You are on page 1of 12

IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO.

5, OCTOBER 2008 755

Dynamic Tomographic Imaging of the Solar Corona


Mark D. Butala, Student Member, IEEE, Farzad Kamalabadi, Member, IEEE, Richard A. Frazin, and Yuguo Chen

Abstract—Three-dimensional (3-D) maps of the electron density to take full advantage of all available information to ultimately
and temperature in the solar atmosphere can be tomographically address open stellar astrophysical questions. In particular, we
reconstructed from two-dimensional images that are measured develop signal processing methods to empirically reconstruct
by a variety of ground-based and space-based instruments. The
electron density and temperature of the solar corona are funda- the 3-D time-varying structure of the solar corona.
mental parameters for understanding the physical mechanisms The corona [2] is the very sparse plasma that forms the upper
that contribute to space weather, or physical phenomena that atmosphere of the Sun, beginning at about 2000 km above the
can, in extreme cases, have adverse effects on Earth and the solar surface. The corona is extremely hot, with temperatures
near-Earth space environment. New signal processing methods in excess of . Despite more than 50 years of research,
are required to take full advantage of the rich and complex suite
of observations that are available from the current generation of the non-thermal processes that heat the corona to such an ex-
Sun-observing spacecraft. In particular, this paper provides signal treme temperature over the span of the relatively thin
models and corresponding optimal state estimation methods for transition region remain unknown [2], though it is clear that the
reconstructing the dynamic 3-D electron density and temperature solar magnetic field provides a conduit for energy transport and
structures in the solar corona with focus on the many challenges storage as well as several possible mechanisms for energy re-
associated with solar tomography.
lease. The solar magnetic field is also central in driving the solar
Index Terms—Astronomy, image processing, Kalman filtering, wind, coronal mass ejections (CMEs; occasional violent solar
statistics, stochastic systems, tomography. explosions originating in the corona), and accelerating solar en-
ergetic particles that all contribute to space weather. The solar
magnetic field also has a strong influence on the structure of the
I. INTRODUCTION corona, a plasma with low , the ratio of plasma to magnetic
pressure. As a consequence, the solar magnetic field traps the
structures of the corona.
This paper addresses the problem of reconstructing the dy-
UCH is still unknown regarding some of the funda-
M mental physical processes of the Sun. For example,
long-standing questions remain regarding the physical mech-
namic 3-D electron density and temperature of the solar corona
from remotely sensed 2-D images. Under the assumption of
quasi-neutrality, the mass density may also be reconstructed
anisms that give rise to many of the geo-effective phenomena because it is approximately proportional to the electron den-
collectively referred to as space weather, that may in ex- sity. These dynamic 3-D reconstructions will contribute to the
treme cases, damage satellites in Earth’s orbit, cause power understanding of fundamental solar-terrestrial plasma physics
distribution failures, or even harm astronauts [1]. To a large and space weather. First, the reconstructions will aid in deter-
degree, our knowledge of the solar environment has been mining the mechanisms that heat the corona and drive the solar
obtained empirically from a variety of dedicated space-based wind. The spatial variation of electron density in the corona is
and ground-based instruments. The current generation of a fundamental parameter for all processes thought to contribute
Sun-observing spacecraft generate a rich and complex suite of to heating because many waves propagate at the Alfvén speed
observations that provide the scientific community with mea- and all wave-particle interaction rates are proportional to the
surements at unprecedented spatial and temporal resolution and local value of the electron density as discussed in [3] and ref-
spectral diversity. New signal processing methods are required erences therein. Empirical determination of the temperature in
the corona as a function of position is important for inferring
where and how energy is deposited in the coronal plasma and is
Manuscript received February 01, 2008; revised July 16, 2008. Current ver- crucial to the understanding of coronal heating. Secondly, the
sion published December 10, 2008. This work was supported in part by the empirical reconstructions of the corona will aid in modeling
National Science Foundation under Grants 055561 SHINE and 0620550 CMG
and by NASA under Grant NNG-04GG32G to the University of Illinois. The the early development and propagation of CMEs. The structure
associate editor coordinating the review of this manuscript and approving it for of the solar corona strongly affects the propagation of a CME
publication was Dr. Amir Leshem.
M. D. Butala is with the Remote Sensing and Space Sciences Group, Coordi-
[4]–[6], the amount of plasma it accumulates in its propagation,
nated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, and its acceleration [7]. Lastly, 3-D tomographic models will
IL 61801 USA (e-mail: butala@uiuc.edu). contribute to modeling the shock acceleration of energetic par-
F. Kamalabadi is with the Department of Electrical and Computer Engi-
neering and the Coordinated Science Laboratory, University of Illinois at
ticles by CMEs. The density and magnetic field strength deter-
Urbana-Champaign, Urbana, IL 61801 USA. mine the compression ratio for CME-driven shock waves, which
R. A. Frazin is with the Department of Atmospheric, Oceanic and Space Sci- in turn determines the time-dependent energy spectrum of shock
ences, University of Michigan, Ann Arbor, MI 48109 USA. accelerated particles. The impact of the density is particularly
Y. Chen is with the Department of Statistics, University of Illinois at Urbana-
Champaign, Champaign, IL 61820 USA. great near the Sun where it is believed that particles are accel-
Digital Object Identifier 10.1109/JSTSP.2008.2005352 erated to the highest energies [8].
1932-4553/$25.00 © 2008 IEEE
756 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

In principle, tomographic analysis [9], [10] may be used to re- both general [36] and cardiac [42] tomography. However, solar
construct the 3-D corona given image measurements (2-D pro- tomography has many aspects that challenge data assimilation,
jections) from multiple point-of-views (POVs) obtained by solar including the dimensionality of the problem, the tomographic
rotation or from spatially distinct sensors [11]–[14]. Until re- nature of the measurements that are corrupted by Poisson
cently, measurements of the Sun have been available from only noise, nonlinear dynamics described by the equations of MHD,
one POV at any given time and tomographic analysis of the and the existence of additional prior knowledge, such as the
corona has relied on solar rotation to provide the necessary mea- empirical structure of the solar magnetic field, that could also
surement diversity. With a synodic period of 27.3 d, almost 14 be incorporated into the analysis. This paper focuses on the
d are required to observe the corona from a single POV over the aforementioned state estimation methods and, specifically, the
180 necessary for a full tomographic analysis. The situation degree to which each addresses the challenges of dynamic solar
has recently changed with the launch of the Solar and Terrestrial tomography.
Relations Observatory (STEREO) [15] in October, 2006. The The remainder of the paper is organized in the following
mission provides two additional simultaneous POVs from dual manner. Section II develops the forward models that relate the
satellites, one that leads Earth’s orbit by 22 and the other remotely sensed measurements to the electron density and tem-
that lags by the same amount. However, almost 5 d are still re- perature in the corona. Next, methods for solving the resul-
quired to view the Sun over 180 when the STEREO satellites tant time-invariant inverse problem are discussed in Section III
are each separated from the Earth by 60 . for later contrast with dynamic methods. Then, the linear state-
The first reconstructions of the 3-D structure of the electron space model is developed in Section IV, as well as the Kalman
density in the solar corona were calculated under the assumption filter and ensemble Kalman filter. Lastly, the general hidden
of azimuthal symmetry for analytical tractability [16]. Today, Markov state-space model and particle filter are developed in
it is possible to reconstruct full time-independent 3-D maps of Section V. Concluding remarks are given in Section VI.
electron density given contemporary computational resources
and signal processing methods [13], [14], [17]. A tomographic II. FORWARD PROBLEM
method for reconstructing the temperature in the corona has
also been recently developed [18]. Most work devoted to tomo- A. Electron Density
graphic reconstruction of the solar corona has so-far assumed The physical relationship between the intensity of remotely
that it is unchanging during the measurement interval. Under observed polarized light and the density of free electrons in the
this assumption, it is possible to reliably reconstruct persistent corona integrated along the measurement line-of-sight (LOS) is
structures in the corona because the large-scale morphology re- given by [16], [43]
mains fairly constant. However, significant reconstruction arti-
facts may result because the corona is very dynamic and exhibits (1)
variation on all observed temporal and spatial scales [17], [19].
Dynamic tomography methods have been developed for
biomedical imaging applications as reviewed in [20]. In this The function is the polarized brightness (pB) measured
context, research has focused on techniques to image the at time from a POV defined by , a 3-D unit vector parallel
quasi-periodic beating heart [21]–[23] and for mitigating to the LOS, and , a 3-D vector perpendicular to that extends
rigid-body patient motion during the scan interval [24]–[26]. from the origin at the center of the Sun to the LOS. The function
However, these motion models cannot adequately account for is the electron density at a specified point in the
the complicated evolution of the solar atmosphere described stellar atmosphere at a distance along the LOS and the set
by the equations of magnetohydrodynamics (MHD) [27], [28] is the range of corresponding to the ray from the observer
that govern the physics of electrically conducting fluids. Some to infinity. The scalar is a combination of several physical
research has been devoted to dynamic solar tomography [29], constants necessary to balance the physical units of the equation.
but not in the framework of signal processing. The Thomson scattering physics are encapsulated in the known
This paper develops data assimilation methods for re- function [16], [43]. Lastly, the function
constructing the time-evolving 3-D structure of the solar is the instrument dependent measurement noise.
atmosphere. Data assimilation is a general systematic math- The statistics of the measurement noise are Poisson,
ematical framework for reconstructing a dynamic object that given that scattering governs the emission of pB light from the
combines empirical measurements with first-principles physical corona and the light sensors are most often charged coupled-de-
knowledge of the object’s dynamics. The corresponding signal vices (CCDs). In practice, the true statistics of the measurement
processing methods fully embrace the stochastic nature of the noise may only be crudely specified given that coronagraphs are
problem and balance uncertainties in the noisy measurements very complicated optical instruments and various systematic ar-
with uncertainties in the inevitably incomplete dynamical tifacts may persist even after calibration. Often, only the vari-
model. The Kalman filter [30], [31], ensemble Kalman filter ance of the noise for each pixel is known.
[32]–[36], and particle filter [37]–[41] are state estimation Measurements of pB are typically in the form of images. Each
methods suitable for data assimilation and each offers tradeoffs pixel of such an image is the intensity of polarized light focused
between model restrictions and computational complexity. The on a sensor, usually a cell of a CCD camera. It will later be
application of these methods to data assimilation is not novel. convenient to represent such an image by the vector , with
For example, the ensemble Kalman filter has been applied to each of its elements equal to a single pixel intensity of the
BUTALA et al.: DYNAMIC TOMOGRAPHIC IMAGING OF THE SOLAR CORONA 757

The choice of the basis functions has a significant impact on


the reconstructions as all features in the solar atmosphere that
cannot be expressed as a linear combination of basis functions
are lost. As an example, compare the very similar, but distin-
guishable spherical (Fig. 3) and cylindrical (Fig. 6 in [17]) voxel
basis reconstructions computed from the same data. The unfor-
tunate byproduct of a voxel basis is that will typically be im-
mense. For example, dividing the reconstructed image domain
into results in . In any case,
it is possible to systematically determine the number of neces-
Fig. 1. Two images of polarized brightness as measured by the Mk4 corona-
sary basis elements through model-order selection [50] or as
R
graph. The inner and outer radii of the images range from 1.1–2.85 , where a hyper-parameter in the estimation process, but at potentially
R 2
is the solar radius equal to 7 10 km. The left and right images were great computational expense.
measured on Oct. 22 and Oct. 29 of 2007.
The finite basis expansion results in the linear system

(4)
pB image. The vector is related to the continuous forward
model (1) through where the measurement operator is an matrix. Each
element of the matrix is given by
(2)
(5)

where is the th element of the vector argument and


is the set of lines-of-sight that are focused onto the th image where is the th component of the matrix argument
pixel. and the symbol should be read “is defined as.” Note that the
Several coronagraphs, specialized telescopes for measuring measurement operator can be broken into two more easily un-
the dim corona, regularly measure pB images suitable for this derstood operators . The diagonal matrix
work. One example is the ground-based Mark-IV (Mk4) coron- accounts for the spatial dependence of the scattering function in
agraph [44] at the Mauna Loa Solar Observatory (MLSO). Two (1). The th row of the matrix is a weighted sum
pB images measured by the Mk4 are shown in Fig. 1. Suitable over a set of elements of and is a discrete representation of
space-based coronagraphs include the Large Angle and Spectro- the line integrals in the continuous forward model (1). As dis-
metric Coronagraph Experiment (LASCO) C2 instrument [45] cussed above, the full statistics of the measurement noise vector
on-board the Solar and Heliospheric Observatory (SOHO) [46] may not be known.
and the COR1 and COR2 coronagraphs available as part of the
Sun Earth Connection Coronal and Heliospheric Investigation B. Temperature
(SECCHI) suite of instruments [47] on both STEREO satel-
Differential emission measure (DEM) analysis is concerned
lites. The best quality measurements of pB are from space-based
with the plasma temperature distribution in the solar atmos-
coronagraphs that are devoid of the sky-scattered light that can
phere. Suitable measurements for DEM analysis are at extreme-
plague terrestrial coronagraph measurements.
ultraviolet (EUV) and X-ray wavelengths. DEM tomography
Closed-form analytical solutions to Fredholm integral equa-
(DEMT) [18] combines DEM and tomographic analyses to re-
tions like (1) are possible only when the kernel is of a
construct the 3-D electron temperature distribution. The forward
simple form. However, the Thomson scattering kernel in (1) is
model is
too complicated and the forward model must be discretized to
approximate the solution through numerical techniques. A gen-
eral expression for the necessary finite basis expansion is

(6)
(3)
Identical to Section II-A, the vectors and define the POV
of the observer relative to the Sun, the vector specifies a
where is the 3-D position vector, is the total number of basis 3-D spatial location along an individual measurement LOS, and
functions, is the th basis function, and the vector has the scalar ensures that physical units balance. The function
components. The corona is an amorphous plasma with un- is the EUV or X-ray intensity measured at wave-
known correlation structure and it is therefore difficult to form a length and time and is the measurement noise.
low-dimensional basis that faithfully captures the corona. How- The function is the local DEM at temperature
ever, research exists in characterizing the corona with consis- . The function depends on the atomic physics of the
tent, but time independent, low-dimensional models [48], [49]. emission process in the stellar plasma and may be calculated
Faced with such difficulties, a voxel basis, an orthogonal set of with tools such as CHIANTI [51].
3-D volume elements that span the reconstructed image domain, Like pB, typical measurements of the EUV or X-ray corona
represents a general choice that makes few prior assumptions. are in the form of images. The pixels of an EUV or X-ray
758 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

where is the 3-D spatial coordinate, is an ma-


trix with th component equal to the local DEM at the
th spatial coordinate and th temperature, is the th
spatial basis function, and is the th temperature basis
function. As argued in Section II-A, it is difficult to determine
a faithful low-dimensional basis for the 3-D structure of the
coronal electron density, and the same is true for the local DEM.
DEMT research has so far focused on voxel spatial and uniform
interval temperature bases [18].
The measurements are related to the elements of the un-
known matrix by

(9)

where the th element of the matrix is the


measurement noise of the th pixel in the th frequency band.
Fig. 2. Four images of the Sun to 1.5 R as measured by the EIT on SOHO.
Note that the terms in the sum over in (9) only operate on the
columns of the matrix for a fixed row. Similarly, the terms
Three of the images were measured on Jan. 11, 2008 at different EUV wave-
lengths: (top left) at 171 
A, (top right) at 195 
A, and (bottom left) at 284 
A. The in the sum over not appearing in the sum over operate
fourth image is at 284  A and was measured on Dec. 15, 2007, 27 days prior to only on the rows of the matrix for a fixed column. Thus, it is
the previous images. Differences between the images in the bottom row are due
to solar dynamics because the Sun rotates 360 =27 d relative to the Earth. possible to separate these two operations. To do so, first define
each element of the matrix by

image at a particular wavelength can be represented by the (10)


vector
and note that the matrix corresponds to the column-wise op-
(7) erations on in (9). Also note the function may only
be evaluated at discrete points by CHIANTI and the integrals in
(10) must be evaluated numerically. Secondly, define the
where the th element of the th column of the matrix matrix by
is the th pixel in the th frequency band measured at time
. Again, is the set of lines-of-sight that are focused onto the (11)
th pixel. The function is the sensitivity or frequency
response of the instrument in the th frequency band.
The matrix corresponds to row-wise operations on in (9)
Measurements of the corona in the EUV are routinely mea-
and is the same discretized representation of line integrals in
sured by dedicated space-based instruments. Ground-based
Section II-A. Finally, (9) can be compactly expressed as
observations of the corona in the EUV are not possible because
Earth’s atmosphere absorbs EUV emissions. Relevant instru- (12)
ments include the Extreme Ultraviolet Imaging Telescope (EIT)
[52] on SOHO, the Extreme Ultraviolet Imaging Spectrometer A matrix identity involving the Kronecker matrix product [57]
(EIS) [53] on Hinode [54], the EUV imager as part of the may be applied to (12) to give
SECCHI instruments [47] on STEREO, and the Atmospheric
Imaging Assembly (AIA) [55] scheduled for launch on the (13)
Solar Dynamics Observatory (SDO). Four images of the EUV where the operator stacks the columns of the matrix argu-
corona measured by the EIT are shown in Fig. 2. X-ray mea- ment into a vector and is the Kronecker matrix product. Note
surements are available, for example, from the X-ray Telescope that (13) is now in the form of a linear system of equations, cf.
(XRT) [56] also on Hinode. (4).
The temperature forward model (6) may be discretized by
choosing both a spatial and temperature basis III. STATIC INVERSE PROBLEM
The forward models developed in Section II each relate an
(8) unknown quantity of interest, either electron density or temper-
ature in the solar corona, to the available measurements. The
BUTALA et al.: DYNAMIC TOMOGRAPHIC IMAGING OF THE SOLAR CORONA 759

associated inverse problem is to then reconstruct a 3-D map of


the solar atmosphere given a set of 2-D images. The reconstruc-
tion method is determined by the overall signal model and the
choice of a suitable optimality objective. The least sophisticated
and computationally demanding is the static model developed
below.

A. Static Signal Model


In the static framework, the temporal variation of the corona
is assumed to be negligible over the measurement interval at the
reconstructed spatial scales. The given measurements in both
cases are samples of the discretized measurement function
denoted by , where is an integer time index and
is a function that maps the time indices to sample times. The
resultant static signal model is Fig. 3. Spherical cut at 1.18 R of a 3-D static reconstruction of the coronal
electron density with units log (cm ). The x and y axes are in units of solar
R
radii ( ). The reconstruction is based on 11 images measured by the Mk4

.. .. .. coronagraph when weather permitted from Oct. 26 to Nov. 8, 2003.


. . . (14)

shown in Fig. 3. Without the second term , (15) is a stan-


where the unknown electron density or temperature has no
dard least squares fit to the data . The second term is a Tikhonov
time dependence. The solar atmosphere is reconstructed by es-
regularization [60]–[62] penalty and is included to stabilize the
timating the properties of the element random vector given
ill-posed solar tomography problem. For solar tomography, the
the element vector of measurements , matrix is some form of derivative operator [14], [17], [18] and
the forward model matrix , the regularization term is a roughness measure where in-
and all known statistical information regarding the ele- creases when is less smooth. From a physical standpoint, the
ment measurement noise vector . electron density and temperature in the corona will, in general,
not vary drastically across small spatial distances and a rough-
B. Static Inverse Problem ness penalty is appropriate as it will dampen large, short dis-
There are several challenges in solving the system of linear tance variations in the reconstructed 3-D maps at the cost of
equations (14). For one, it is well-known [10] that tomographic smoothing the overall reconstruction to some degree. The reg-
inverse problems are ill-posed, meaning direct inversion of the ularization parameter controls the tradeoff between agreeing
linear system (14) may result in significant reconstruction ar- with the data and favoring a smooth reconstruction. Cross val-
tifacts due to noise sensitivity. The situation is pronounced in idation [62], [63], one method to determine the regularization
solar tomography because of the low measurement sampling parameter that represents the best tradeoff based on the data
rate, due either to telemetry constraints in space-based observa- , has been successfully applied in the context of solar tomog-
tions or the averaging necessary in ground-based observations raphy [14], [17]. For electron temperature reconstruction, note
to achieve a desired signal-to-noise ratio. In practice, only a that [18] decomposes (15) into a series of two lower dimensional
single daily pB observation has been available for electron den- linear inverse problems. However, even when the electron tem-
sity reconstruction from the Mk4 and LASCO C2 coronagraphs. perature problem is decomposed, static tomography of the solar
The situation is changing with the launch of new missions, such electron density and temperature remains a demanding problem,
as the dual STEREO satellites, which offer higher spatial and requiring hours or even days of computation on a contemporary
temporal resolution measurements. The benefits from increased workstation.
data rates has been studied in [58]. Another challenge is that the The formulation (15) is deterministic, meaning any statis-
linear system (14) is of high dimension, on the order of tical knowledge of the unknown and measurement noise is dis-
[14], [17]. Complicating the situation, new high-resolution mea- regarded. In the Bayesian stochastic framework, the unknown
surements from STEREO and other future solar missions will and measurement noise are regarded as random vectors and
further increase the dimensionality of the problem by one or , each with a probability density function (pdf) and the recon-
more orders of magnitude. struction problem reduces to statistical inference or estimation
The above challenges have motivated the following method of the random vector given the data. A static estimate may be
for reconstructing persistent 3-D structures in the corona under computed by minimizing the mean squared error (MSE)
the static model [14], [17], [59]
(16)
(15)
where is called the minimum MSE (MMSE) estimate,
where is the standard vector norm with . is the statistical expectation operator, and is an arbitrary
An example of electron density reconstructed in this manner is function of the data. The MMSE estimate is a least squares fit
760 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

to the data in the stochastic sense because (16) will be casting the future evolution of the state, and smoothed estimates
as close as possible to the unknown quantity in the sense of based on a complete data-set may each be of interest.
average minimum squared difference. Predictive and smoothed estimates are usually found by further
An explicit connection can be made between the regular- processing of the filtered estimates [31], which we exclusively
ized least squares reconstruction (15) and the linear MMSE focus on below.
(LMMSE) estimate. The LMMSE estimate is the
MMSE estimate (16) that is restricted to be an affine function A. Linear Additive Noise Signal Model
of the data, or, expressed mathematically, . The linear additive noise state-space model, which assumes
The LMMSE estimate is given by [64] that the dynamics of the time evolving unknown state are linear
in nature, is summarized by
(17)
(21)
where is a vector norm when is positive (22)
definite, the unknown and noise are uncorrelated, has
mean and covariance with The discrete-time Markov random process is a
, and is zero mean with covariance . If time-series representation of the state of the solar atmosphere.
, the vector with all elements equal to 0, , the The forward model (22) is unchanged from Section III. The
state-transition model (21) is a linear stochastic description of
identity matrix scaled by , and with full
how the unknown state changes from one time index to the next.
column rank, then (15) and (17) are equivalent. In other words,
The matrix models the deterministic linear compo-
deterministic regularized least squares is equivalent to LMMSE
nent of the evolution and is derived from first principles physics.
estimation when the unknown and noise have special distribu-
The state noise is a random process that accounts
tions.
for any uncertainty in the model . The first and second-order
An important property, used later in this paper, is that (17)
statistics of the uncorrelated and zero-mean measurement and
may be solved in closed form [64]
state noise are also included as part of the above state-space
(18) model with and ,
where the symbol is the Kronecker delta function.
(19)
The dynamics of the solar atmosphere are modeled by the
In addition, the LMMSE estimator error covariance , equations of MHD. However, incorporation of MHD models
a measure of the uncertainty in , can be expressed in into (21) poses two major difficulties. First, while the MHD
closed form [64] equations are functions of the electron density and temperature,
they also depend, for example, on the energy density, vector ve-
(20) locity, and vector magnetic field at each spatial location in the
solar atmosphere. Including these additional parameters into the
where and is the signal model would significantly increase the dimensionality of
LMMSE estimate error. the already large problem. Secondly, (21) demands a linearized
representation of the MHD equations, a process that may be
IV. LINEAR DYNAMIC INVERSE PROBLEM questionable at the time scales under consideration. A less so-
The primary criticism of the approach in Section III is that phisticated model for would only include the differential ro-
the static assumption is invalid. The bottom row of images in tation of the Sun to account for the fact that the solar atmosphere
Fig. 2 demonstrates that the dynamics of the solar corona can rotates at different rates depending on latitude. A very simple
be significant over a period of a 27 d full rotation where the model that is often used in practice [65] is the random walk
differences between the two images are due to changes in the model where and the evolution of is then assumed
solar atmosphere. Unphysical smearing artifacts may result in to be entirely stochastic. Despite the simplicity of the random
the static reconstructions due to such changes [17]. Depending walk model, it can be superior to the static model [66].
on the situation, the solar atmosphere may change significantly
at time scales of even less than a day. B. Kalman Filter
In Section III, electron density and temperature estimation The Kalman filter (KF) is the recursive state estimation pro-
involved statistical inference of the unknown random vector cedure for computing LMMSE estimates of the electron den-
given the data . To mitigate the reconstruction artifacts men- sity or temperature state given the measurements and
tioned above and more directly address scientific questions re- are denoted . Only filtered estimates and one-step
garding the time-evolving solar atmosphere, the electron den- predictions are required by the KF. The KF esti-
sity and temperature are now modeled as discrete-time Markov mator error covariance is denoted where
random processes. The resulting estimation problem involves is the estimate error. Examples of KF tomo-
the inference of the random vector , the state of the solar at- graphic reconstructions of a dynamic low-dimensional object
mosphere, at each time index given the set of measurements are given in [66] and in Fig. 4. However, as discussed below,
. Depending on the situation, filtered esti- the KF is computationally feasible only when the state dimen-
mates computed in real-time, predictions fore- sion is relatively small.
BUTALA et al.: DYNAMIC TOMOGRAPHIC IMAGING OF THE SOLAR CORONA 761

The recursive KF process begins with an initial state prior and C. Ensemble Kalman Filter
prior error covariance denoted by and . Monte Carlo approximations to the LMMSE estimates given
These quantities are an initial estimate of the state and a mea- by the KF may be computed with the ensemble Kalman filter
sure of the confidence in that initial estimate, respectively. In (EnKF). The attraction of the EnKF is that it can, in certain ap-
practice, a static reconstruction often serves as the initial prior plications, provide reasonable state estimates at relatively low
state estimate [36], [66]. The KF computes the LMMSE esti- computational expense when the state dimension is large. The
mates in a two-stage process. First, in the measurement up- tradeoff is that the estimates given by the EnKF are random due
date, the one-step prediction of the state at time is corrected to the Monte Carlo nature of the algorithm, biased because of lo-
in the Bayesian sense to incorporate information from the mea- calization, and biased when is finite [67]. The biases may not
surement at time . The measurement update is be severe as demonstrated in a dynamic tomography example in
[36] and in Fig. 4, discussed below. At least in principle, these
(23) biases may be controlled by the careful application of covari-
ance tapering and ensemble inflation [67], but precise theoret-
(24)
ical analysis and practical methods are open topics of research.
(25) The idea behind the EnKF is to process an ensemble of
samples, each denoted , in such a manner that the ensemble
where the matrix is commonly referred to as the sample mean
Kalman gain. Comparing (18)–(20) to (23)–(25), note that the
KF measurement update is equivalent to the LMMSE estimate
(29)
of the unknown state based on the prior statistical information
and . The measurement update (24) can also be
expressed as the quadratic optimization and ensemble sample covariance

(26) (30)

Just as (19) is the unique minimum of (17), (24) is the unique approximate the LMMSE state estimate and estimator error
minimum of (26). From (26), each LMMSE estimate is the covariance given by the KF. Tremendous computational
balance between two terms. The first term is the measurement savings relative to the KF are possible, especially if only a small
discrepancy and is weighted by the measurement uncertainty number of ensemble members relative to the state dimension
. The second term, the prior discrepancy, is weighted by the are necessary to compute the estimate to a desired con-
uncertainty in the prior . As mentioned in [66], it is not fidence level.
difficult to add a Tikhonov penalty term to the KF measurement The EnKF is initialized by forming an ensemble of in-
update to incorporate further prior knowledge, such as smooth- dependent and identically distributed (i.i.d.) Gaussian samples,
ness, by forming an augmented linear state-space signal model each denoted
and then applying (23)–(25).
The second stage of the KF, the time update, calculates the (31)
prediction of the state based on the current LMMSE state
The EnKF, like the KF, then proceeds in a two step recursion. In
estimate and the state transition model (21). The equations
the measurement update, each sample of the prior ensemble is
for the predicted estimate and predicted estimate error covari-
updated to incorporate information from the most recent mea-
ance are
surement in the Bayesian sense. The measurement update be-
gins by computing the sample Kalman gain based on the sample
(27) covariance of the prior ensemble
(28)
(32)
These predictions are the prior information necessary for the
next measurement update step in the recursive state estimation where the element matrix is discussed in more detail
process. below and the matrix operator is the Hadamard or element-by-
The estimates computed by the KF are LMMSE optimal. element matrix product [57]. Only the matrix
However, they become difficult to compute when the state di- is required in the computation of (32) so the
mension is large. The primary computational challenge is sample covariance is never explicitly stored or operated
the storage and processing of the state estimate error upon. To form the posterior ensemble, each sample is processed
covariance matrix because will be large for solar tomog- with
raphy. For example, when the solar atmosphere is divided into
, the matrix requires over 3.6 TB (33)
of storage if each element is a 32-bit number, and an immense
amount of computational effort is required to update the matrix where . A Tikhonov penalty term can be
according to (25) and (28). added to the EnKF measurement update by augmenting the
762 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

signal model in the same way as discussed in Section IV-B


for the KF. The time update step of the EnKF simulates the
evolution of each sample from the current time index to the
next according to the state-transition model (21) with

(34)

and .
The choice of covariance taper matrix is critical for suc-
cessful application of the EnKF to high dimension problems.
Covariance tapering is necessary in practice because the sample
covariance is systematically prone to significant error in
the computationally desirable situation when the sample size
is small relative to the state dimension [68]. The tapered es-
timate is a regularized estimate of the error covari-
ance matrix that incorporates prior knowledge to dampen
such systematic errors. Typically, physical processes are not
spatially correlated past some known correlation length. This
prior knowledge should be used in the design of . A covari-
ance taper matrix with elements that are greatest near the diag-
onal and 0 some number of bands away will eliminate unphys- Fig. 4. Three dynamic tomography reconstructions of electron density based
ical long distance correlations in the regularized estimate on a MHD simulation of the emergence of magnetic flux in the solar corona. The
that are the result of systematic errors in the sample co- 2
top left image (truth) is one 32 32 frame of the simulation. The EnKF (top
right), KF (bottom left), and LKF (bottom right) reconstructions are comparable
variance . The taper matrix suggested in [69] is often used to the truth, but the EnKF required only 30 s of computation to reconstruct the
because of its properties: 1) it is easily computed; 2) it represents entire 256 frame simulation, whereas the KF and LKF each required nearly 45
m.
the prior knowledge of an exponential decay in correlation with
distance that is reasonable for many physical processes; 3) it
is sparse, which results in many further computational simpli- The LKF measurement update can also be expressed as the
fications in the implementation of the EnKF; and 4) it is posi- quadratic optimization
tive definite, which guarantees that the product is
also positive definite and a valid covariance matrix by the Schur
product theorem [57]. Analysis of the quality of tapered sample
(38)
covariance estimates is an active field of research
The LKF time update is given by
[67], [68], [70] and the best choice of for a given applica-
tion remains an open and challenging question. Some guidance (39)
is offered as an implication of the convergence of the EnKF dis-
cussed below. (40)
The EnKF converges to a well-defined limit as the number of
Comparing (26) and (38), the only difference between the mea-
Monte Carlo samples increases. As proven in [71], the EnKF
surement updates of the KF and LKF is the weighting in the
estimate without tapering (when , the matrix with
prior discrepancy term. Because the EnKF is an approximation
all elements equal to 1) converge in probability to the LMMSE
to the LKF, the same comparison may be made between the KF
estimate . The inductive proof in [71] may be extended for
and EnKF.
more general , and, in this case, the EnKF may be shown
As noted above, the EnKF estimates converge to LMMSE
to converge to the localized KF (LKF). Like the KF, the initial
estimates when . An implication is that if the product
LKF prior estimate and estimate error covariance are given by
is approximately equal to then it is reasonable
and . The LKF measurement update is to expect that the EnKF estimates will be approximately equal to
given by LMMSE estimates in the limit as the ensemble size increases.
This suggests that the choice of taper matrix is highly dependent
on the structure of the error covariance matrix and should
be designed such that .
(35) The images in Fig. 4 compare the EnKF, KF, and LKF in a
numerical experiment. The unknown state is the th 32 32
(36) frame of a 256 frame MHD simulation of the emergence of mag-
netic flux in the solar corona. The measurement is 3 sets of
46 parallel line integrals from POVs separated by 60 to mimic
the observations now available with the launch of STEREO.
(37) The POVs are at an angle that sweeps through 360 a total
BUTALA et al.: DYNAMIC TOMOGRAPHIC IMAGING OF THE SOLAR CORONA 763

of 4 times over the 256 frames. A small (roughly 1%) amount where is the posterior pdf of the state of the stellar
of white Gaussian noise has been added to the measurements atmosphere conditioned on the measurements. The simple form
( and is the noise variance). The state transition op- of (44) belies numerous challenges. For one, (44) requires the
erator is which models the dynamics as a random walk closed-form evaluation of an -dimensional integral, an ana-
and is a fixed 9-band positive definite matrix that models a lytical intractable problem except in special cases. Straightfor-
1 pixel-length temporal correlation in the state noise. The ini- ward numerical evaluation of (44) is also intractable because nu-
tial prior estimate is a static reconstruction. All three esti- merical quadrature becomes exponentially more computation-
mates of the truth (top left) image are of comparable quality. ally demanding with the state dimension . Secondly, except
The LKF estimate shows the bias introduced by covariance ta- in special cases, it is impossible to determine the posterior pdf
pering relative to the LMMSE optimal KF. The EnKF in closed form through analysis of the given pdfs
approximates the LKF, but is much less computationally expen- (41)–(43). The posterior pdf may be approximated numerically,
sive. Whereas the KF and LKF each required nearly 45 m of but this also involves -dimensional numerical quadrature.
computation to reconstruct the entire 256 frame simulation, the Computationally tractable approximate methods must be
EnKF required only 30 s. used to reconstruct the solar atmosphere under the general
hidden Markov signal model. Some research has been devoted
V. NONLINEAR DYNAMIC INVERSE PROBLEM to the application of the EnKF to problems with nonlinear
The linear additive noise state-space signal model in measurement and state transition models [34], [72], [73].
Section IV-A imposed some significant restrictions on the However, it is difficult to make definitive statements regarding
models for the dynamics of the solar atmosphere and mea- the optimality, convergence, and general statistical properties
surements. For one, only the first and second order statistics of such methods as they no longer operate within the signal
of the initial state were modeled and any additional physically model in which they were derived. The situation is similar to
motivated prior knowledge was disregarded. Additional prior the case when the KF is applied to a nonlinear problem through
knowledge could be incorporated through Tikhonov regular- linearization, a process called the extended KF [31], [74], and
ization, but it is unclear how to express information such as the optimality is lost in the process. A more general class of se-
empirical solar magnetic field as a quadratic penalty. Secondly, quential Monte Carlo methods called particle filters (PFs) take
it was impossible to incorporate a more sophisticated and a more direct approach at solving the general state estimation
realistic signal dependent Poisson measurement noise model. problem. PFs cleverly circumvent the analytically intractable
Most importantly, the nonlinear MHD equations cannot be posterior pdfs by using a method called sequential importance
directly incorporated. The general signal model developed sampling. PFs operate on samples from a model proposal dis-
below is free of these restrictions. Unfortunately, the generality tribution that, when properly combined and weighted, form an
comes at such great computational expense that existing signal approximation that converges to the posterior pdf. Except under
processing methods quickly become intractable for the high special circumstances, unfortunately, the PF is computationally
dimensional solar tomography application. tractable only for relatively low-dimensional state-estimation
problems. To date, the PF has been limited to state estimation
A. Hidden Markov Signal Model problems with only 10s or 100s of state variables [75]–[77] and
is hence currently unsuitable for solar tomography.
In the hidden Markov state-space signal model, the unknown
quantity of interest is defined as the discrete-time Markov VI. CONCLUSIONS
random process and the measurements by the
random process . The model is fully defined by the This paper has addressed the tomographic reconstruction
pdfs of the 3-D electron density and temperature structure of the
solar corona. Signal processing methods were discussed in
(41) the context of three signal models: static, linear dynamic,
and nonlinear dynamic. Each approach, a signal model and
(42)
corresponding signal processing algorithm, offers a tradeoff
(43) between computational complexity and the degree to which
the model captures the true nature of the solar atmosphere,
The pdf encompasses all statistical prior knowl-
measurements, and additional side information. Static recon-
edge regarding the initial state . The state transition pdf
struction is tractable with existing techniques, but fails to model
is a statistical description for the evolution of the
the dynamic corona. The linear dynamic framework models
state random process from one time index to the next. Finally,
the linear component of such dynamics, but current algorithms
the forward model pdf is the statistical relationship
are too computationally demanding for the dimensionality of
between a measurement and the state.
solar tomography. However, a parallel implementation of the
B. Nonlinear Dynamic Inverse Problem Methods EnKF is possible and a computer cluster may then be leveraged
to enable the reconstruction of the dynamic 3-D corona. The
A dynamic tomographic reconstruction may be computed as nonlinear framework may incorporate a full MHD model for
the MMSE estimate the state evolution to improve the reconstructions. However,
PFs are currently limited to small dimensional problems. New
(44) developments [78] and the application of covariance tapering
764 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

concepts to PFs may enable the fully general reconstruction of [23] N. Aggarwal and Y. Bresler, “Spatio-temporal modeling and adaptive
the time-evolving solar atmosphere. acquisition for free-breathing cardiac magnetic resonance imaging,” in
Proc. 2nd SIAM Conf. Imaging in Science, Salt Lake City, UT, May
2004.
ACKNOWLEDGMENT [24] J. L. Duerk and O. P. Simonetti, “Review of MRI gradient waveform
design methods with application in the study of motion,” Concepts
The authors would like to thank R. J. Hewett for carefully Magn. Reson., vol. 5, pp. 105–122, 1993.
proofreading the document. The MHD simulation of the [25] T. Ernst, O. Speck, L. Itti, and L. Chang, “Simultaneous correction
emerging magnetic flux is courtesy of Dr. W. Manchester of for interscan patient motion and geometric distortions in echoplanar
imaging,” Magn. Reson. Med., vol. 42, pp. 201–205, 1999.
the Department of Atmospheric, Oceanic and Space Sciences [26] Z.-P. Liang and P. C. Lauterbur, Principles of Magnetic Resonance
at the University of Michigan. The authors would also like to Imaging. New York: Wiley-IEEE Press, 2000.
thank the reviewers for their thorough and insightful remarks. [27] H. Alfvén, “Existence of electromagnetic-hydrodynamic waves,” Na-
ture, vol. 150, pp. 405–406, 1942.
[28] P. A. Davidson, An Introduction to Magnetohydrodynamics. Cam-
REFERENCES bridge, U.K.: Cambridge Univ. Press, 2001.
[1] D. N. Baker, “What is space weather?,” Adv. Space Res., vol. 22, pp. [29] T. Dunn, B. V. Jackson, P. P. Hick, A. Buffington, and X. P. Zhao,
7–16, 1998. “Comparative analysis of the CSSS calculation in the UCSD tomo-
[2] M. J. Aschwanden, Physics of the Solar Corona. Berlin, Germany: graphic solar observations,” Sol. Phys., vol. 227, pp. 339–353, 2005.
Springer-Verlag, 2004. [30] R. E. Kalman, “A new approach to linear filtering and prediction prob-
[3] S. R. Cranmer and A. A. van Ballegooijen, “Alfvénic turbulence in the lems,” J. Basic Eng.—Trans. ASME, vol. 82, pp. 35–45, 1960.
extended solar corona: Kinetic effects and proton heating,” Astrophys. [31] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood
J., vol. 594, pp. 573–591, 2003. Cliffs, NJ: Prentice-Hall, 1979.
[4] P. Riley, J. A. Linker, and Z. Mikić, “An empirically-driven global [32] G. Evensen, “Sequential data assimilation with a nonlinear
MHD model of the solar corona and inner heliosphere,” J. Geophys. quasi-geostrophic model using Monte Carlo methods to forecast
Res., vol. 106, pp. 15 889–15 902, 2001. error statistics,” J. Geophys. Res., vol. 99, pp. 10 143–10 162, 1994.
[5] D. Odstrcil et al., “Merging of coronal and heliospheric numerical two- [33] G. Burgers, P. J. van Leeuwen, and G. Evensen, “Analysis scheme
dimensional MHD models,” J. Geophys. Res., vol. 107, p. SSH 14-1, in the ensemble Kalman filter,” Mon. Weather Rev., vol. 126, pp.
2002. 1719–1724, 1998.
[6] W. B. Manchester, IV et al., “Modeling a space weather event from the [34] P. L. Houtekamer and H. L. Mitchell, “A sequential ensemble Kalman
Sun to the Earth: CME generation and interplanetary propagation,” J. filter for atmospheric data assimilation,” Mon. Weather Rev., vol. 129,
Geophys. Res., vol. 109, p. A02107, 2004. pp. 123–137, 2001.
[7] N. Lugaz, W. B. Manchester, IV, and T. I. Gombosi, “The evolution of [35] T. M. Hamill, J. S. Whitaker, and C. Snyder, “Distance-dependent
coronal mass ejection density structures,” Astrophys. J., vol. 627, pp. filtering of background error covariance estimates in an ensemble
1019–1030, 2005. Kalman filter,” Mon. Weather Rev., vol. 129, pp. 2776–2790, 2001.
[8] I. V. Sokolov et al., “A new field line advection model for solar particle [36] M. D. Butala, R. A. Frazin, Y. Chen, and F. Kamalabadi, “A Monte
acceleration,” Astrophys. J., vol. 616, pp. L171–L174, 2004. Carlo technique for large-scale dynamic tomography,” in Proc.
[9] A. C. Kak and M. Slaney, Principles of Computerized Tomographic ICASSP, Honolulu, HI, 2007, vol. 3, pp. 1217–1220.
Imaging. New York: IEEE Press, 1988. [37] N. J. Gordon, D. J. Salmond, and A. F. M. Smith, “Novel approach to
[10] F. Natterer and F. Wübbeling, Mathematical Methods in Image Recon- nonlinear/non-Gaussian Bayesian estimation,” IEE Proc. F, vol. 140,
struction. Philadelphia, PA: SIAM, 2001. pp. 107–113, 1993.
[11] M. D. Altschuler and R. M. Perry, “On determining the electron density [38] J. S. Liu and R. Chen, “Sequential Monte Carlo methods for dynamic
distribution of the solar corona from K-coronameter data,” Sol. Phys., systems,” J. Amer. Stat. Assoc., vol. 93, pp. 1032–1044, 1998.
vol. 23, pp. 410–428, 1972. [39] A. Doucet, N. de Freitas, and N. Gordon, Sequential Monte Carlo
[12] J. M. Davila, “Solar tomography,” Astrophys. J., vol. 423, pp. 871–877, Methods in Practice. New York: Springer-Verlag, 2001.
1994. [40] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tu-
[13] S. Zidowitz, “Coronal structure of the whole sun month: A tomographic torial on particle filters for online nonlinear/non-Gaussian Bayesian
reconstruction,” Space Phys., vol. 104, pp. 9727–9734, 1999. tracking,” IEEE Trans. Signal Process., vol. 50, no. 2, pp. 174–188,
[14] R. A. Frazin and P. Janzen, “Tomography of the solar corona. II. Ro- Feb. 2002.
bust, regularized, positive estimation of the three-dimensional electron [41] P. M. Djurić et al., “Particle filtering,” IEEE Signal Process. Mag., vol.
density distributions from LASCO-C2 polarized white-light images,” 20, no. 5, pp. 19–38, May 2003.
Astrophys. J., vol. 570, pp. 408–422, 2002. [42] A. K. George, M. D. Butala, R. A. Frazin, F. Kamalabadi, and Y.
[15] J. T. Mueller, H. Maldonado, and A. S. Driesman, “STEREO: The chal- Bresler, “Time-resolved CT reconstruction using the ensemble Kalman
lenges,” Acta Astronaut., vol. 53, pp. 729–738, 2003. filter,” in Proc. ISBI, Paris, France, 2008, pp. 1489–1492.
[16] H. C. van de Hulst, “The electron density of the solar corona,” Bull. [43] M. D. Altschuler, “Reconstruction of the global-scale three-dimen-
Astron. Inst. Netherlands, vol. 11, pp. 135–150, 1950. sional solar corona,” in Image Reconstructions From Projections: Im-
[17] M. D. Butala, R. A. Frazin, and F. Kamalabadi, “Three-dimensional plementation and Applications. Berlin, Germany: Springer-Verlag,
estimates of the coronal electron density at times of extreme solar ac- 1979, pp. 105–145.
tivity,” J. Geophys. Res., vol. 110, p. A09S09, 2005. [44] D. F. Elmore, J. T. Burkepile, A. J. Darnell, A. R. Lecinski, and A. L.
[18] R. A. Frazin, F. Kamalabadi, and M. A. Weber, “On the combination Stanger, “Calibration of a ground-based solar coronal polarimeter,” in
of differential emission measure analysis and rotational tomography Proc. SPIE, 2003, vol. 4843, pp. 66–75.
for 3D solar EUV imaging,” Astrophys. J., vol. 628, pp. 1070–1080, [45] G. E. Brueckner et al., “The large angle spectroscopic coronagraph
2005. (LASCO),” Sol. Phys., vol. 162, pp. 357–402, 1995.
[19] R. A. Frazin and F. Kamalabadi, “Rotational tomography for 3D re- [46] V. Domingo, B. Fleck, and A. I. Poland, “The SOHO mission: An
construction in the post-SOHO era,” Sol. Phys., vol. 228, pp. 219–237, overview,” Sol. Phys., vol. 162, pp. 1–37, 1995.
2005. [47] R. A. Howard, J. D. Moses, D. G. Socker, K. P. Dere, and J. W.
[20] S. Bonnet et al., “Dynamic X-ray computed tomography,” Proc. IEEE, Cook, “Sun earth connection coronal and heliospheric investigation
vol. 91, no. 10, pp. 1574–1587, Oct. 2003. (SECCHI),” Adv. Space Res., vol. 29, pp. 2017–2026, 2002.
[21] N. P. Willis and Y. Bresler, “Optimal scan for time varying tomo- [48] T. J. Bogdan and B. C. Low, “The three-dimensional structure of mag-
graphic imaging I: Theoretical analysis and fundamental limitations,” netostatic atmospheres. II—Modeling the large-scale corona,” Astro-
IEEE Trans. Image Process., vol. 4, no. 5, pp. 642–653, May 1995. phys. J., vol. 306, pp. 271–283, 1986.
[22] N. P. Willis and Y. Bresler, “Optimal scan for time varying tomo- [49] T. Neukirch, “On self-consistent three-dimensional analytic solutions
graphic imaging II: Efficient design and experimental validation,” IEEE of the magnetohydrostatic equations,” Astron. Astrophys., vol. 301, pp.
Trans. Image Process., vol. 4, no. 5, pp. 654–666, May 1995. 628–639, 1995.
BUTALA et al.: DYNAMIC TOMOGRAPHIC IMAGING OF THE SOLAR CORONA 765

[50] B. Sharif and F. Kamalabadi, “Optimal resolution in algebraic image [76] A. Doucet, S. Godsill, and C. Andrieu, “On sequential Monte Carlo
reconstruction,” in Bayesian Inference and Maximum Entropy Methods sampling methods for Bayesian filtering,” Stat. Comput., vol. 10, pp.
in Science and Engineering—Proc. MaxEnt 2005, K. H. Knuth, A. E. 197–208, 2000.
Abbas, R. Morris, and P. Castle, Eds., Melville, NY, 2005, vol. 803, [77] T. Bengtsson, P. Bickel, and B. Li, “Curse-of-dimensionality revis-
Amer. Inst. of Phys., pp. 199–206. ited: Collapse of the particle filter in very large scale systems,” Prob-
[51] E. Landi et al., “CHIANTI—An atomic database for emission ability and Statistics: Essays in Honor of David A. Freedman, vol. 2,
lines—VII. New data for X-rays and other improvements,” Astrophys. pp. 316–334, 2008.
J. Suppl., vol. 162, pp. 261–280, 2006. [78] N. Vaswani, “Particle filtering for large dimensional state spaces with
[52] J.-P. Delaboudinière et al., “EIT: Extreme-ultraviolet imaging tele- multimodal observation likelihoods,” IEEE Trans. Signal Process., vol.
scope for the SOHO mission,” Sol. Phys., vol. 162, pp. 291–312, 1995. 56, no. 10, pp. 4583–4597, Oct. 2008.
[53] J. L. Culhane et al., “The EUV imaging spectrometer for Hinode,” Sol.
Mark D. Butala (S’00) received the H.B.E.E. de-
Phys., vol. 243, pp. 19–61, 2007.
gree in electrical engineering from the University of
[54] T. Kosugi et al., “The Hinode (Solar-B) mission: An overview,” Sol.
Delaware, Newark, in 2002 and the M.S. degree in
Phys., vol. 243, pp. 3–17, 2007.
electrical and computer engineering from the Uni-
[55] M. A. Weber, E. E. Deluca, L. Golub, and A. L. Sette, “Temperature versity of Illinois at Urbana-Champaign (UIUC) in
diagnostics with multichannel imaging telescopes,” in Proc. IAU Symp. 2004, where he is currently pursuing the Ph.D. de-
No. 223, A. V. Stepanov, E. Benevolenskaya, and A. Kosovichev, Eds., gree.
Cambridge, MA, 2004, pp. 321–328. Since 2003, he has been a Research Assistant in
[56] L. Golub et al., “The X-ray telescope (XRT) for the Hinode mission,” the Remote Sensing and Space Sciences Group at the
Sol. Phys., vol. 243, pp. 63–86, 2007. Coordinated Science Laboratory. His research inter-
[57] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. New York: ests include the theory of remotely sensed image for-
Cambridge Univ. Press, 1991. mation, tomography and biomedical imaging, and Monte Carlo and statistical
[58] R. A. Frazin, A. M. Vásquez, F. Kamalabadi, and H. Park, “Three-di- signal processing methods.
mensional tomographic analysis of a high-cadence LASCO-C2 polar- Mr. Butala received a National Science Foundation Graduate Research Fel-
ized brightness sequence,” Astrophys. J., vol. 671, pp. L201–L204, lowship in 2002, a UIUC Department of Electrical and Computer Engineering
2007. Distinguished Fellowship in 2005, and a UIUC Computational Science and En-
[59] R. A. Frazin, S. R. Cranmer, and J. L. Kohl, “Empirically determined
O
gineering Fellowship and a Massachusetts Institute of Technology Lincoln Lab-
anisotropic velocity distributions and outflows of ions in a coronal oratories Graduate Fellowship, both in 2008.
streamer at solar minimum,” Astrophys. J., vol. 597, pp. 1145–1157,
2003.
[60] A.-I. N. Tikhonov and V. Y. Arsenin, Solutions of Ill-Posed Prob- Farzad Kamalabadi (M’02) received the B.S. de-
lems. Washington, DC: Winston, 1977. gree in computer systems engineering from the Uni-
[61] G. Demoment, “Image reconstruction and restoration: Overview of versity of Massachusetts, Amherst, in 1991, and the
common estimation structures and problems,” IEEE Trans. Acoust., M.S. and Ph.D. degrees in electrical engineering from
Speech, Signal Process., vol. 37, pp. 2024–2036, 1989. Boston University, Boston, MA, in 1994 and 2000,
[62] W. C. Karl, “Regularization in image restoration and reconstruction,” respectively.
in Handbook of Image and Video Processing, A. Bovik, Ed. San In 2000, he joined the University of Illinois at
Diego, CA: Academic, 2000, ch. 3.6, pp. 141–160. Urbana-Champaign (UIUC), where he is currently
[63] G. H. Golub, M. Heath, and G. Wahba, “Generalized cross-validation an Associate Professor in the Department of Elec-
as a method for choosing a good ridge parameter,” Technometrics, vol. trical and Computer Engineering, an Affiliated
21, pp. 215–223, 1979. Faculty Member in the Department of Statistics, and
[64] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation a Research Associate Professor at the Coordinated Science Laboratory. In fall
Theory. Upper Saddle River, NJ: Prentice-Hall, 1993. 2002, he was a Visiting Fellow with SRI International, and in summer 2003,
[65] Y. Zhang, A. Ghodrati, and D. H. Brooks, “An analytical comparison he was a National Aeronautics and Space Administration (NASA) Faculty
of three spatio-temporal regularization methods for dynamic linear in- Fellow with the Jet Propulsion Laboratory, California Institute of Technology,
verse problems in a common statistical framework,” Inverse Probl., vol. Pasadena. In 2008, he was a Visiting Professor at the Institut National de
21, pp. 357–382, 2005. Recherche en Informatique et en Automatique (INRIA), Sophia Antipolis,
[66] R. A. Frazin, M. D. Butala, A. Kemball, and F. Kamalabadi, France. His research interests include multidimensional and statistical signal
“Time-dependent reconstruction of nonstationary objects with tomo- processing, inverse problems in image formation, sensor array processing, and
graphic or interferometric measurements,” Astrophys. J., vol. 635, pp. tomography, and the development of optical and radio sensing and imaging
L197–L200, 2005. techniques and their applications to space physics, astronomy, and medicine.
[67] R. Furrer and T. Bengtsson, “Estimation of high-dimensional prior and Dr. Kamalabadi received a National Science Foundation (NSF) Fellowship
posterior covariance matrices in Kalman filter variants,” J. Multivariate in 1994, a Best Student Paper Award in 1997, a NASA Fellowship in 1998,
Anal., vol. 98, pp. 227–255, 2007. a NSF Faculty Early Career Development (CAREER) Award in 2002, and a
[68] P. J. Bickel and E. Levina, “Regularized estimation of large covariance UIUC College of Engineering Xerox Award for Outstanding Faculty Research
in 2006. He is currently a member of the Users and Scientific Advisory Com-
matrices,” Ann. Stat., vol. 36, pp. 199–227, 2008.
mittee of the Arecibo Observatory and has served in various organizational ca-
[69] G. Gaspari and S. E. Cohn, “Construction of correlation functions
pacities, including Session Organizer and Chair for the International Union of
in two and three dimensions,” Q. J. Roy. Meteor. Soc., vol. 125, pp.
Radio Science special sessions on ionospheric imaging in 2003 and 2004.
723–757, 1999.
[70] J. R. Guerci, “Theory and application of covariance matrix tapers for
robust adaptive beamforming,” IEEE Trans. Signal Process., vol. 47, Richard A. Frazin received the B.S. degree in
no. 4, pp. 977–985, Apr. 1999. physics, math, and astronomy from the University
[71] M. D. Butala, J. Yun, Y. Chen, R. A. Frazin, and F. Kamalabadi, of Wisconsin-Madison in 1992, the M.S. in physics
“Asymptotic convergence of the ensemble Kalman filter,” in Proc. from Georgia State University, Atlanta, in 1994, and
ICIP, San Diego, CA, 2008. the Ph.D. degree in astronomy from the University
[72] G. Evensen and P. J. van Leeuwen, “An ensemble Kalman smoother of Illinois at Urbana-Champaign (UIUC) in 2002.
for nonlinear dynamics,” Mon. Weather Rev., vol. 128, pp. 1852–1867, He was an Astrophysicist at the Smithsonian
2000. Astrophysical Observatory in Cambridge, MA,
[73] G. Evensen, “The ensemble Kalman filter: Theoretical formulation and from 2001–2002 and a Postdoctoral Researcher at
practical implementation,” Ocean Dynam., vol. 53, pp. 343–367, 2003. UIUC in the Department of Electrical and Computer
[74] A. Gelb, Applied Optimal Estimation. Cambridge, MA: MIT Press, Engineering from 2002 to 2006. In 2006, he joined
1974. the faculty of the Department of Atmospheric, Oceanic and Space Sciences at
[75] F. Daum and J. Huang, “Curse of dimensionality and particle filters,” the University of Michigan. His research interests include solar physics, inverse
in Proc. IEEE Aerospace Conf., Big Sky, MT, 2003, vol. 4, pp. problems in remote sensing (especially tomography of the Sun’s corona) and
1979–1993. statistical signal processing.
766 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 2, NO. 5, OCTOBER 2008

Yuguo Chen received the B.S. degree in mathematics


from the University of Science and Technology of
China, Hefei, in 1997 and the Ph.D. degree in statis-
tics from Stanford University, Stanford, CA, in 2001.
He is an Associate Professor in the Department
of Statistics at the University of Illinois at Ur-
bana-Champaign (UIUC). Prior to joining UIUC,
he was an Assistant Professor at the Institute of
Statistics and Decision Sciences at Duke University,
Durham, NC, from 2001 to 2005. His research
interests include Monte Carlo methods, statistical
genetics, and dynamic systems.

You might also like