Images:

I downloaded the leavstacked multiframe images from the archive listing at WSA (e.g., w20050413_02994_sf_st.fit). The table below shows examples for the following cases.

The ring-shaped artefacts (cross talk) are already described here, but see also the example w20050412_02874_sf_st1_ringe.png in the table for the different possible sizes.

There are also the striped patterns caused by dust on the field lens as described here (no individual example here, but see the K-band images in w20050413_03010_sf_st3_diffem.png and w20050413_03010_sf_st4_spurjhk3.png).

In some cases the image shows a funny raster pattern (w20050410_02494_sf_st3_raster.png), that is, every star shows such a pattern slightly offset. In addition, there is a larger patch unconnected with any stars. These patterns can be picked up by the source detection algorithm (detection as galaxies shown only). Maybe this only happened with the very early images?

Another example shows a dark raster instead of a bright one (w20050418_01425_sf_st1_raster2.png), though I'm not sure whether this is an artefact of the display and is an extreme case of the dark patches described below.

The images show a very uneven background (e.g., w20050412_02874_sf_st1_darkpatches.png, but see also others). Part of it is due to the density of very faint and unresolved stars at low latitudes, but the appearance of the pattern reminds me of other observations in the plane where the sky subtraction (also affected by star crowding) seemed to be resposible for this. There is probably not much one can do about it, however, people working at such high stellar densities should be aware of the uncertainties in the background, which makes it difficult to discriminate these variations from faint diffuse emission of similar scales.

There are large dark stripes in the images (I vaguely remember this problem being mentioned at the last meeting in London). These are both obvious with gaia as well as at large zoom-ins with ds9. In addition, at smaller zooms with ds9 a chequered pattern appears with must be due to the sampling technique used by ds9 (it varies when the window is moved across the field). w20050413_03010_sf_st4_checkered.png shows this chequered pattern for the three passbands J (left), H (middle) and K (right) - the K-band is washed out due to differences in the scaling, sorry.

Detections:

I'm mainly interested in galaxies, and consequently, all kind of extended emission; so my investigations focus on this. So far I only found the time to look at two multiframes of one tile (18_0), each with 4 images and 3 passbands. This tile is at high latitude and low extinctions (in terms of the Galactic plane!). Next, I'm going to look at a tile at high extinctions...

Scanning those eight images by eye (simultaneously in JHK) I found 20 galaxies (these are obvious, I didn't try to look at faint and small objects - here it is almost impossible to judge by eye unless one can confirm it otherwise) and 5 extended objects that could be galaxies or Galactic Nebulae. w20050413_02994_sf_st1_gal.png shows an example of a galaxy.

I compared these findings with the catalogues in the archive. First I looked at the (raw) catalogues that come with the images (e.g., w20050413_03010_sf_st_cat.fits). These give a class (1 for galaxies) and a statistics parameter. Since there is no documentation yet, I can only guess that this statistics paramater is a measure of how good the classification is. So I first plotted all class=1 detection on an image, and found it was swamped with detections (see w20050410_02366_sf_st1_detall.png: the thicker blueish-green circles are those with statistics > 4.5). It is obvious that most detections are triggered by blended stars. I have then looked at the histogram of the statistics parameter and arbitrarily set a limit at 10 (where the number of detections has dropped significantly). But even here, most of the detections are no extended emssion (see w20050410_02366_sf_st3_det10.png: the detection in the centre is a galaxy). This is all J-band only.

There are a couple of examples that show the variations of spurious detections with statistics > 10: w20050413_03010_sf_st4_spurjhk2.png and w20050413_03010_sf_st4_spurjhk3.png. In these examples the J-band image is to the left, the H-band image in the middle, and the K-band image (rather bright, sorry, because of the difference in scaling). The detections are in green for the J-band, cyan for the H-band and red for the K-band. I plotted all three regions on the J-band image, but only the respective regions on the H- and K-band images for clarity. The most interesting detections are those that are detected in more than one band, which will further reduce the number of spurious detections. However, one has to keep in mind that due to the intrinsic variations in colours galaxies may not be visible in the K-band while being faint in J. At high extinctions, however, we would expect to see galaxies in the K-band more often than in the J-band (which is due to the higher extinction in J, which is 0.21 times the extinction in the B-band, while in the K-band the extinction is only 0.09 times A_B).

I have therefore extracted the statistics parameter for all the galaxies that I have found, and it indicates that they can have statistics parameters as low as 8. Only in one case the J-band statistics was ~3 (see w20050413_02994_sf_st3_galfaint.png), which may be counted as 'not detetced in J' and is consistent with the fact that the J-band object is just at the detection limit, while the H- and K-band objects are clearly above.

This limit, that is, statistics > 8 (based on only 20 objects!), still leaves too many spurious detections in the catalogue to comfortably live with.

Next, I downloaded the merged catalogues for this area (through 'menue query' at WSA). w20050413_03010_sf_st3_mcat_all.png (J-band) shows all objects with pgal > 0.85, that means, there are a few objects with class other than 1 (in cyan). But the majority is defined as galaxy. Again, there are clearly too many spurious detections, in particular in connection with the edge of the image and around brighter stars (since I could only download objects selected by RA and DEC and not by image, multiframe or tile, there are also detections beyond the edge of the image).

Splitting these detections into the individual bands shows that only few of these merged class objects have a class other than 1 in one passband (no example shown).

Restricting the pgal paramater to pgal > 0.999 is much better, but there are still a lot of spurious detections (blended stars mainly, but also one of the rings caused by the saturated stars, see w20050413_03010_sf_st3_mcat_999.png). There are actually two galaxies just slightly above and to the left of the centre, they have pgal=0.99965715.

Finally a few example of larger extended objects (note, that the detection regions are only shown for statistics > 10, as explained above).

• dark cloud: w20050410_02366_sf_st3_darkclouds.png
• prob gal: w20050413_03010_sf_st3_diffem.png
• others? w20050413_03010_sf_st3_diffem2.png and w20050413_03010_sf_st3_diffem3.png (shows also 2 galaxies detected at the bottom of the image)
• prob gal: w20050413_03010_sf_st_diffem4.png (the green circle in the J-band is nopt a detection region but only indicates the object).
• 2 galaxies or parts of a larger nebula? w20050413_02994_sf_st3_extd.png

Conclusions:

As far as I understood, the image quality is going to improve. But to detect small and large extended objects with higher reliability, there is still a lot of work to be done. It is extremely difficult to avoid detecting blended objects (in particular around bright stars) as extended objects. It may be possible to make use of other parameters (e.g. the relation of peak magnitude to isophotal area) to further restrain the detections. Another possibility may be the use of neural networks. This has been tried before, but unfortunately my computational knowledge is too small to be of any help here. I can only point out a couple of works, one regarding SExtractor:

• Bertin, E., Arnouts, S: 1996, A&AS 117, 393
• Bazell, D., Peng, Y: 1998, ApJS 116, 47
• Andreon, S, Cargiulo, G., et al.: 2000, MNRAS 319, 700
• Cortiglioni, F., Maehoenen, P, et al.: 2001, ApJ 556, 937

Of course, there is more, but these seem to be either presnt a package or discuss various methods.

Regarding (large) dark clouds one could make use of stellar density contours (cf w20050410_02366_sf_st3_darkclouds.png) and look out for sharp gradients. This, of course, might also help with detecting stellar clusters.

-- AnjaSchroeder - 15 Aug 2005