HMI Tracking Tests

Michael Turmon, JPL, April 2011

The movies below are made with the current masks, just like the earlier movies from late 2010. These masks use magnetogram information only (no intensity information). The masks are believed to be of good quality.

We use masks generated at a 720s cadence. The faster cadence (relative to MDI) made it important to use temporal information more carefully, so we enhanced the tracker to retain more information about the configuration of past AR outlines. The movies below are made with this new tracker. We are still tuning two or three parameters, but we believe these results are good, and much improved from the MDI tracker.

Movie Layout

The movies have 1024x1024 frames; they may be cropped by the inline movie player in your browser. If this is getting in the way, download the movie and play it locally.

The disk is gray, and pixels labeled as "active" within the gray disk are black. The active pixels shown are generated by simply downsampling the activity mask by 4:1 in both directions. This can be misleading, especially near the limb, where very few active pixels are needed to cause an AR to be found.

The colored blobs, together with the light gray outline that may be around the blob, represent the full content of different HARPs. Small clumps of black pixels that are outside the HARP's are below-threshold. The color of the HARP-blob is related to its track number (counting up from one).

The white bounding box surrounding the HARP encircles all within-HARP pixels. Sometimes a white bounding box can enclose two separated blobs. (The blobs within a box should always be the same color.) This happens when, at a later frame, these two regions were merged. A small white square hangs inside the white box to remind us that this is a merged region. If more than one region was merged, the height of the white indicator lengthens.

New regions are shown with a white square hanging outside the white bounding box.

Sometimes the bounding box is dotted. This means that the corresponding HARP was not seen in that frame, but is being kept around in case it reappears. This is called a "placeholder" region and generally happens for highly decayed ARs.

More on Filtered HARPs

Earlier movies used only a colored blob to represent HARP outline. The new tracker is more careful to combine past outlines of the HARP with the new outline at this particular image in the sequence. This combination is shown in the movie as follows. The filtered HARP outline is the colored blob plus the light-gray border around it. Sometimes the border vanishes (e.g., a growing HARP). Very stable ARs will have a very narrow border, which can look like a thin gray line. Unstable ARs, usually decaying, can be mostly gray. The instantaneous outline, as deduced from this image alone, is the colored blob alone. For unstable ARs, the colored blob may disappear entirely, and then reappear later, as the region dips below and then back above a detection threshold.

The best way to think of the distinction between the filtered outline and the instantaneous outline is as data assimilation or filtering in a forward time series model:

filtered_outline(t+1) = a × instantaneous_outline + (1-a) × filtered_outline(t)

The new filtered_outline is the posterior, the old one is the prior, and the instantaneous_outline is the new information from just this frame (the "data term" or "likelihood term" in filtering theory).

The colored-plus-gray color coding allows us to show both the instantaneous_outline (which is 0/1) and the filtered_outline (which is continuous-valued) on the same map. We're only showing the nonzero part of the filtered_outline, because its precise value is not very important to the current tracker. That is, the light-gray regions show where filtered_outline is greater than zero, but do not show what value it has.

The weight a is a function of the time since the last observation: a = exp(- rate × dt). As a decreases from 1 to zero, memory is lengthened. The movies below use parameters that retain about 20 frames at the nominal rate.

Text Insets

All four corners of the movie have text insets. I changed the format a little from past movies.

Top left: time and date of T_REC, time difference since the prior image (in DD+HH:MM format, typically 00+00:12 for 720s cadence imagery), and frame number in overall image sequence.

Bottom left: Total number of HARPs in this frame, plus counts of special HARPs: new regions (rare), merged regions, and two types of regions which had an all-zero instantaneous_outline in this frame ("past" and "placeholder" regions).

Top right: For the top half of the image, shows the HARP ID of the current region sets. A "^" next to the ID means it is the largest on the disk, and a "v" means it is the smallest. A "!" means it is new. A "+" means it will later be merged.

Bottom right: Region list for the bottom half of the image, in format just described.

Next to the track number appears a colored square corresponding to the track shown.

Files Here

I'm processing all the present HMI 720s imagery into movies, using a common parameter set, to see if the parameters are OK. Tracking a long time period is also good to identify systems robustness issues. It is taking about 40s-60s per image, which translates into 40 to 60 hours to process tracks per month.

My top-level conclusion is that the temporal memory is not long enough at present, which causes intermittent ARs to be double-counted. The key parameter here is final_time, the length of time a region is allowed to linger, after not being seen, before it is declared final and that track is closed out. This is presently 3 hours and should probably be closer to 24 hours.

Tag Size Starts at Ends at Notes
2010-may 37M 2010/05/01 2010/06/01 Notes
2010-jun 34M 2010/06/01 2010/07/01 Notes
2010-jul 44M 2010/07/01 2010/08/01 Notes
2010-aug 49M 2010/08/01 2010/09/01 Notes
2010-sep 51M 2010/09/01 2010/10/01 Notes
2010-oct 45M 2010/10/01 2010/11/01 Notes
2010-nov 43M 2010/11/01 2010/12/01 Notes
2010-dec 50M 2010/12/01 2011/01/01 Notes
2011-jan 44M 2011/01/01 2011/02/01 Notes
2011-feb 47M 2011/02/01 2011/03/01 Notes
2011-mar 55M 2011/03/01 2011/04/01 Notes
2011-apr 47M 2011/04/01 2011/04/18 Notes

Times of Interest

Notes on May 2010 movie

This movie has several times when a longer memory for ARs would allow combining two small and intermittent ARs into one longer AR. I have not taken the time yet to list them. From the point of view of the memory-length question, this is the most revealing movie of the set.

Notes on June 2010 movie

Nothing noted yet.

Notes on July 2010 movie

Nothing noted yet.

Notes on December 2010 movie

This movie breaks the earlier sequence: AR numbering starts at 1001.

There is a temporal gap of 2:12 (11 nominal images) at 2010/12/09 14:48. The current code should be insensitive to gaps up to a few days in width. I've made an effort to put all parameters in terms of time differences rather than frame number differences.

The region numbered 1021 has a 3-way merge.

Region 1038 (frame 001730) is probably a CR. I have not checked.

Regions 1035 and 1039 (frames 1710-1760) should probably have been combined. This seems like the most glaring mistake during this month, again due to the memory issue.

Notes on January 2011 movie

Nothing noted yet.

Notes on February 2011 movie

No notes yet.

Notes on March 2011 movie

Huge artifact in two magnetograms at 2011/03/04 12:36 (frame 11087). This kind of artifact can destroy an entire month's results because an overlap with existing regions can cause all on-disk regions to be merged. Here the limited memory noted above might have been beneficial because the AR dies out quickly.

Several cosmic rays near the limb (frame 11327, 11808, 12643?, 13814) caused length-1 HARPs.

There are several data gaps of from one to three hours.