0806.4803/ms.tex
1: % COMPILING:
2: %
3: % latex -interaction=nonstopmode ms
4: % bibtex ms
5: % latex  ms
6: % latex  ms
7: % dvipdfmx ms
8: %
9: % Note: When switching between aastex and emulateapj, the first call
10: % to LaTeX may throw an error. Use 'r' to force it to run, or use
11: % -interaction=nonstopmode, or clear out all the auxiliary files before
12: % running LaTeX.   CCK 2008-Jul-13
13: 
14: % ApJ Emulation: 
15: \documentclass[dvipdfm,letter]{emulateapj} % comment out for MS to ApJ!
16: \usepackage{mathptmx}                      % comment out for MS to ApJ!
17: \usepackage{hyperref}                      % comment out for MS to ApJ!
18: 
19: %\documentclass[manuscript]{aastex} % comment out when in emulation mode!
20: \usepackage{natbib}
21: \usepackage{amsmath}
22: \usepackage{graphicx}
23: 
24: 
25: \citestyle{aa} % As required by ApJ: (Bester 1998) rather than (Bester, 1998)
26: 
27: 
28: % Attempt to reprogram LaTeX's algorithm for float placement (didn't help)
29: %    % See p.105 of "TeX Unbound" for suggested values.
30: %    % See pp. 199-200 of Lamport's "LaTeX" book for details.
31: %    %   General parameters, for ALL pages:
32: %    \renewcommand{\topfraction}{0.9}	% max fraction of floats at top
33: %    \renewcommand{\bottomfraction}{0.8}	% max fraction of floats at bottom
34: %    %   Parameters for TEXT pages (not float pages):
35: %    \setcounter{topnumber}{2}
36: %    \setcounter{bottomnumber}{2}
37: %    \setcounter{totalnumber}{4}     % 2 may work better
38: %    \setcounter{dbltopnumber}{2}    % for 2-column pages
39: %    \renewcommand{\dbltopfraction}{0.9}	% fit big float above 2-col. text
40: %    \renewcommand{\textfraction}{0.07}	% allow minimal text w. figs
41: %    %   Parameters for FLOAT pages (not text pages):
42: %    \renewcommand{\floatpagefraction}{0.7}	% require fuller float pages
43: %	% N.B.: floatpagefraction MUST be less than topfraction !!
44: %    \renewcommand{\dblfloatpagefraction}{0.7}	% require fuller float pages
45: %	% remember to use [htp] or [htpb] for placement
46: 
47: %\slugcomment{Not to appear in any journal, as it won't LaTeX.}
48: \shorttitle{Coronal Tomography}
49: \shortauthors{Charles C.\ Kankelborg}
50: 
51: \begin{document}
52: 
53: \title{Coronal Tomography}
54: \author{Charles C. Kankelborg}
55: \affil{Department of Physics, Montana State University, Bozeman, MT 59717}
56: \email{kankel@solar.physics.montana.edu}
57: 
58: 
59: \begin{abstract}
60: A simple, yet powerful, algorithm for computed tomography of the solar corona is
61: presented and demonstrated using synthetic EUV data. A minimum of three
62: perspectives are required. These may be obtained from \textit{STEREO}/EUVI plus
63: an instrument near Earth, e.g.\ \textit{TRACE} or \textit{SOHO}/EIT.
64: \end{abstract}
65: 
66: \keywords{methods: data analysis --- Sun: corona --- techniques: image processing}
67: 
68: \section{Introduction}
69: 
70: Observations by the Extreme Ultraviolet Imager \citep[EUVI,][]{Wuelser2004} aboard
71: \textit{STEREO} provide the first simultaneous, stereoscopic image pairs of the solar
72: corona and transition region. Ideally, these data are simple projections through an
73: optically thin corona. However, the 3D distribution of emission is difficult to
74: estimate with only two projections \citep[see][and references therein]{Gary1998}. This difficulty may be explained as follows. Consider an object  $I(x,y,z)$
75: on domain $D$. The coordinates $x$ and $y$ need not be orthogonal, but $z$ is
76: orthogonal to $x$ and $y$. Given two projections $f(x,z) = \int_D I\,dy$ and
77: $g(y,z)=\int_D I\,dx$, a plausible reconstruction of $I$ is
78: \begin{equation}\label{eq:separable}
79:    I'(x,y,z) = \frac{f(x,z)\,g(y,z)}{T(z)},
80: \end{equation}
81: where $T(z) = \int_D I(x,y)\,dx\,dy = \int_D f\,dx = \int_D g\,dy$ is
82: the total emission of a plane of constant $z$. Unfortunately, this solution
83: fails utterly in practice. For each pair of sources in $I$,  $I'$ introduces a
84: pair of ``ghost'' artifacts. These are systematic errors, independent of the
85: noise and apparently unavoidable. Traditional regularization strategies are
86: not fruitful: $I'$ is positive, is as smooth as the observed images $f$ and
87: $g$, and is precisely the maximum entropy solution. More information is
88: therefore required to guide the tomographic reconstruction.
89: 
90: Previously described approaches to the \textit{STEREO} coronal tomography problem rely
91: on assumptions about the geometry of the coronal plasma distribution. The triangulation
92: method \citep{Gary1998,Aschwanden2005,Aschwanden2008} assumes loops with circular cross-section, and
93: relies on the identification of the same loops in both images. The magnetic tomography
94: approach \citep{WiegelmannInhester2006} also assumes loops with circular cross-section,
95: and incorporates magnetic field extrapolations to constrain loop geometries.  These
96: methods are powerful, but they require assumptions about things that one might
97: reasonably hope to learn from the 3D reconstruction.
98: 
99: I propose that EUV images taken from a third perspective---e.g.\
100: \textit{TRACE} or \textit{SOHO}/EIT---may provide adequate additional
101: constraints for coronal tomography, without any assumptions about loop
102: geometry or magnetic fields. I describe a simple computed tomography
103: algorithm, fast enough to run in real time, and demonstrate its performance
104: using synthetic data with three viewpoints. 
105: 
106: 
107: \section{Algorithm} \label{sec:smart}
108: 
109: The Smooth Multiplicative Algebraic Reconstruction Technique (SMART) presented here has
110: been developed to solve a mathematically analogous problem of reconstructing spectra
111: for the \textit{MOSES} sounding rocket payload \citep{KankelborgThomas2001,Fox2003}.
112: Iterative multiplicative algebraic reconstruction techniques (MART), perhaps inspired
113: by the separable solution to the two-view problem (equation \ref{eq:separable}), have
114: been available for many years \citep{OkamotoYamaguchi1991,Verhoeven1993}.
115: \cite{Gary1998} used MART along with volume constraints derived from magnetic field
116: extrapolations to reconstruct a pair of loops from two simulated XUV images. The unique
117: algorithmic features of SMART are iterative smoothing and an adaptive correction
118: strategy. These refinements improve numerical stability and promote convergence to a
119: compromise between smoothness and goodness-of-fit, leading to a reduced chi-squared of
120: unity. 
121: 
122: In the $N$-view tomography problem, an object $I(x,y,z)$ on domain $D$ is
123: known only by $N$ projections $f_m$, taken at angles $\theta_m$:\footnote{In our coordinates, $\theta_m$ is a right-handed rotation of the object; it therefore corresponds to the eastward heliographic longitude of the observer.}
124: \begin{equation}
125:    f_m(x,z) = \int_D \mathcal{R}(\theta_m)\,I(x,y,z)\,dy.
126: \end{equation}
127: The operator $\mathcal{R}(\theta_m)$ rotates the object $I$ by angle
128: $\theta_m$ about the $z$ axis.\footnote{Rotation $\mathcal{R}$ could
129: incorporate compound angles with altitude, azimuth and roll. The extension of
130: SMART to the general case is straightforward.} SMART uses the projections $f_m$
131: to estimate $I(x,y,z)$  by the following steps: 
132: \begin{enumerate}
133:    \item \label{step:guess} Create an initial guess, $I'(x,y,z) = 1$ on $D$. 
134:    \item Initialize correction weights, $\gamma_m=\frac{1}{N}$.
135:    \item \label{step:project} $I' \leftarrow I' * K$
136:       (smoothing kernel $K$ defined by eq.\ \ref{eq:kernel}).
137:    \item Calculate projections $f'_m(x,z) = \int_D \mathcal{R}(\theta_m)\,I'\,dy$.
138:    \item Calculate correction factors, 
139:       \begin{equation} \label{eq:correctionfactor}
140:          C_m(x,y,z) = \mathcal{R}(-\theta_m)\,\frac{f'_m(x,z)}{f_m(x,z)}.
141:       \end{equation}
142:       Note that a nontrivial $y$-dependence is introduced through the
143:       rotation.
144:    \item \label{step:correct} Apply corrections weighted by $\gamma_m$,
145:       \begin{equation}\label{eq:correct}
146:          I' \leftarrow I' \,\prod_m C_m^{\gamma_m}.
147:       \end{equation}
148:    \item Calculate reduced chi-squared for each projection, 
149:       \[
150:          \chi_{R,m}^2 = \frac{1}{N_x N_y N_z}\sum_{xyz} \frac{(I-I')^2}{I'}.
151:       \]
152:    \item \label{step:feedback} Adjust correction weights, $\gamma_m$,
153:       using a control law designed to drive $\chi_{R,m}^2 \rightarrow 1$.
154:    \item Repeat steps \ref{step:project}-\ref{step:feedback}
155:       until converged.
156: \end{enumerate}
157: 
158: The $I'$ array is oversized so that rotations will not move any of the emission
159: outside the volume. The subphotospheric portion of the volume is zeroed in step
160: \ref{step:guess}, and will remain zero since the corrections are multiplicative
161: (eq. [\ref{eq:correct}]).
162: 
163: Since the correction factors (eq.\ [\ref{eq:correctionfactor}]) are ratios of
164: positive numbers, the reconstruction is always positive. The strength of the
165: $m$th applied correction is governed by $\gamma_m$. Step \ref{step:feedback}
166: sets up feedback to establish a dynamic equilibrium between smoothing and
167: correction, so that $\chi_{Rm}^2$ tends toward unity. Our control law, which has two adjustable parameters $(a,b)$, modifies $\gamma_m$ for iteration $n+1$ using a linear combination of the previous and current values of $\chi_{Rm}^2$:
168: \begin{align} \label{eq:control}
169:    \gamma_m^{n+1} &= \gamma_m^n + a\,X_m^n + b\,\Delta X_m^n,\\ 
170:    X_m^n &\equiv \log (\chi_{Rm,n}^2),
171:    \quad \Delta X_m^n \equiv X_m^n - X_m^{n-1}.
172: \end{align}
173: 
174: The algorithm is implemented in IDL, with rotations $\mathcal{R}(\theta_m)$ perfomed
175: via cubic  convolutional interpolation. It is possible that spurious negative voxels
176: could be introduced during the rotation, but negative values are eliminated from our
177: projections by thresholding: $f_m \ge 0$.
178: 
179: The normalized smoothing kernel, $K(x,y,z)$, is defined on the discrete space
180: of voxels as follows:
181: \begin{equation} \label{eq:kernel}
182: %   K = \frac{1}{1+s(1+2a)^2}
183: %      \begin{bmatrix}
184: %         sa^2 & sa  & sa^2 \\
185: %         sa   & 1+s & sa   \\
186: %         sa^2 & sa  & sa^2
187: %      \end{bmatrix}, \quad a=\frac{2}{5}.
188:    K_{ijk} = \frac{\delta_{ij}\,\delta_{jk} 
189:       + s\,c^{(i+j+k)}}{1+s(1+4c+12c^2+8c^3)},  
190:       \quad c \equiv \frac{2}{5}.
191: \end{equation}
192: This form of $K$ is not crucial, but it is designed to be very nearly
193: isotropic. The adjustable smoothing parameter, $s \in (0,1]$, affects the
194: rate of convergence but has no discernible effect on the result.
195: 
196: 
197: 
198: \section{Synthetic Data} \label{sec:data}
199: 
200: 
201: A volume of synthetic coronal emission was prepared as a test target for the SMART
202: algorithm. The test target is more complex than any that have proven tractable for
203: previous approaches to tomographic analysis for STEREO. It is not an attempt at
204: detailed atmospheric modeling, but resembles a small active region. I began with a
205: potential field defined by four sub-photospheric magnetic charges. The resulting
206: line-of-sight photospheric magnetic field is shown in figure \ref{fig:magnetogram}. The
207: field lines in the figure emerge from five small, square ``heated'' patches in the
208: postive polarities on the photospheric plane. A large  number of field linese were
209: traced from random points in the volume to both footpoints (or to the boundary).
210: Emission was added to each voxel crossed by the field line, in proportion to the
211: arbitrary heating value at its positive photospheric footpoint. Figure \ref{fig:volume}
212: shows the volume model projected along each of the three axes of figure
213: \ref{fig:magnetogram}. The coronal flux tubes have complicated geometries, including a
214: broad, vertical fan of emission above the postive pole that is associated with a
215: coronal magnetic null at a height of $\sim 53$ grid points. All images of EUV emission
216: in this letter are square-root scaled to bring out faint features. 
217: 
218: \begin{figure}[htb]
219:    \begin{center}
220:       \includegraphics[width=\columnwidth]{magnetogram_fig.eps}
221:    \end{center}
222:    \caption{Vertical magnetic field (shaded plane) and field lines within the emission
223:    from the model corona. five ``Heated'' patches  within the positive photospheric
224:    magnetic poles are colored in black. Note the sharp deflection of the upppermost
225:    field line from the coronal null.}
226:    \label{fig:magnetogram}
227: \end{figure}
228: 
229: 
230: \begin{figure}[htb]
231:    \begin{center}
232:       \includegraphics[width=\columnwidth]{corona3view.eps}
233:    \end{center}
234:    \caption{The model coronal volume viewed along each of the three axes of
235:    figure \ref{fig:magnetogram}. Intensities are square-root
236:    scaled.}\label{fig:volume}
237: \end{figure}
238: 
239: \begin{figure*}[htb]
240:    \begin{center}
241:       \includegraphics[width=\textwidth]{projections_horiz.eps}
242:    \end{center}
243:    \caption{Synthetic data taken from three virtual spacecraft. Intensities are
244:    square-root scaled.}\label{fig:projections}
245: \end{figure*}
246: 
247: SMART was tested on numerous synthetic observations of the model coronal volume,  each
248: time placing the model active region at a different random northern heliographic
249: latitude over $[0^{\circ},30^{\circ}]$, a random heliographic longitude over
250: $[-40^{\circ}, 40^{\circ}]$, and a random tilt over the interval $[-180^{\circ},
251: 180^{\circ}]$. For simplicity, in our coordinate system the solar equator coincides
252: with the equatorial plane. Observations were projected for three distant virtual
253: instruments in the ecliptic, observing from heliographic longitudes $-40^{\circ},
254: 0^{\circ}$, and $40^{\circ}$.  The twin \textit{STEREO} spacecraft will reach similar
255: separation angles in October, 2008.  The images were normalized so that the brightest
256: pixel among the three images had 3000 counts. The $\theta =
257: (-40^{\circ},0^{\circ},40^{\circ})$ projections shown in figure \ref{fig:projections}
258: correspond to an example observation with the region placed at latitude
259: $16.51^{\circ}$\,N, longitude $14.65^{\circ}$\,W, and tilt $29.63^{\circ}$
260: counter-clockwise. Poisson noise was applied to the images prior to passing them to the
261: SMART algorithm for inversion.  Intensities are square-root scaled to show the noise
262: more clearly. The mean value of the nonzero pixels is 216 counts.
263: 
264: 
265: 
266: \section{Inversion Results} \label{sec:results}
267: 
268: The synthetic data were inverted using 15 SMART iterations. This was typically
269: sufficient to converge to $\chi^2_{Rm}=1 \pm 0.01$. The rate of convergence is affected
270: by the smoothing parameter and by the two adjustable parameters in the control law. The
271: examples shown in this Letter used $s=0.5$, $a=0.05$, $b=0.16$. Within broad limits,
272: the results are insensitive to these parameters. For example, if the smoothing is
273: removed altogether, numerical instabilities arise; if it is made too strong, then it
274: will not be possible to reach $\chi^2_{Rm}=1$. For our computational volume of $169^3$
275: voxels, an inversion runs in a few minutes on a laptop computer.
276: 
277: % \begin{figure}[!htbp]
278: %    \begin{center}
279: %       \includegraphics[width=0.7\columnwidth]{eclipticPan.eps}
280: %    \end{center}
281: %    \caption{Comparison of true and reconstructed coronal volumes, panning
282: %    in the ecliptic plane. Intensities are square-root scaled.}
283: %    \label{fig:eclipticPan}
284: % \end{figure}
285: % \begin{figure}[!htbp]
286: %    \begin{center}
287: %       \includegraphics[width=0.7\columnwidth]{eclipticPan2.eps}
288: %    \end{center}
289: %    \caption{Same as figure \ref{fig:eclipticPan}, but with loops
290: %    nearly parallel to the ecliptic.}
291: %    \label{fig:eclipticPan2}
292: % \end{figure}
293: % \begin{figure}[!htbp]
294: %    \begin{center}
295: %       \includegraphics[width=0.7\columnwidth]{outpan.eps}
296: %    \end{center}
297: %    \caption{Comparison of true and reconstructed coronal volumes, moving
298: %    up the central heliographic meridian at five different latitudes. 
299: %    Intensities are square-root scaled.}
300: %    \label{fig:outpan}
301: % \end{figure}
302: % \begin{figure}[!htbp]
303: %    \begin{center}
304: %       \includegraphics[width=0.7\columnwidth]{outpan2.eps}
305: %    \end{center}
306: %    \caption{Same as figure \ref{fig:outpan}, but with loops
307: %    nearly parallel to the ecliptic.}
308: %    \label{fig:outpan2}
309: % \end{figure}
310: 
311: \begin{figure}[!p]
312:    \begin{center}
313:       \includegraphics[width=\columnwidth]{threeview1.eps}
314:    \end{center}
315:    \caption{Comparison of true and reconstructed coronal volumes from three
316:       orthogonal points of view. Intensities are square-root scaled.}
317:    \label{fig:threeview1}
318: \end{figure}
319: \begin{figure}[!p]
320:    \begin{center}
321:       \includegraphics[width=\columnwidth]{threeview2.eps}
322:    \end{center}
323:    \caption{Same as figure \ref{fig:threeview1}, but with loops
324:    nearly parallel to the ecliptic.}
325:    \label{fig:threeview2}
326: \end{figure}
327: 
328: The simulated data in figure \ref{fig:projections} gives rise to an accurate
329: reconstruction. 
330: 
331: Figure \ref{fig:threeview1} shows reconstruction results for the data in figure
332: \ref{fig:projections}, compared to noise-free visualizations of the coronal volume
333: model. The Earth viewpoint at the top of figure \ref{fig:threeview1} corresponds to the
334: middle panel of figure \ref{fig:projections}. Comparing these two figures shows that
335: the photon noise has been largely suppressed in the SMART reconstruction. The success
336: at recovering 3D geometry is best illustrated by the east-west and south-north
337: projections, which are $50^{\circ}$ and $90^{\circ}$, respectively, from the nearest
338: available observing angles in the synthetic data.  All of the true features have been
339: recovered. Square-root scaling helps to bring out the artifacts, which are few and
340: faint. There is slight blurring, and minimal ghosting. These results are typical of hundreds of realizations tried so far. 
341: 
342: A second example, with loops nearly parallel to the ecliptic plane, is given in figure
343: \ref{fig:threeview2}. The horizontal loop orientation is very challenging because only
344: the ends of horizontal features provide any depth cues.  Altough views from within the
345: ecliptic plane are reproduced well, the example shows relatively poor reconstruction of
346: an out-of-ecliptic view (lower panel). Animated versions of figures
347: \ref{fig:threeview1} and \ref{fig:threeview2} are provided in the electronic version of
348: the Journal. 
349: 
350: 
351: 
352: \section{Discussion and Conclusions} \label{sec:discussion}
353: 
354: Coronal tomography is possible with as few as three viewpoints, making no prior
355: assumptions about coronal morphology or magnetic fields.  The test cases demonstrate
356: recovery of complex geometry without reference to magnetic field extrapolations or
357: assumptions about loop geometry. Loops that run in an east-west direction, however,
358: provide insufficient depth cues for three instruments confined to the  ecliptic.
359: 
360: The SMART algorithm described here provides a noise-insensitive tomographic
361: reconstruction by finding an optimal balance between goodness of fit and local
362: smoothness.  \textit{STEREO} will obtain data at large separation angles in Fall 2008.
363: Observations from \textit{SOHO} and/or \textit{TRACE} at that time should allow the
364: best opportunity for application of SMART to coronal tomography. 
365: 
366: 
367: \section*{Acknowledgments}
368: 
369: I thank Dana Longcope for many helpful suggestions made during the preparation
370: of this Letter. This work was supported by NASA grant NNX07AG6G.
371: 
372: \clearpage
373: 
374: \bibliographystyle{apj}
375: \bibliography{apj-jour,CCKletter}
376: 
377: 
378: 
379: 
380: \end{document}
381: