Pith. sign in

REVIEW 2 major objections 34 references

Masking Algorithm for CCD Bleeding in Korea Microlensing Telescope Network Images

T0 review · 2 major / 0 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pixel-level mask for KMTNet bleed trails raises source completeness from 94% to 99% and cuts photometric error nearly tenfold.

desk verdict Practical, pipeline-ready KMTNet bleeding mask with real Gaia-validated gains; single-field Fβ tuning of n/CL is the main soft spot, not a load-bearing flaw. read the letter →

arxiv 2607.09148 v1 pith:BZBRBN3X submitted 2026-07-10 astro-ph.IM

Jiseop Shin , Mankeun Jeong , Myungshin Im , Seo-Won Chang , Sang-Mok Cha , Seong-Kook Lee This is my paper
classification astro-ph.IM
keywords CCDbleedingbleedtrailsKMTNetpixelmaskingsourcecompletenessphotometricaccuracyimageprocessingastronomicaldetectors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Bright stars on KMTNet CCDs leave long column-aligned bleed trails that ruin aperture placement and brighten neighboring sources. This paper gives a simple pixel-level recipe that finds the start of each trail with a brightness cut plus a bleeding index, then ends the trail with a background threshold and a short continuity rule. The two termination numbers are chosen by maximizing an F-beta score that prefers purity over recall. After the mask is applied and the masked pixels are interpolated, mean completeness versus Gaia rises from 94.08% to 98.61% and the RMS magnitude offset of carefully selected bleed-affected stars falls from 0.623 mag to 0.068 mag. The method is already running in the KS4 pipeline and is offered as a practical, model-free tool for any KMTNet user who needs cleaner photometry or fewer false detections.

What carries the argument

The bleeding index (BI): the background-subtracted sum of twenty pixels lying downstream of a bright candidate along the expected bleed direction. BI > 500, together with a conservative ADU threshold, flags true bleed-generating pixels; trail length is then set by the pair (n=0.4, CL=6) that maximizes an F-beta purity-weighted score.

What would settle it

Re-tune n and CL independently on several other KS4 fields spanning different bands and sites; if the new optima differ substantially from (0.4, 6) and the completeness or RMS gains shrink, the claimed transferability fails.

Watch

Extended reading notes

Core claim

An optimized pixel-level bleeding mask (bleeding threshold ~50 000 ADU, bleeding index >500, detection threshold n=0.4, continuity length CL=6) followed by simple interpolation restores proper aperture placement and removes the systematic brightening of nearby sources, lifting mean source completeness from 94.08% to 98.61% and reducing the RMS photometric offset of 125 selected bleed-affected stars from 0.623 mag to 0.068 mag relative to Gaia XP synthetic magnitudes.

Load-bearing premise

The two termination numbers chosen on one crowded I-band stack remain near-optimal for every other field, band, site and seeing condition.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. The paper presents a pixel-level algorithm for identifying and masking column-aligned bleeding trails in KMTNet CCD images. Bleed-generating pixels are selected with a conservative ADU threshold (Sbl ≈ 50 000 ADU, lowered for one chip) plus a bleeding index BI that sums excess flux in a 20-pixel segment along the expected trail direction; trails are then extended until CL consecutive pixels fall below a background-relative threshold nσ. Termination parameters are chosen by maximizing an Fβ-like score (β = 0.5) that balances purity against relative recovery of photometrically biased sources, using Gaia DR3 as the external reference. After SExtractor interpolation cleaning on the crowded KS4 field 1084, mean completeness versus Gaia rises from 94.08 % to 98.61 % and the RMS photometric offset of 125 selected bleed-affected sources falls from 0.623 mag to 0.068 mag. Stability of BI is checked visually across sites, bands and lunar conditions, and residual failure modes (premature termination, interpolation artifacts) are discussed.

Significance. Bleeding has been a known, unaddressed systematic in KMTNet data products used for microlensing, transients and the KS4 survey. A transparent, model-independent pixel mask that demonstrably recovers completeness and photometry against an independent catalog is therefore of immediate practical value. The algorithm is already incorporated into the KS4 DR1 pipeline and public code is provided, which strengthens reproducibility and community utility. The work is incremental rather than conceptually novel, but the quantitative Gaia-based validation and the explicit treatment of residual failure modes make it a useful contribution to the instrumentation and survey-processing literature.

major comments (2)
  1. Section 3.1.3 and Table 1: the optimal termination parameters (n = 0.4, CL = 6) are selected solely by maximizing the Fβ-like score on a single stacked I-band mosaic of the most crowded field (1084). No quantitative re-optimization or multi-field score table is provided for other bands, sites or less-crowded fields. While Section 5.1 shows that the BI threshold is visually stable, the load-bearing claim that the same (n, CL) pair remains near-optimal under varying background and seeing rests only on that limited check. A modest multi-field or multi-band score comparison (even on a subset of the 24 images already examined) would substantially strengthen the generality of the adopted mask.
  2. Section 4.2: the photometric-improvement sample is defined by several empirical cuts (ΔELLIPTICITY > 0.1, NIMAFLAGS_ISO > 0, d_bleed > 5 pix, 14 < MAG_AUTO_I < 21), yielding only 125 sources. The dramatic RMS reduction (0.623 → 0.068 mag) is therefore reported for a carefully pre-selected subset rather than for the full population of bleed-affected sources. The paper should state more clearly that the quoted RMS applies only to this regime, and ideally report the corresponding statistic (or a failure rate) for sources that fail the d_bleed cut, so that the practical impact on a typical catalog is not overstated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: parameters tuned via Fβ ranking on Gaia-matched photometry, but claimed completeness and RMS gains are independent external metrics, not forced by construction.

full rationale

The paper is a methods paper that defines a pixel-level bleed mask (Sbl, BI>500, n, CL), ranks candidate (n, CL) pairs with an Fβ-like score that uses Gaia XP synthetic magnitudes only to label Badphot vs Goodphot sources (Eqs. 4–6, Table 1), then reports two separate performance numbers against the same external catalog: (i) Gaia-source recovery fraction rising from 94.08 % to 98.61 % and (ii) RMS Δm of a differently selected sample of 125 sources falling from 0.623 mag to 0.068 mag. Neither metric is algebraically identical to the Fβ score or to any fitted parameter; the score merely chooses which mask is applied. Completeness is a detection statistic, while the photometric RMS uses an independent selection cut (ΔELLIPTICITY>0.1, NIMAFLAGS_ISO>0, distance >5 pix). Self-citations (Jeong et al. 2026, Chang et al. 2026) describe the surrounding KS4 pipeline into which the mask is inserted; they do not supply the uniqueness or functional form of the mask itself. No equation equates a claimed improvement to an input by construction, no uniqueness theorem is imported, and no ansatz is smuggled via citation. The derivation chain is therefore self-contained against an external benchmark.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central performance claims rest on a handful of hand-chosen or grid-searched numerical thresholds plus standard CCD and catalog assumptions. No new physical entities are postulated; the bleeding index is an operational diagnostic, not a new force or particle.

free parameters (6)
  • bleeding threshold Sbl = 50000 ADU (40000 for CTIO T)
    Conservative ADU cut (50 000; 40 000 for CTIO T-chip) chosen by visual inspection of all chips/sites; not derived from full-well capacity.
  • bleeding index threshold = 500
    BI > 500 adopted after visual inspection of BVRI sources on the N-chip (Fig. 4).
  • detection threshold n and continuity length CL = n=0.4, CL=6
    Grid-searched and selected by maximum Fβ-like score (β = 0.5) on one stacked I-band field (Table 1).
  • NIMAFLAGS_ISO cut = >190
    Empirical threshold (>190) used both to define bleed-affected sources for scoring and to select unsaturated stars for FWHM; set by visual inspection (Fig. 5).
  • Fβ weight β = 0.5
    Set to 0.5 to favor purity over recall; stated preference, not data-driven.
  • photometric selection cuts (ΔELLIPTICITY, d_bleed) = ΔE>0.1, d>5 pix
    ΔE > 0.1 and d_bleed > 5 pix chosen empirically to isolate the regime where photometry is expected to improve (Section 4.2).
assumptions (4)
  • domain assumption Bleed trails always propagate strictly along detector columns away from the readout register.
    Used throughout Section 3 to justify independent column scanning; standard for this CCD architecture but not re-derived.
  • domain assumption Gaia DR3 is effectively complete and photometrically reliable over 14.5 ≤ G ≤ 20.5 in the test field after proper-motion correction.
    Underpins both completeness and photometric-offset claims (Section 4); supported by gaiaunlimited estimates but still an external assumption.
  • domain assumption Sigma-clipped median and std of the full-chip image give an adequate local background and noise for the termination test.
    Equation 2; standard practice but can fail near bright sources or in highly structured backgrounds.
  • ad hoc to paper Sources with NIMAFLAGS_ISO > 190 are sufficiently contaminated by bleeding to be labeled bleed-affected for mask scoring.
    Adopted after visual inspection (Section 3.1.3); directly controls the Badphot/Goodphot counts that enter the Fβ score.
invented entities (1)
  • Bleeding Index (BI)
    purpose: Scalar diagnostic that sums background-subtracted flux in a 20-pixel segment downstream of a candidate bright pixel to decide whether that pixel truly launches a bleed trail.
    Defined in Equation 1; operational construct with no independent physical status outside the algorithm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Masking Algorithm for CCD Bleeding in Korea Microlensing Telescope Network Images." pith.science (2026). https://pith.science/paper/BZBRBN3X

@misc{pith2026260709148,
  author       = {Pith},
  title        = {Pith review of: Masking Algorithm for CCD Bleeding in Korea Microlensing Telescope Network Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZBRBN3X}},
  note         = {Machine review of arXiv:2607.09148}
}
read the original abstract

Astronomical CCD images are often affected by bleeding from bright sources, producing column-aligned streaks that degrade source detection and photometry. We present a pixel-level masking algorithm for bleeding in Korea Microlensing Telescope Network (KMTNet) images, in which these streaks extend along detector columns away from the readout register. The algorithm identifies bleed-generating pixels using a conservative bleeding threshold and a bleeding index that quantifies excess signal along the expected bleeding direction. The extent of each bleed trail is then determined by termination criteria based on a background-relative detection threshold and a continuity condition. To optimize these termination parameters, we generate multiple candidate masks and evaluate them with an F-beta-like score that favors high purity while preserving recovery of sources whose photometry is biased by bleeding. Using the final mask, we apply interpolation-based cleaning and assess the performance in terms of completeness and photometric accuracy in a crowded KMTNet field, using Gaia DR3 as a reference. The source completeness improves from 94.08% to 98.61%, primarily by enabling proper aperture placement that was previously hindered by bleeding. For selected bleeding-affected sources, the root-mean-square photometric offset decreases from 0.623 mag to 0.068 mag. These results demonstrate that the algorithm provides a practical framework for mitigating bleeding artifacts in KMTNet images. We further examine the stability of the algorithm across different observing conditions and discuss limitations of the current masking and cleaning procedure.

Figures

Figures reproduced from arXiv: 2607.09148 by the authors.

Figure 2
Figure 2. Flowchart illustrating the bleeding masking algorithm. Each parameter is explained in Section 3.1. bright pixels and extends along the detector column away from the readout register. The algorithm uses this spatial pattern by scanning each column independently and performing two tasks: identifying the onset of each bleeding streak and de￾termining where the streak terminates. Two parameters de￾termine where the blee… view at source ↗
Figure 1
Figure 1. Stacked I-band image of field 1084 in KS4, obtained at CTIO. Dashed lines indicate the boundaries of the four CCD chips. The bleed trails extend upward in the M and T-chips and downward in the K and N-chips, away from the readout registers. We refer to these preprocessed frames as single-epoch images throughout this paper. To develop and tune the masking algorithm, we selected the KS4 field 1084, centered at RA = 17… view at source ↗
Figure 3
Figure 3. Images of two sources with bleeding (above) and without bleeding (below) are shown together with their ADU profiles along the column direction (y-axis). For each source, the bleed-generating pixel (peak flux) is marked with a red cross. The pixels used to compute BI are indicated by the yellow dashed line in the image and by yellow stars and a shaded region in the ADU profile. The green dashed line in the profile ma… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: BV RI images of sources with increasing BI values from left to right. In each panel, the candidate bleed-generating pixel (peak flux) is marked with a red cross, and the 20-pixel segment used to compute BI is indicated by a yellow dashed line. The 20-pixel segments for…
Figure 5
Figure 5. Figure 5: Sources are shown as NIMAFLAGS_ISO increases from left to right. The NIMAFLAGS_ISO value (NIMF) is indicated in the upper￾right corner of each panel, and the target source is marked with a yellow cross. From visual inspection, we confirmed that sources with NIMAFLAGS_I…
Figure 6
Figure 6. Figure 6: Stacked images combining the K-, M-, T-, and N-chips (left) and the corresponding bleed-cleaned image after interpolation (right). The remaining column-aligned artifacts in the right panel are residual bleed trails after interpolation (Section 5.2.2) or crosstalk (Kim …
Figure 7
Figure 7. Figure 7: Source completeness of the 1084 field in the I-band, evaluated relative to Gaia DR3. The upper row shows results from the original image, and the lower row shows results from the bleed-cleaned image. The left panels present completeness as a function of Gaia G magnitud…
Figure 9
Figure 9. Figure 9: ∆m of sources in the bleed-cleaned image, as a function of distance from the bleeding mask. The samples are restricted to sources with ∆ELLIPTICITY > 0.1 between the original and bleed￾cleaned images and NIMAFLAGS_ISO > 0. jects near the KMTNet limiting magnitude (Chan…
Figure 8
Figure 8. Figure 8: APERTURES CHECKIMAGEs from SExtractor for three rep￾resentative sources. In each row, the left and right panels correspond to the original and bleed-cleaned images, respectively, and the red dashed circle marks the target source. The top row shows a source recovered ne…
Figure 10
Figure 10. Figure 10: Comparison of KMTNet I-band MAG_AUTO and synthetic I-band magnitudes derived from the Gaia XP spectra. The upper panel shows the photometric comparison, and the lower panel shows the offset ∆m. Blue and red points indicate measurements from the original and bleed-clea…
Figure 11
Figure 11. Figure 11: FWHM distributions of unsaturated, bleed-unaffected sources derived from five images for each site (SAAO, SSO, and CTIO) in each of the B, V , R, and I bands. The orange dashed line in each panel indicates the median FWHM. Original Image 1 591 2 273 3 -351 4 -478 5 -5…
Figure 12
Figure 12. Figure 12: Example of residual bleeding in the cleaned image caused by premature termination of the masking procedure. From left to right, the panels show: (1) a segment of the original bleed trail containing a short sequence of anomalously depressed pixels; (2) the same column …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references

  1. [1]

    R., et al

    Afsariardchi, N., Moon, D.-S., Drout, M. R., et al. 2019, ApJ, 881, 22

    2019

  2. [2]

    P., Ishiguro, M., Jin, S., et al

    Bach, Y. P., Ishiguro, M., Jin, S., et al. 2019, JKAS, 52, 71

    2019

  3. [3]

    2010, SWarp: Resampling and Co-adding FITS Images

    Bertin, E. 2010, SWarp: Resampling and Co-adding FITS Images

    2010

  4. [4]

    1996, A&AS, 117, 393

    Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393

    1996

  5. [5]

    2023, A&A, 669, A55

    Cantat-Gaudin, T., Fouesneau, M., Rix, H.-W., et al. 2023, A&A, 669, A55

    2023

  6. [6]

    2018, JKAS, 51, 129

    Chang, S.-W., Byun, Y.-I., Shin, M.-S., et al. 2018, JKAS, 51, 129

    2018

  7. [7]

    2026, JKAS, 59, 179

    Chang, S.-W., Im, M., Jeong, M., et al. 2026, JKAS, 59, 179

    2026

  8. [8]

    2018, PASJ, 70, S7

    Coupon, J., Czakon, N., Bosch, J., et al. 2018, PASJ, 70, S7

    2018

Show all 34 references
  1. [9]

    J., Bertin, E., et al

    Desai, S., Mohr, J. J., Bertin, E., et al. 2016, A&C, 16, 67 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1 GaiaCollaboration,Vallenari,A.,Brown,A.G.A.,etal.2023,A&A, 674, A1

    2016

  2. [10]

    2022, A&A, 664, A13

    Gould, A., Han, C., Zang, W., et al. 2022, A&A, 664, A13

    2022

  3. [11]

    2022, AJ, 163, 43

    Hwang, K.-H., Zang, W., Gould, A., et al. 2022, AJ, 163, 43

    2022

  4. [12]

    J., et al

    Im, M., Yoon, Y., Lee, S.-K. J., et al. 2017, ApJL, 849, L16 1https://github.com/jmk5040/KMTNet_ToO.git 2https://github.com/JiseopShin17/Bleeding-Masking-Algorithm Shin et al. 12 Masking Algorithm for CCD Bleeding

    2017

  5. [13]

    2026, JKAS, 59, 157

    Jeong, M., Im, M., Kim, J., et al. 2026, JKAS, 59, 157

    2026

  6. [14]

    2009, PKAS, 24, 83

    Kim, D.-J., Lee, C.-U., Kim, S.-L., et al. 2009, PKAS, 24, 83

    2009

  7. [15]

    2013, PKAS, 28, 1

    Kim, D.-J., Lee, C.-U., Kim, S.-L., et al. 2013, PKAS, 28, 1

    2013

  8. [16]

    Kim, J., Im, M., Paek, G. S. H., et al. 2021, ApJ, 916, 47

    2021

  9. [17]

    2018, JKAS, 51, 89

    Kim, J., Karouzos, M., Im, M., et al. 2018, JKAS, 51, 89

    2018

  10. [18]

    2025, AJ, 169, 261

    Kim, S.-L., Lee, C.-U., Hong, K., et al. 2025, AJ, 169, 261

    2025

  11. [19]

    2024, ApJS, 275, 46

    Kim, Y., Kim, M., Im, M., et al. 2024, ApJS, 275, 46

    2024

  12. [20]

    Lee, D., Paek, G. S. H., Chang, S.-W., et al. 2025, AJ, 170, 337

    2025

  13. [21]

    2025, Wide Field Camera 3 Instrument

    Marinelli, M., & Green, J. 2025, Wide Field Camera 3 Instrument

    2025

  14. [22]

    A., Menanteau, F., et al

    Morganson, E., Gruendl, R. A., Menanteau, F., et al. 2018, PASP, 130, 074501

    2018

  15. [23]

    Q., Moon, D.-S., Drout, M

    Ni, Y. Q., Moon, D.-S., Drout, M. R., et al. 2022, NatAs, 6, 568

    2022

  16. [24]

    Paek, G. S. H., Im, M., Jeong, M., et al. 2025, ApJ, 981, 38

    2025

  17. [25]

    Paek, G. S. H., Im, M., Jeong, M., et al. 2026, ApJ, 1000, 53

    2026

  18. [26]

    Paek, G. S. H., Im, M., Kim, J., et al. 2024, ApJ, 960, 113

    2024

  19. [27]

    2025, JINST, 20, P07031

    Polin, D., Snyder, A., Lage, C., et al. 2025, JINST, 20, P07031

    2025

  20. [28]

    2016, JKAS, 49, 73

    Shin, I.-G., Ryu, Y.-H., Udalski, A., et al. 2016, JKAS, 49, 73

    2016

  21. [29]

    2008, Instrument Science Report HLA 2008-01

    Stankiewicz, M., Gonzaga, S., & Whitmore, B. 2008, Instrument Science Report HLA 2008-01

    2008

  22. [30]

    2017, Nature, 551, 71 Viaña, J., Hwang, K.-H., de Beurs, Z., et al

    Troja, E., Piro, L., van Eerten, H., et al. 2017, Nature, 551, 71 Viaña, J., Hwang, K.-H., de Beurs, Z., et al. 2025, AJ, 169, 159

    2017

  23. [31]

    Z., Magnier, E

    Waters, C. Z., Magnier, E. A., Price, P. A., et al. 2020, ApJS, 251, 4

    2020

  24. [32]

    Winecki, D., & Kochanek, C. S. 2024, ApJ, 971, 61

    2024

  25. [33]

    2021, AJ, 162, 163

    Zang, W., Hwang, K.-H., Udalski, A., et al. 2021, AJ, 162, 163

    2021

  26. [34]

    K., et al

    Zhang, J., Zang, W., Jung, Y. K., et al. 2023, MNRAS, 522, 6055 Shin et al. 13

    2023

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.