Vous êtes sur la page 1sur 12

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO.

5, MAY 2007

1303

Comparative Study of Semi-Implicit Schemes for


Nonlinear Diffusion in Hyperspectral Imagery
Julio M. Duarte-Carvajalino, Student Member, IEEE, Paul E. Castillo, and Miguel Velez-Reyes, Senior Member, IEEE

AbstractNonlinear diffusion has been successfully employed


over the past two decades to enhance images by reducing undesirable intensity variability within the objects in the image, while enhancing the contrast of the boundaries (edges) in scalar and, more
recently, in vector-valued images, such as color, multispectral, and
hyperspectral imagery. In this paper, we show that nonlinear diffusion can improve the classification accuracy of hyperspectral imagery by reducing the spatial and spectral variability of the image,
while preserving the boundaries of the objects. We also show that
semi-implicit schemes can speedup significantly the evolution of
the nonlinear diffusion equation with respect to traditional explicit
schemes.
Index TermsHyperspectral imaging, nonlinear diffusion, partial differential equations (PDEs), preconditioning, remote sensing,
scale space, semi-implicit schemes, vector image processing.

I. INTRODUCTION

EMOTE-SENSING imaging can provide synoptic,


repetitive, consistent, and comprehensive environmental
monitoring of the earth ecosystem, revealing patterns and
relationships unavailable when using traditional data-gathering
techniques. Hyperspectral imaging technology has the potential
of extracting more plentiful and accurate environmental information, given the enhanced discrimination capabilities of high
spectral resolution imagery. However, the natural variability
of the material spectra, noise, and degradation added by the
transmission media and sensor system reduce the separability
of the different structures in hyperspectral imagery, reducing
the accuracy of segmentation and classification algorithms.
For two decades, techniques based on partial differential
equations (PDEs) have been used in image processing for
image segmentation, deblurring, restoration, smoothing, and
multiscale image representation. Among these techniques, parabolic PDEs have found a lot of attention for image smoothing
and image restoration purposes.
This paper introduces image smoothing of hyperspectral images, using a regularized nonlinear diffusion PDE. We show that
Manuscript received December 20, 2005; revised January 24, 2007. This
work was supported in part by CenSSIS, the Center for Subsurface Sensing
and Imaging Systems, under the Engineering Research Center Program of
the National Science Foundation (NSF) Award Number EEC-9986821 and
the NSF-EPSCOR fellowship, from the PR program. The associate editor
coordinating the review of this manuscript and approving it for publication was
Dr. Jacques Blanc-Talon.
J. M. Duarte-Carvajalino and M. Velez-Reyes are with the Laboratory of Applied Remote Sensing and Image Processing (LARSIP), University of Puerto
Rico, Mayagez, PR 00681-9048 USA (e-mail: jmartin@ece.uprm.edu; miguel.
velez@ece.uprm.edu).
P. E. Castillo is with the Department of Mathematics, University of Puerto
Rico, Mayagez, PR 00681-9018 USA (e-mail: castillo@math.uprm.edu).
Digital Object Identifier 10.1109/TIP.2007.894266

semi-implicit discretization schemes have better performance


(in terms of accuracy and CPU time) than traditional explicit
schemes to solve the nonlinear diffusion PDE on hyperspectral
imagery. We also show that nonlinear diffusion can be used
to reduce the spatial and spectral variability in hyperspectral
imagery, improving classification accuracy (Sections IV and
V). We extend approximated semi-implicit schemes such as:
additive operator splitting (AOS) and alternating direction
implicit (ADI) schemes to vector-valued images. Additionally,
we also evaluate the use of the preconditioned conjugated
gradient (PCG) linear solvers as an alternative to AOS and ADI
schemes.
The performance of the vector-valued nonlinear diffusion
PDE is studied using four hyperspectral images. The first
is a synthetic image that allows a controlled environment to
measure CPU speed and accuracy in the solution of the PDE.
The second and third images are real hyperspectral images
acquired using the NASA AVIRIS hyperspectral sensor1 on the
Northwest Indian Pines test site (1992) and the Cuprite mining
district (1997) in Nevada. The last hyperspectral image is an
indoor image taken with the SOC700 Hyperspectral Imager by
the Surface Optics Corporation.2 The real hyperspectral images
are used here for the evaluation of the effect of nonlinear
diffusion on image classification.
To our knowledge, this is the first extension of semi-implicit
schemes to discretize and solve the nonlinear diffusion on hyperspectral imagery, detailing the effect of nonlinear diffusion
on the spatial and spectral variability of the image, as well as its
effect on classification accuracy, using real and synthetic hyperspectral images.
Nonlinear diffusion generates a scale-space representation of
the image (see Section II), which constitutes a powerful tool for
image processing, in computer vision. The scale-space representation of an image facilitates the detection of different structures in the image at their appropriate image scale. Hyperspectral remote sensing can benefit enormously from the scale-space
framework, which has been largely limited in their applications
to grayscale and color images to improve object recognition
from images taken on remote-sensing platforms.
II. BACKGROUND
A. Classification of Hyperspectral Imagery
Remote-sensing sensors are producing high spectral and spatial resolution imagery, increasing their potential for environmental monitoring, precision farming, insurance and car navigation at global and local scales. More recently, hyperspec1http://aviris.jpl.nasa.gov
2http://surfaceoptics.com

1057-7149/$25.00 2007 IEEE

1304

tral imaging technology has found applications beyond earth


remote sensing in agriculture, medicine, biology, pharmaceuticals, forensics, color vision, target detection, archaeology, and
many others near field applications. However, classification of
HSI imagery is primarily made on a pixel by pixel basis with
classification accuracy figures in the range 80%85%, and they
have not changed significantly in recent decades [1]. The natural variability of the material spectra and the noise added by
the transmission media and sensor system make necessary the
use of statistical methods for information extraction and pattern
recognition on hyperspectral imagery.
Statistical parametric and nonparametric classification
methods that derive directly from the Bayes rule suffer from
the Hughes phenomena [2], that is, in order to estimate accurately the density distribution of each class in high-dimensional
feature spaces, a prohibitively large amount of training samples
is required. Hyperspectral imagery with hundreds of channels
presents such a challenge.
Instead of processing remotely sensed imagery at the pixel
level, it has been proposed [3] to segment the images as a disjoint set of regions that have homogenous distinctive spectral
response and spatial uniformity. State of the art, object-based,
and object-oriented segmentation algorithms have been recently
used for remotely sensed multispectral imagery [4][7], but
little has been done on hyperspectral imagery due to the large
dimensionality of the data.
The underlying assumption in nonlinear diffusion is that the
true image is piecewise smooth and the original image is corrupted by noise. The scale-space framework introduced by the
diffusion equation has been also used for image compression
[8], in conjunction with level sets to detect movement (optic
flux) in image sequences [9], information extraction [10] and
image restoration [11], registration [12], and segmentation integrating level sets in a common framework [13].
B. Image Smoothing Using PDEs
Image smoothing by parabolic PDEs can be seen as a continuous transformation of the original image into a space of progressively smoother images identified by the scale or level
of image smoothing, in terms of pixel resolution [14]. However, the structures on an image can be of any size, that is, they
can be located at different image scales, in the continuum scale
scale generated by the PDE. The adequate selection of an image
scale smoothes out undesirable variability at lower scales that
constitute a source of error in segmentation and classification
algorithms.
Perona and Malik [15] proposed a nonlinear diffusion PDE
defined in such a way that forward diffusion (smoothing)
occurs more likely within the image structures; meanwhile,
backward diffusion (sharpening) may occur on their boundaries
(edges). Later, the pioneering work of Alvarez et al. [14] proved
that every scale scale that satisfies some natural architectural
(recursivity, causality, regularity, locality, and consistency),
information-reducing, and invariance properties (stability and
shape-preserving) is governed by a second order PDE with
the original image as its initial condition. They also showed
that the PeronaMalik nonlinear diffusion equation was ill

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

posed, though it can be made well posed if the flux is always


non-negative.
In 1996, Weickert [16] established the well posedness and
axiomatic requirements that a discretized diffusion equation
must satisfy in order to share similar scale-space properties
as the continuous PDE. The usual discretization scheme of
the nonlinear diffusion PDE is explicit, because it is simple to
implement. However, explicit schemes are limited by numerical stability conditions to small scale steps, which make them
computationally expensive. Otherwise, semi-implicit schemes
much better stability properties, so they can evolve at larger
scale steps [16].
C. Nonlinear Diffusion PDE
The nonlinear diffusion equation proposed by Perona and
Malik for scalar images is given by the following PDE, with
reflecting boundaries [16]

(1)
is the smoothed image, at the spatial position
where
given by the coordinates vector
, at scale , defined on
, with boundary
. The remaining terms in
the domain
, which is a nonlinear
(1) are the diffusion coefficient
is the original, noisy
function of the image gradient
image and
is the derivative along the normal to
that defines the boundary conditions of the PDE.
By a convenient selection of the diffusion coefficient in (1),
the intensity of the image is allowed to diffuse within the image
structures, eliminating, thus, the intraobject variability, while
preventing diffusion on the edges, characterized by a high
intensity gradient. However, as Alvarez et al. showed [14],
[17], [18], the diffusion coefficients proposed by PeronaMalik
lead to ill posedness of (1). Another shortcoming of the
PeronaMalik equation is that noise on the edges may be
amplified by backward diffusion. Alvarez et al. [17] showed
that the PeronaMalik equation can be made well posed,
by smoothing isotropically the image, before computing the
image gradient used by the diffusion coefficient. Equation (2)
corresponds to the regularized version of the PeronaMalik
PDE, where, for simplicity, we have dropped the dependence
on the spatial coordinates, , and time , and
is a smoothed version of obtained by convolving the image
of variance . With the
with a zero-mean Gaussian kernel
same boundary conditions as (1), the regularized PeronaMalik
equation is given by
(2)
In our computations, we use the nonlinear diffusion coefficient
proposed by Weickert [19]
(3)

DUARTE-CARVAJALINO et al.: COMPARATIVE STUDY OF SEMI-IMPLICIT SCHEMES

where segmentation-like results are obtained using


is the value that makes the flux
increasing for
and decreasing for
, but always non-negative.

, and

D. Explicit Scheme
For a 2-D scalar image with
posed as

, (2) can be decom-

1305

the Thomas algorithm [19]. Additionally, we use several preconditioning techniques to accelerate the convergence of the conjugated gradient method, which is the optimum iterative method
, provided that
to solve large sparse linear systems,
is symmetric positive definite, which is regularly the case of discretized PDEs.
1) Additive Operator Splitting (AOS) Method: AOS approximates the solution of (7) as [19]

(4)
Let us call
, the number of pixels in the image along the
axis and
the number of pixels along the axis. Numbering
the pixels of the image in major column format, the explicit
discretization of (4), in matrix-vector notation, is given by [19]
(5)
, being
the discretization of time
where
and
the discretization of the spatial coordinates,
is
, corresponding to the image (taken in
a vector of length
and
are both matrices
major column format) at scale
, being the identity matrix and
of size
the matrix of diffusion coefficients at scale
given by

(8)
. Since,
and
where
are both tridiagonal matrices, and can be obtained in linear
time using the Thomas algorithm.
2) Alternating Direction Implicit (ADI) Methods: Here, we
will consider, the three most widely used ADI methods [20],
[21]: Locally one-dimensional (LOD), DouglasRachford, and
PeacemanRachford. The simplest approximation to (7) is
given by ADI-LOD
(9)
The DouglasRachford method solves (7) as
(10)

(6)
In (6),
and
are, respectively, the image intensity and difand scale
fusion coefficient at coordinates
with
being the indices of the image, in
major column format. Here, we make the usual assumption that
so that
is the scale step.
E. Semi-Implicit Schemes
From consistency and stability considerations [19], [20], the
, which
explicit scheme indicated in (5) requires that
constitutes a severe limitation on the step size. Alternatively, we
can use semi-implicit discretization schemes, given by [19]
(7)
and are defined as before. Semi-implicit schemes
where
are unconditionally stable for all values of [19], [20]. Howequaever, (7) requires us to solve a linear system with
tions and unknowns, at each iteration step. The extra computational cost required to update the solution is compensated by
the numerical stability of semi-implicit schemes that allow us to
choose much larger scale steps, limited only by the accuracy of
the computed solution.
We consider here the AOS and ADI semi-implicit methods
that decompose (7) as a sum (AOS) or a product (ADI) of two
tridiagonal systems, which can be solved, in linear time, using

AOS and the ADI schemes considered until now are only
first order accurate in scale. A scheme that is second order
accurate in scale, for the isotropic diffusion equation, is the
PeacemanRachford scheme given in (11) [20]. However, this
scheme does not achieve second order accuracy in scale when
it is used on the nonlinear anisotropic diffusion equation, since
the diffusion coefficients are computed at the previous step
[21], tough, better accuracies can be expected using this scheme
is close to , i.e., at small scale steps
if
(11)
3) Preconditioned Conjugated Gradient (PCG): The conjugated gradient (CG) can be considered as an acceleration of
steepest descent to solve the linear system
, when
is symmetric positive definite [20]. The basic idea of precondiby [22]
tioning is to replace the system
(12)
where matrix
is called the preconditioner of and
is a matrix with better condition number than , such that the
conjugated gradient method converges faster, and the operation
must be performed fast, for any vector .

1306

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

The simplest preconditioner for


is based on the
Symmetric successive over-relaxation method (SSOR), which
[20]
has an explicit formula for the preconditioner

and minimal rate of change. Hence, the strength of an edge on


that measures
a vector-valued image is a function
and
. A possible choice for
the dissimilarity between
is
, which reduces to

(13)
being a lower triangular matrix and
. For any vector , the product
is equivalent
as
to solve the system
where

(14)
(16)
can be computed in linear time using forward
where
and backward substitution, since
and
are lower
and upper tridiagonal triangular matrices, respectively.
Another preconditioner commonly used in practice is the inas the
complete Cholesky factorization that approximates
product
, where
and is the error in the
approximation. In our work, we use the incomplete Cholesky
factorization with 0 drop tolerance as indicated in [23], which
means that has the same sparcity pattern as the lower triangular part of .
AOS and ADI methods provide also an approximation to matrix , as given by (9)(11); hence, they can be used also as
preconditioners. In particular, ADI is usually run with a fixed
number of times as a preconditioning step as in [24].
III. EXTENSION TO HYPERSPECTRAL IMAGERY
A hyperspectral image is an especial case of multispectral images in the sense that we have now hundreds of bands, instead
of tents of bands as it is usual in multispectral images, providing
much more information about the physical nature of the underlying substrate.
The first problem one face trying to extend the methods used
in computer vision for grayscale image processing to vectorvalued images is to extend the concept of gradient. The first
formal treatment of gradient in vector-valued images is due to
Di Zenzo in 1986 [25].
be a vector-valued image, with
Let
components
. Hence, the
first fundamental form in differential geometry is given by [26]

Of course, other possible choices are available for


as
in the Beltrami flow framework [27]. However, we use (16) for our extension of the nonlinear diffusion PDE to hyperspectral images, because it can be
interpreted as the Euclidean distance between two close vectors
). One can think that, in
(recall that we assumed
hyperspectral images, other similarity metrics, commonly used
in remote sensing, could also be used here instead of the image
gradient, but we will not explore those possibilities here any
longer.
We can represent a hyperspectral image as a matrix
, where
is the number of pixels
in the image, and each vector
corresponds to the spectral
signature of the th pixel in the image, taken in major column
format. Hence, the semi-implicit scheme given in (5) becomes
now for a hyperspectral image
(17)
where and are matrices of size
is the identity matrix and is defined by the diffusion coefficients as

(18)

(15)
For a unit vector
is a measure of the
rate of change in the image on the direction. The extrema
are obtained by the eigenvalues of the matrix
of
in the directions given by the eigenvectors. Let
be, respectively, the maximum and minimum values of the rate of
and
the respective directions of maximal
change in

(number of specNotice that we introduce in (18) the factor


tral bands) to normalize the measure of dissimilarity between
two vector valued pixels.
The AOS scheme can be extended to hyperspectral imagery
in a straightforward manner, by changing to in (8) as

(19)

DUARTE-CARVAJALINO et al.: COMPARATIVE STUDY OF SEMI-IMPLICIT SCHEMES

We can find the unknown matrices and in (19), using the


Thomas algorithm with the spectral vectors instead of scalars,
updating simultaneously all image bands. ADI methods can be
extended to vector valued images, similarly as we did here with
the AOS scheme.
Let us consider now the extension of PCG methods to vector
valued images. Since the CG method works with the image,
taken in major column format, as a single vector, we will use
a different, but equivalent, matrix representation of the hyper, where each
spectral image, namely
corresponds to each image band, taken in major
column
column format. Since matrix is the same for all image bands,
is also the same and (12) is now
the preconditioner
(20)
Solving (20) independently for each image band requires
solving
times a system of
equations and unknowns
at each iteration step. In order to speed up the process, we
propose here to update all image bands, simultaneously, based
on the mean value of the image, along the spectral direction,
i.e., update all image bands, based on the scalar image
(21)
Additionally, ADI and AOS schemes, used as preconditioners,
require the inclusion of a reduction factor , in order to avoid instability on the conjugated gradient method at high scale steps. If
, where
is the preconditioner in (22),
we want to find
shown at the bottom of the page, and the AOS preconditioner in
(23), shown at the bottom of the page, is the ADI-LOD preconditioner. The PeacemanRachford and DouglasRachford ADI
schemes are more expensive computationally and more sensitive to the scale step than ADI-LOD, and, hence, they are not
used here as preconditioners.
IV. EXPERIMENTS
The numerical methods indicated in Section II were implemented in Matlab, using the extensions to vector-valued images
explained in Section III. The classification was performed with
Multispec3 freeware software developed by Landgrebe. All the
hyperspectral images were normalized in the [0 1] range.
The choice of an optimum threshold value, , has been
addressed by several authors [27][29], and it is still an open
problem. However, in this work, we are concerned with the
relative running time and accuracy of each of the schemes
3http://dynamo.ecn.purdue.edu/~biehl/MultiSpec/

1307

considered, versus the explicit scheme, and, hence, finding the


is not relevant. The stopping scale is chosen
best value of
here as convenient integer value that facilitates the comparison
between the different schemes.
in (2), since 99% of the
It seems reasonable to select
, that is,
pixel from the center,
Gaussian area is within
which corresponds to the same stencil used by PeronaMalik
in the discretization of the nonlinear diffusion equation, i.e., a
3 3 grid. As Weickert [19] suggested, the regularization of the
image can be done efficiently using isotropic diffusion, on each
.
step with
Otherwise, since
, then
and the regularization of the image can be done with a single step of the
explicit scheme, which is stable and computationally cheaper
. In our experithan the semi-implicit schemes for
ments,
, which is a value that preserves the edges
on all images considered. Even though this value seems very
small to produce appreciable smoothing, when it is repeated
on each iteration step, it is enough to avoid enhancing of
impulsive noise (characterized by a very high peak or valley
surrounded by a smooth neighborhood) without destroying the
image edges.
The four hyperspectral images used in our experiments are as
follows.
1) The Indian Pines image Fig. 1(a) taken with the AVIRIS
(Airborne Visible/Infrared Imaging Spectrometer) sensor,
flown by NASA/Ames on June 12, over an area 6 mi west
of West Lafayette, IN. This image contains 145 145
pixels and 220 spectral bands in the 4002500-nm range,
for which ground truth exists. We disregard bands 13, 58,
77, 103110, 148166, and 218220, from the original
image either because they do not contain information,
they were too noisy or present strong illumination differences due to the sensor; hence, our Indian Pines
image has 145 145 pixels and 185 spectral bands in the
4102430-nm range.
2) A synthetic hyperspectral test image Fig. 1(b) made from
real pixels extracted from the Indian Pines image that fills
simple geometric figures: triangle, ellipse, donut, and a
common background. This image has 150 150 pixels and
the same number of bands that the Indian Pines image.
The pixels belonging to each geometric figure and background were selected at random and with uniform probability, from the pixels belonging to four different crops in
the Indiana Pines image: the Corn-min field (triangle), the
Soybeans-notill field (donut), the Soybeans-min field (ellipse), and the Hay-windrowed field (the background).
3) The Cuprite image Fig. 1(c) taken over the mining district, 2 km north of Cuprite, Nevada, with the AVIRIS

(22)
(23)

1308

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

Fig. 2. Superimposed spectra showing the spectral variability within each object and background on the synthetic image.

A. Performance in Terms of the Accuracy of the Computed


Solution

Fig. 1. Grayscale (spectral mean) representation of: (a) Indian Pines, (b) Synthetic, (c) Cuprite, and (d) Noisy False Leaves hyperspectral images.

(Airborne Visible/Infrared Imaging Spectrometer) sensor,


flown by NASA/Ames on June 19, 1997. This image contains five scenes for a total of 640 2378 pixels and 224
bands in the 3702500-nm range. We selected a portion of
the fourth scene, in the Cuprite image, of size 500 500
pixels, that corresponds to part of the mineral mapping in
the Cuprite mining district, reported by the U.S. Geological
Survey (USGS) spectroscopy laboratory in 1995, using the
expert system algorithm Tetracorder [30] and signatures of
60 sampled fields in the region. We use the USGS images
as ground truth.4 We selected from this image 50 bands:
172221 that correspond to the 20002480-nm vibrational
absorption region used by the USGS to mapping minerals
in the Cuprite image.
4) The False Leaves indoor image Fig. 1(d) of size 640 640
pixels and 120 bands in the 402908-nm range, obtained
by the Surface Optics Company using the SOC-700 hyperspectral imager. We selected a portion of this image of
size 540 575 pixels that contains all the objects present in
the original image. Additionally, and given that this spectrometer has a high spectral resolution, we selected only 30
bands of the original image, by taking one of each four consecutive bands. Since, this image has a high signal-to-noise
ratio (SNR) and none of the atmospheric effects that affect remote-sensed images, such as those taken with the
AVIRIS sensor; we add white Gaussian noise with zero
mean and
, of amplitude 10% relative to the maximum amplitude in the image, on each image band, and
then renormalization to the [0 1] range.
4http://speclab.cr.usgs.gov/PAPERS/tetracorder

The synthetic image is used to quantify the numerical performance in terms of the accuracy achieved by the different semiimplicit methods implemented, as the scale step increases, relative to a reference image generated using a very small scale step
and the semi-implicit CrankNicholson scheme
[20], which is a second order accurate scheme, both in scale and
space.
that preserves the edges in the
The highest value of
synthetic image, while reduce most of the internal variability
within the image objects is 0.015. Otherwise, the accuracy
of the explicit scheme at its maximum possible step size,
and the accuracy of AOS, ADI and PCG semi-imand
were all
plicit schemes for
compared to the reference image. We perform 1000 iterations
of the CrankNicholson scheme so that the real evolution in
; hence, the explicit
scale of the PDE is
should be run
times and the
scheme using
100, 20, 10, 5,
semi-implicit schemes should be run
and 2 times.
The best values for in the PCG-SSOR scheme were found,
simply by sweeping in the 0.01 to 2.0 range at intervals of
0.1. The values of found by this mean were 0.5, 0.4, 0.3, 0.15,
and 0.05 for
and
, respectively,
and they also correspond to the best values for the synthetic
and real hyperspectral images used. Finally, AOS and ADI-LOD
schemes used as preconditioners were implemented as indicated
1, 0.5,
in (27) and (28), where best results were found using
and
,
0.25, 0.125, and 0.025 for
respectively.
Fig. 2 shows the synthetic image and the spectral variability
within each image object and background, obtained by superposition of the spectrums of each pixel within each image region.
Fig. 3 shows the strong reduction on the spectral variability
within each image region, after nonlinear diffusion, while preserving the edges. Table I indicates the reduction on the variance within each image region. Fig. 4 shows the classification
map using the spectral angle mapper (SAM) in Multispec and
all image bands available.

DUARTE-CARVAJALINO et al.: COMPARATIVE STUDY OF SEMI-IMPLICIT SCHEMES

Fig. 3. Superimposed spectra showing the spectral variability within each object and background on the smoothed synthetic image.

1309

Fig. 5. Square error on the computed solution of each algorithm, relative to the
CrankNicholson scheme.

TABLE I
REDUCTION IN THE SPATIAL/SPECTRAL VARIABILITY

explore in the next set of experiments the performance of our


algorithms in terms of speed up and classification accuracies, as
the scale step increases, using real hyperspectral images.
the disk storage of an image of size
Finally, if we call
(see Section III), then AOS and ADI-LOD requires
disk space, the other ADI methods require
disk space, and the PCG methods require
disk space.
These values must be kept in mind when selecting between these
methods to solve the semi-implicit PDE (20), since typical hyperspectral images requires GBytes of disk storage.
B. Performance in Terms of the Classification Accuracy

Fig. 4. Classification of the synthetic image using SAM on (a) original and
(b) smoothed images.

Fig. 5 shows the square error of each one of the numerical


methods implemented here, relative to the CrankNicholson
scheme.
From Fig. 5, it can be noticed that all schemes have a larger
. In practice, we
error than the explicit scheme at
produces visible artifacts
found that a square error above
in the smoothed image. Hence, one could conclude that using
the PeacemanRachford and DouglasRachford schemes we
cannot achieve scale steps larger than 15 times the explicit
scheme without producing visible artifacts in the image. Similarly, AOS can only achieve a scale step 25 times larger than the
explicit scheme, result this that agrees with the ones reported
by [19], while the PCG method initialized with ADI-LOD can
reach higher step values than the ones used here. The remaining
in this image.
methods can achieve scale steps up to
Fig. 5 gives us an idea of the accuracy of the computed solution, as we increase the scale step, in the semi-implicit scheme.
However, in practice, the quality of the computed solution not
necessarily translates into higher classification accuracies. We

In order to test classification accuracy on the original and


smoothed hyperspectral images, we need training and testing
samples. The Indian Pines and Cuprite images are two of the
few hyperspectral remote-sensed images with reported ground
truth. Ground truth is very scarce in remote sensing, given the
costs involved in its acquisition. The False Leaves is an indoor
image with objects that can be easily identified. This image owes
its name to the fact that there are some plastic leaves that cannot
be distinguished from the real ones in the visible range; hence,
we must use a suitable combination of bands that include the
near infrared wavelengths to detect visually the false leaves and
select the corresponding training and testing samples.
Fig. 6(a) shows the ground truth available for the Indian Pines
image consisting of 16 classes, of which ten correspond to different kind of crops, five correspond to vegetation, and one corresponds to a building. Fig. 6(b) shows the training and testing
samples selected for 14 of the 16 classes identified on the Indian
Pines image. The other two classes (Oats and Alfalfa) were not
sampled since there are not enough training and testing samples
to perform the classification using classical statistical classification methods.
Fig. 7 shows the ground truth available for the Cuprite image,
which consists of 25 classes of minerals, grouped in five categories: sulfates, carbonates, Kaolinites, Clays, and other minerals. Fig. 8(a) shows the training and testing samples selected
on 11 classes of the Cuprite image. They are Calcite, Kaolinite and Semectite or Muscovite, K-Alumnite, Kaolinite, Alunite and Kaolinite or Muscovite, Calcite and Kaolinite, Chal-

1310

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

Fig. 6. Indian Pines image: (a) Ground truth and (b) training and testing samples (RGB shown corresponds to bands 29, 15, and 12).

Fig. 7. Ground truth Cuprite image.

Fig. 8. Training and testing samples on (a) Cuprite image (RGB corresponds
to bands 183, 193, and 207) and (b) False Leaves image (RGB corresponds to
bands 90, 68, and 29).

cedony, Na-Montmorillonite, Chlorite and Muscovite or Montmorillonite, High-Al Muscovite, and Med-Al Muscovite. We
consider the different kinds of Alunites as a single class, given
that it is extremely difficult to obtain pure training and testing

samples in this image. The remaining classes were not sampled


given that they do not provide enough training and testing samples or because they were too difficult of localize within the
Cuprite image, even with the help of the wavelengths recommended by the USGS to identify some of the minerals in this
image [see Fig. 8(a)].
Fig. 8(b) shows the training and testing samples on each one
of the different objects that can be identified in the Fake Leaves
image using the channels indicated. The classes in this image are
the wall, the jar, the flowerpot, the true leaves [seen as red leaves
on Fig. 8(b)], the false leaves [seen as dark leaves on Fig. 8(b)],
the metallic case, plastic label, paper label, and lens cover (dark
red) of the featured SOC-700 hyperspectral imager that appears
in the image.
We use all the classical classifiers available in Multispec [31]:
maximum likelihood (ML), Fisher linear likelihood (FLL), Euclidean distance (ED), extraction and classification of homogeneous objects (ECHO), SAM, and matched filter (MF) to evaluate how each classifier is affected by the nonlinear diffusion
process.
Since the smoothed Indian Pines image has 185 spectral
bands and the statistical classifiers employed here require more
training pixels than spectral bands in the image [31], we selected 20 bands using the SVD subset band selection algorithm
implemented at the UPRM Matlab toolbox [32] on each one of
the smoothed images.
The best classification results for the Indian Pines image were
and
runs of the explicit scheme
obtained using
. Hence, the semi-implicit methods were also run
at
and
that correspond to
for
10, 5, 2, and 1 step, respectively. For the Cuprite
image, we selected a value of
0.015 and 50 steps of the
explicit scheme, and, hence, we use the same values of as in
the Indian Pines image for the semi-implicit methods. On the
other hand, we obtained good classification results in the noisy
0.015 and 100 runs of the explicit
Fake Leaves image using
scheme; hence, the semi-implicit methods were run for 20, 10,
5, and 2 steps, respectively.
The classification results are shown in Tables IIIV for each
image and numerical method implemented. In these tables,
stands for the speedup relative to the explicit method, i.e., the
ratio between the running time of the explicit method and the
running time of the semi-implicit methods. The running time of
all the algorithms is given in minutes using a PC with 1.5-Gb of
RAM, CPU of 2.8 GHz, and Matlab for windows.
The highest classification accuracies and speedups are indicated in the tables, for each method, in bold and cursive.
The highest speedups were chosen as the maximum speedup
that keeps the classification accuracy very close or above the
classification accuracy achieved with the explicit scheme. Of
course, the best performance is for those methods that achieve
classification accuracies above the explicit method and high
speedups.
From Tables IIIV, one can see that all the smoothed images
achieve higher classification accuracies than using the original
image, on all classifiers, except for the ML classifier on the
Cuprite and Fake Leaves images. The bad performance of ML
on these images can be explained by the fact that ML becomes

DUARTE-CARVAJALINO et al.: COMPARATIVE STUDY OF SEMI-IMPLICIT SCHEMES

1311

TABLE II
CLASSIFICATION ACCURACIES, INDIAN PINES IMAGE

TABLE III
CLASSIFICATION ACCURACIES, CUPRITE IMAGE

very unstable when the region of the training samples is too uniform. This effect affects more the Cuprite and False Leaves im0.015) than in the
ages given that the smoothing is higher (
0.012) and also because
case of the Indian Pines image (
these images have more bands.
On the other hand, the FLL classifier benefits from the reduction in the variability within the image classes [31], and, hence,
it has the highest classification accuracies on all the images.
ECHO classifier is based in Multispec on either a quadratic
or Fisher linear spectral-spatial algorithm. The results indicated
on Tables IIIV for ECHO correspond to the maximum value
between the two possible classifiers, which was almost always
FLL for the smoothed images and quadratic for the original images. In general, ECHO is just a little superior than FLL in classification accuracy. The difference between ECHO and FLL reduces as the smoothing increases, as can be appreciated on Ta0.015, meanwhile the difference
bles III and IV, where
0.012. This is
is higher in the Indian Pines image, where
due to the fact that ECHO tries to homogenize the image before classifying it, by choosing a small window (2 2 pixels in
our simulations). Hence, if the region within the objects is already smooth, due to diffusion, the difference between ECHO
and FLL is reduced.
The remaining classifiers, ED, SAM, and MF are, in general,
very insensitive to the scale step, but, in general, they do not
achieve good classification accuracies, except for the SAM classifier on the Cuprite image. The relative good performance of

SAM on this image agrees with the reported studies on mineral classification using the spectral angle and a high number of
bands [33].
In terms of the implemented numerical methods, AOS and
achieving high speedups and
ADI are very insensitive to
on all the images
classification accuracies up to
analyzed here. On the other hand, the DouglasRachford and
PeacemanRachford methods are sensitive to the scale step,
,
achieving high classification accuracy only up to
which limits their speedup. Despite of this limitation, these
schemes are of higher accuracy than AOS and ADI and they
achieved the highest classification accuracies on all the images.
PCG methods are very insensitive to the scale step and all
behave similarly in terms of classification accuracy. The best
classification accuracies and speed-ups are achieved by PCGCholesky initialized by ADI-LOD. These results are similar to
the ones obtained in terms of the accuracy of the computed solution (Fig. 4). This means that the accuracy of the computed
solution affects the classification accuracy in the case of ML,
FLL, and ECHO classifiers.
It is noteworthy, though, that AOS has a better performance
than the expected from Fig. 4, for the Indian Pines image. We
believe that this occurs because K is small here; hence, the magnitude of the error is lower. AOS is also symmetric, and, hence,
the error introduced can be reduced by a classifier as ECHO,
which tends to average out random variations in a small window.
It is also fortunate that the Indian Pines image consists of large

1312

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

TABLE IV
CLASSIFICATION ACCURACIES, FALSE LEAVES IMAGE

Fig. 9. Indian Pines classification map: (a) original; (b) smoothed.

Fig. 10. Cuprite classification map: (a) original; (b) smoothed.

patches of uniform regions allowing that ECHO obtains a greater


reduction on the artifacts introduced by large scale steps. Since
ADI-LOD is not symmetric, the artifacts introduced are not reduced by ECHO. Otherwise, the fortunate situation for AOS in
the Indian Pines image does not applies for the other real images, and, hence, its performance is not as good there.
Otherwise, from Tables IIIV, we can see that the maximum
speedup reduces as the complexity of the image increases, in
terms of higher object intravariability. The Cuprite and False
Leaves images are more complex than the Indian Pines image,
because the Cuprite image is not a patchy image as it happens
with the Indiana Pines image, and we add a nondepreciable
amount of noise to the False Leaves image. This higher complexity was dealt in our case by increasing the value of , which
means that there is more diffusion in the image, and, hence, for
a given , the change is higher with respect to the original image
and the errors in the computed solution (artifacts) are of higher
magnitude, affecting more the classification accuracy and reducing the speedup that we can achieve on those images.
In order to see the effect of nonlinear smoothing on the classification of the full real hyperspectral images used here, in
Figs. 911, we present the classification maps of the original and
smoothed images that achieved the highest classification accuracies on Tables IIIV. It can be noticed from Figs. 10 and 11
that not only the testing samples improve their classification accuracy, but that the smoothed images also produce classification
maps that look more accurate.

Fig. 11. Fake Leaves classification map: (a) noisy; (b) smoothed.

V. CONCLUSION
PDE-based methods for image enhancement, segmentation, and restoration have a large history of success for scalar
and color images in computer vision, but it has been disregarded in segmentation and classification of hyperspectral
imagery. Recently, Lennon et al. [34], [35] implemented the
PeronaMalik nonlinear diffusion equation to smooth a hyperspectral image and classify it using support vector machines.
However, they used the original, unregularized explicit scheme
of PeronaMalik, given in (1) and used only 17 bands.
This work shows that PDE-based image processing methods
can improve significantly image enhancing, segmentation, and
classification in hyperspectral imagery at a low computational

DUARTE-CARVAJALINO et al.: COMPARATIVE STUDY OF SEMI-IMPLICIT SCHEMES

cost, using semi-implicit schemes. Traditional statistical classification methods are very robust at low-dimensional spaces, but
they require an enormous amount of data at higher dimensional
spaces, as is the case of hyperspectral imagery. Otherwise, parabolic PDEs offer a well-sounded, common framework to perform image smoothing, restoration and object-based segmentation and classification, with accuracy and highly parallelizable discretizations that can speedup PDE image processing in
high-dimensional spaces.
In particular, this work shows that nonlinear diffusion can enhance significantly image classification accuracies by reducing
both, the spatial and spectral variability in hyperspectral imagery. AOS and ADI semi-implicit schemes offer high performance in terms of accuracy and speedup of the computed solution of the nonlinear PDE, when the complexity of the image
is not high in terms of highly variability within the image objects. When the complexity of the image increases, more accurate methods such as the Douglas and Peaceman schemes and
PCG methods can achieve accuracies and speedups superior to
the less accurate AOS and ADI-LOD methods, justifying their
higher computational cost.
PCG linear solvers are less sensitive to the scale step as the
approximated ADI and AOS schemes, which mean that higher
can be used. However, PCG methods also require
values of
more space, and finding a good preconditioner is still an art as
we could corroborate here. In fact, PCG-methods also depend
on the size of the image, making it difficult to generalize them
to all image sizes. Even though image complexity can reduce
sensibly the speedup that can be achieved with the numerical
methods presented here; we achieved significant speedups of
10 and higher on all the images used, over the explicit scheme,
which justifies their use in hyperspectral imagery.
REFERENCES
[1] G. Wilkinson, Are remotely sensed image classification techniques
improving? Results of a long term trend analysis, presented at the
IEEE Workshop on Advances in Techniques for Analysis of Remotely
Sensed Data, Oct. 2728, 2003.
[2] D. A. Landgrebe, Hyperspectral image data analysis, IEEE Signal
Process. Mag., vol. 19, no. 1, pp. 1728, Jan. 2002.
[3] R. L. Ketting and D. A. Landgrebe, Classification of multispectral
image data by extraction and classification of homogeneous objects,
IEEE Trans. Geosci. Electron., vol. GE-14, no. 1, pp. 1926, Jan.
1976.
[4] A. Marangoz, M. Oruc, and G. Buyuksalih, Object-oriented image
analysis and semantic network for extracting the roads and buildings
from IKONOS pan-sharpened images, Int. Arch. Photogramm. Remote Sens., vol. 35, no. B3, 2004.
[5] W. Tadesse, T. L. Coleman, and T. D. Tsegaye, Improvement of land
use and land cover classification of an urban area using image segmentation from Landsat ETM+ data, presented at the Proc. 30th Int. Symp.
Remote Sensing of the Environment, Honolulu, HI, 2003, PS I-35.
[6] J. Schiewe, L. Tufte, and M. Ehlers, Potential and problems of multiscale segmentation methods in remote sensing, GeoBIT/GIS J. Spatial
Inf. Decision Making, vol. 6, pp. 3439, 2001.
[7] U. C. Benz, P. Hofmann, G. Willhauck, I. Lingenfelder, and M.
Heynen, Multi-resolution, object-oriented fuzzy analysis of remote
sensing data for GIS-ready information, J. Photogramm. Remote
Sens., vol. 58, pp. 239258, Jan. 2004.
[8] T. Sziranyi, I. Kopilovic, and B. P. Toth, Anisotropic diffusion as a
preprocessing step for efficient image compression, in Proc. 14th Int.
Conf. Pattern Recognition, 1998, vol. 2, pp. 15651567.
[9] L. Alvarez, J. Weickert, and J. Sanchez, Scale-space approach to nonlocal optical flow calculations, in Proc. 2nd Int. Conf. Scale-Space
Theories in Computer Vision, 1999, vol. 1682, pp. 235246.

1313

[10] I. Pollak, A. S. Willsky, and Y. Huang, Nonlinear evolution equations


as fast and exact solvers of estimation problems, IEEE Trans. Signal
Process., vol. 53, no. 2, pp. 484498, Feb. 2005.
[11] D. Tschumperle and R. Deriche, Vector-valued image regularization
with PDEs: A common framework for different applications, IEEE
Trans. Pattern Anal. Mach. Intell., vol. 27, no. 4, pp. 506517, Apr.
2005.
[12] B. Fischer and J. Modersitzki, Fast diffusion registration, in Contemporary Mathematics: Inverse problems, Image Analysis and Medical
Imaging. Philadelphia, PA: AMS, vol. 313, pp. 117129.
[13] R. Malladi and J. A. Sethian, A unified approach to noise removal,
edge enhancement, and shape recovery, IEEE Trans. Image Process.,
vol. 5, no. 11, pp. 15541568, Nov. 1996.
[14] L. Alvarez, F. Guichard, P. L. Lions, and J. M. Morel, Axioms and fundamental equations of image processing, Arch. Rational Mech., vol.
123, pp. 199257, Sep. 1993.
[15] P. Perona and J. Malik, Scale-space and edge detection using
anisotropic diffusion, IEEE Trans. Pattern Anal. Mach. Intell., vol.
12, no. 7, pp. 629639, Jul. 1990.
[16] J. Weickert, Anisotropic diffusion in image processing, Ph.D. dissertation, Univ. Kaiserslautern, Kaiserslautern, Germany, 1996.
[17] L. Alvarez, P. L. Lions, and J. M. Morel, Image selective smoothing
and edge detection by non-linear diffusion, SIAM J. Numer. Anal., vol.
29, no. 1, pp. 182193, Jun. 1992.
[18] L. Alvarez, P. L. Lions, and J. M. Morel, Image selective smoothing
and edge detection by non-linear diffusion II, SIAM J. Numer. Anal.,
vol. 29, no. 3, pp. 845866, June 1992.
[19] J. Weickert, B. M. ter Haar Romeny, and M. A. Viergever, Efficient
and reliable schemes for nonlinear diffusion filtering, IEEE Trans.
Image Process., vol. 7, no. 3, pp. 398410, Mar. 1998.
[20] J. C. Strikwerda, Finite Difference Schemes and Partial Differential
Equations. Philadelphia, PA: SIAM, 2004.
[21] D. Barash, T. Schlik, M. Israeli, and R. Kimmel, Multiplicative operator splittings in nonlinear diffusion: From spatial to multiple steps, J.
Math. Imag. Vis., vol. 19, pp. 3348, Jul. 2003.
[22] R. Barret, Templates for the Solution of Linear Systems: Building
Blocks for Iterative Methods. Philadelphia, PA: SIAM, 1994.
[23] Y. Saad, Iterative Methods for Sparse Linear Matrices, 2nd
ed. Philadelphia, PA: SIAM, 2003.
[24] P. Castillo and Y. Saad, Preconditioning the matrix exponential
operator with applications, J. Sci. Comput., vol. 13, no. 3, Sep.
1998.
[25] S. Di Zenzo, A note on the Gradient of a multi-image, Comput. Vis.,
Graph., Image Process., vol. 33, no. 1, pp. 116125, Jan. 1986.
[26] G. Sapiro, Geometric Partial Differential Equations and Image Analysis. Cambridge, U.K.: Cambridge Univ. Press, 2001.
[27] M. J. Black, G. Sapiro, D. H. Marimont, and D. Heeger, Robust
anisotropic diffusion, IEEE Trans. Image Process., vol. 7, no. 3, pp.
421431, Mar. 1998.
[28] F. Voci, S. Eiho, N. Sugimoto, and H. Sekiguchi, Estimating the gradient threshold in the PeronaMalik equation, IEEE Signal Process.
Mag., vol. 21, no. 3, pp. 3965, May 2004.
[29] P. Mrzek and M. Navara, Selection of optimal stopping time for nonlinear diffusion filtering, Int. J. Comput. Vis., vol. 52, no. 23, pp.
189203, MayJun. 2003.
[30] R. N. Clark, Imaging spectroscopy: Earth and planetary remote
sensing with the USGS Tetracorder and expert systems, J. Geophys.
Res., vol. 108, no. E12, pp. 5-15-44, Dec. 2003.
[31] D. A. Landgrebe, Signal Theory Methods in Multispectral Remote
Sensing. Hoboken, NJ: Wiley, 2003.
[32] E. Arzuaga-Cruz, A MATLAB toolbox for hyperspectral image analysis, in Proc. IEEE Int. Geosci. Remote Sens. Symp., 2004, vol. 7, pp.
48394842.
[33] G. Girouard, A. Bannari, A. El Harti, and A. Desrochers, Validated
spectral angle mapper algorithm for geological mapping: Comparative
study between quickbird and landsat-TM, in Proc. 20th ISPRS Congress: Geo-Imagery Bridging Continents, Istanbul, Turkey, 2004, pp.
599604.
[34] M. Lennon, G. Mercier, and L. Hubert-Moy, Nonlinear filtering of
hyperspectral images with anisotropic diffusion, in IEEE Int. Geosci.
Remote Sens. Symp., 2002, vol. 4, pp. 24772479.
[35] M. Lennon, G. Mercier, and L. Hubert-Moy, Classification of
hyperspectral images with nonlinear filtering and support vector
machines, in IEEE Int. Geosci. Remote Sens. Symp., 2002, vol. 3,
pp. 16701672.

1314

Julio M. Duarte-Carvajalino (S07) received the


B.S.E.E. degree (cum laude) from the Universidad
Industrial de Santander (UIS), Colombia, in 1995,
and the M.Sc. degree in electric engineering from
the University of Puerto Rico, Mayagez (UPRM),
in 2003. He is currently pursuing the Ph.D. degree
in the Computing and Information Sciences and
Engineering doctorate program, UPRM.
From 1995 to 1998, he worked in electric construction. From 1998 to 2000, he was an Assistant
Professor at the Universidad Tecnolgica de Bolivar,
Colombia. His research interests are in computer vision, especially image
processing using PDEs.
Prof. Duarte-Carvajalino was awarded with the National Science FoundationEPSCOR fellowship for three consecutive years, from 2003 to 2007. He was
also included in the National Deans List of 2003/2004 and 2004/2005 and the
Chancellor List of 2004/2005 and 2005/2006.

Paul E. Castillo received the licence degree in


mathematics from UNAH, Tegucigalpa, Honduras,
in 1988, and from USTL Montpellier II, France,
in 1989; the M.Sc. degree in computational mathematics from the University of Puerto Rico, Mayagez
(UPRM), in 1995; and the M.Sc. degree in computer
science and the Ph.D. degree in scientific computation from the University of Minnesota, Minneapolis,
in 2001.
He was a Postdoctorate at Lawrence Livermore
National Laboratory, Livermore, CA, from 2001
to 2003, where he worked in the development of a high-order finite-element
code (FEMSTER). In 2003, he joined the Department of Mathematical Sciences, UPRM. His research interests include numerical analysis, in particular,
discontinuous Galerkin methods, adaptive finite-element techniques, and the
development of mathematical software for solving physical problems.

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 5, MAY 2007

Miguel Velez-Reyes (S81M92SM00) received


the B.S.E.E. from the University of Puerto Rico,
Mayagez (UPRM), in 1985, and the M.Sc. and
Ph.D. degrees from the Massachusetts Institute
of Technology, Cambridge, in 1988 and 1992,
respectively.
In 1992, he joined the faculty of the UPRM where
he is currently a Professor. He has held faculty internship positions with AT&T Bell Laboratories, Air
Force Research Laboratories, and the NASA Goddard Space Flight Center. His teaching and research
interests are in the areas of model-based signal processing, system identification, parameter estimation, and remote sensing using hyperspectral imaging.
He has authored over 60 publications in journals and conference proceedings.
He is the Director of the UPRM Tropical Center for Earth and Space Studies,
a NASA University Research Center, and the Associate Director of the Center
for Subsurface Sensing and Imaging Systems, a National Science Foundation
Engineering Research Center lead by Northeastern University.
Dr. Velez-Reyes was one of 60 recipients from across the United States and
its territories of the Presidential Early Career Award for Scientists and Engineers
(PECASE) from the White House in 1997. He is a member of the Academy of
Arts and Sciences of Puerto Rico and a member of the Tau Beta Pi, Sigma Xi,
and Phi Kappa Phi honor societies.

Vous aimerez peut-être aussi