A., RAMIREZ, J.A., TAKAHASHI, R.H.C. and SALDANHA, R.R., 2001,
Improvements in genetic algorithms. IEEE Transactions on Magnetics, 3, pp.
3414–3417.
WALD, L., 2000, Quality of high resolution synthesized images: is there a simple criterion? In
Proceedings of the International Conference on Fusion of Earth Data, 26–28 January,
pp. 99–105 (Nice: SEE Greca).
WALD, L., 2002, Data Fusion: Definitions and Architectures. Fusion of Images of Different
Spatial Resolution (Paris, France: Les Presses de l’Ecole des Mines).
WALD, L., RANCHIN, T. and MANGOLINI, M., 1997, Fusion of satellite images of different
spatial resolutions: assessing the quality of resulting images. Photogrammetric
Engineering and Remote Sensing, 63, pp. 691–699.
WANG, Z. and BOVIK, A.C., 2002, A universal image quality index. IEEE Signal Processing
Letters, 9, pp. 81–84.
YAO, H. and TIAN, L., 2003, A Genetic-Algorithm-based Selective Principal Component
Analysis (GA-SPCA) method for high-dimensional data feature extraction. IEEE
Transactions on Geoscience and Remote Sensing, 41, pp. 1469–1478.
ZHAN, H., LEE, Z., SHI, P., CHEN, C. and CARDER, K.L., 2003, Retrieval of water optical
properties for optically deep waters using genetic algorithms. IEEE Transactions on
Geoscience and Remote Sensing, 41, pp. 1123–1128.
ZHANG, Y., 1999, A new merging method and its spectral and spatial effects. International
Journal of Remote Sensing, 20, pp. 2003–2014.
3292 PAN-sharpening by genetic algorithm
Dow
nloa
ded
by [
Uni
vers
ity o
f M
iam
i] a
t 01:
26 2
4 Se
ptem
ber
2013