Analysis of Galactic cirrus filaments in HSC-SSP high-resolution deep images using artificial neural networks

This study utilizes convolutional neural networks and ensemble learning on deep HSC-SSP images to identify significantly more Galactic cirrus filaments than previous surveys, revealing that sky background over-subtraction in coadd images dims faint objects near these structures and necessitating their inclusion in background estimation algorithms.

The Gist

Imagine you are trying to take a super-clear photo of a distant, faint firefly in a dark forest. But there's a problem: the forest itself has a thin, wispy mist (Galactic cirrus) that drifts in front of your camera lens. This mist is made of cosmic dust, and it's so faint that it's hard to see, but it's bright enough to blur your photo of the firefly.

This paper is about a team of astronomers who built a smart digital detective to find and map this cosmic mist in a massive new set of telescope photos.

Here is the breakdown of their adventure, explained simply:

1. The Problem: The "Cosmic Fog"

Astronomers use powerful telescopes (like the Subaru Telescope in Japan) to take incredibly deep pictures of the universe. They want to see the faintest, oldest galaxies. But our own Milky Way galaxy is filled with "cirrus clouds"—not made of water vapor like Earth's clouds, but of dust and gas.

These clouds are tricky. They are so faint that standard computer programs often mistake them for empty space or, worse, they mess up the calculation of the "background darkness" of space. If the computer gets the background wrong, it might make a real galaxy look dimmer than it is, or hide it entirely.

2. The Solution: A Team of AI Detectives

In the past, finding these clouds was like trying to find a specific needle in a haystack by looking at the whole haystack at once. It was slow and prone to errors.

The authors decided to use Artificial Intelligence (AI), specifically a type called Convolutional Neural Networks. Think of this AI as a highly trained dog that has been shown thousands of pictures of "mist" and "not-mist."

• The Training: They taught the AI using a smaller, well-known set of photos (from the SDSS survey).
• The Upgrade: They then took this AI and gave it a new, much sharper set of photos (from the HSC-SSP survey).
• The Ensemble Trick: Instead of relying on just one AI dog, they trained nine slightly different AI dogs. When they needed to identify a cloud, they asked all nine dogs, and they took a "majority vote." If 7 out of 9 dogs said, "That's a cloud!" then it was a cloud. This "teamwork" approach made the detection much more accurate.

3. The Discovery: Seeing the Invisible

Because the new telescope photos are about four times deeper (they can see fainter things) than the old ones, their AI team found 4.5 times more cirrus clouds than ever before.

They created a map (a catalog) of these clouds across a huge patch of the sky. It's like drawing a weather map for the Milky Way, showing exactly where the "cosmic fog" is thick and where the air is clear.

4. The Big Surprise: The "Over-Subtraction" Glitch

Here is the most interesting part of the story. When the astronomers compared their new, high-resolution maps with the old ones, they noticed something weird.

In the new, deep photos, the cosmic dust clouds looked dimmer than they should have.

The Analogy: Imagine you are trying to measure the brightness of a candle in a room. You have a machine that tries to measure the "background light" of the room to subtract it from your reading.

• If the room has a thin, invisible fog (the cirrus), the machine gets confused.
• It thinks the fog is part of the "background" and tries to subtract it too aggressively.
• The result? The machine subtracts the fog and some of the candle's light, making the candle look darker than it really is.

The authors found that in these deep images, the computer was "over-subtracting" the background. Near large clouds, this error could make faint objects look 0.5 magnitudes dimmer. That's a significant difference! It means that without fixing this, we might be missing or misjudging the brightness of very faint galaxies.

5. Why This Matters

This paper is a "user manual" for future astronomers.

• The Map: They are giving away their catalog of clouds so other scientists can avoid these "foggy" areas or correct their data.
• The Warning: They are telling everyone, "Hey, when you look at these super-deep photos, remember that the background subtraction might be too aggressive near dust clouds. You need to account for this!"
• The Future: As we prepare for even bigger telescopes (like the Vera C. Rubin Observatory), knowing how to spot and correct for this "cosmic fog" will be essential to finding the universe's faintest secrets.

In a nutshell: The authors built a super-smart AI team to map the invisible dust clouds in our galaxy. They discovered that these clouds are much more common than we thought, and they found a glitch in how we process deep-space photos that makes faint objects look too dark. They've shared their maps and warnings to help everyone see the universe more clearly.

Technical Summary

Here is a detailed technical summary of the paper "Analysis of Galactic cirrus filaments in HSC-SSP high-resolution deep images using artificial neural networks."

1. Problem Statement

Galactic cirrus clouds are diffuse, filamentary structures of interstellar dust that pose a significant challenge for deep astronomical imaging. They contribute to the Diffuse Galactic Light (DGL) and can obscure or mimic faint extragalactic sources (e.g., tidal features, low-surface-brightness galaxies).

• The Challenge: Accurately identifying and segmenting these filaments in deep, wide-field surveys is difficult due to their fractal nature, low surface brightness, and the presence of aggressive sky background subtraction algorithms that often create artifacts (over-subtraction) around these structures.
• The Gap: Previous catalogs (e.g., in SDSS Stripe 82) were limited by shallower depth and resolution. There was a need to extend this analysis to the deeper, higher-resolution data provided by the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) to create a more comprehensive catalog and understand the impact of cirrus on background estimation.

2. Methodology

The authors employed a deep learning approach based on Convolutional Neural Networks (CNNs) and Ensemble Learning to automate the segmentation of cirrus filaments.

Data Preparation

• Dataset: Used HSC-SSP Public Data Release 3 (DR3) "global-sky" coadd images in the $g$, $r$, and $i$ bands.
• Coverage: Analyzed 4,331 fields across five regions (Intersection, Fall plus, Fall minus, Spring, North), covering approximately 1,050 deg².
• Preprocessing:
  • Masking: Created "combined masks" by merging HSC-SSP source masks with "neural masks" generated by a conditional GAN. This ensured all non-cirrus sources (stars, galaxies) were masked out, leaving only the cirrus and background.
  • Cleaning: Generated "cleaned" $r$-band images by zeroing out masked regions and interpolating the flux from the boundaries to remove source contamination while preserving cirrus morphology.
  • Ground Truth: Manually created high-quality cirrus segmentation maps (cirrus maps) for 70 fields using DS9 contour tools at the 29 mag arcsec$^{-2}$ isophote level. These served as the training labels.

Neural Network Architecture & Training

• Architecture: Utilized a U-Net encoder-decoder architecture with MobileNetV2 as the encoder.
• Input: 3-channel ($g, r, i$) or 4-channel ($g, r, i$, and cleaned $r$) images.
• Training Strategy:
  • Fine-tuning: Attempts to fine-tune models trained on SDSS Stripe 82 data failed to achieve high performance due to differences in pixel scale (0.168" vs 0.396") and data depth.
  • Training from Scratch: Models were trained from scratch on the HSC-SSP Intersection dataset.
  • Data Augmentation: Applied symmetry group transformations (rotations, reflections) and min-max normalization.
  • Ensemble Learning: To overcome the limitations of small dataset size and improve robustness, the authors trained multiple models and combined their predictions using pixel-by-pixel majority voting. The best ensemble consisted of nine top-performing 4-channel models.

Post-Processing

• Artifact Removal: Manually removed false detections caused by scattered light patterns and other artifacts from the ensemble-generated maps.
• Validation: Used Intersection over Union (IoU), Precision, and Recall metrics.

3. Key Contributions

1. First Deep Optical Cirrus Catalog for HSC-SSP: Created the first comprehensive catalog of Galactic cirrus filaments for the HSC-SSP DR3 data, covering a significantly larger area and greater depth than previous SDSS-based studies.
2. Ensemble Learning Framework: Demonstrated that an ensemble of nine U-Net models significantly outperforms individual models, achieving an IoU of 0.48 (compared to ~0.43 for single models and 0.576 for the best SDSS model, though direct comparison is complicated by data differences).
3. Discovery of Sky Over-Subtraction: Identified and quantified a systematic sky background over-subtraction in HSC-SSP coadd images specifically around large cirrus filaments.
4. Public Release: Released the rectified cirrus maps and filament catalogs via a GitLab repository for community use.

4. Key Results

Detection Performance and Statistics

• Detection Rate: The HSC-SSP dataset revealed 4.5 times more cirrus clouds than the overlapping SDSS Stripe 82 data, attributed to the ~1 magnitude deeper depth of HSC-SSP.
• Consistency: In the overlapping "Intersection" region, the spatial distribution of identified cirrus showed high consistency with previous SDSS maps (IoU = 0.105, which is close to the theoretical maximum given the difference in detected fractions).
• Filament Properties:
  • Size: Typical filament sizes are approximately 0.5–1.0 arcmin² (at 6-arcsec resolution), consistent with previous findings.
  • Surface Brightness: Most filaments are well above the 3$\sigma$ noise limit, though small filaments are susceptible to noise and over-subtraction artifacts.
  • Correlation: Filaments with statistically significant correlations ($p < 0.01$) between $g, r,$ and $i$ bands show positive correlation coefficients ($\rho > 0.5$), confirming their nature as diffuse dust structures rather than noise.

The Over-Subtraction Issue

• Observation: A pixel-by-pixel comparison between HSC-SSP and SDSS Stripe 82 revealed that HSC-SSP fluxes are systematically lower (dimmer) in cirrus regions.
• Quantification: Analysis of boundary layers adjacent to filaments showed a median pixel value deficit.
  • For objects with a surface brightness of $m = 29$ mag arcsec$^{-2}$ near large filaments, the over-subtraction causes a dimming of ~0.5 magnitudes in the $r$-band.
  • The over-subtraction is most severe 50–150 pixels away from the filament edges.
• Implication: This suggests that current "global-sky" subtraction algorithms in HSC-SSP (and potentially other deep surveys) fail to account for the extended nature of cirrus, leading to biased photometry for faint background objects.

5. Significance and Future Impact

• Astronomical Data Quality: The study highlights a critical systematic error in deep optical surveys: the over-subtraction of sky background near cirrus. This affects the photometric accuracy of faint extragalactic sources.
• Algorithmic Improvement: The authors propose that future sky background estimation algorithms must explicitly account for cirrus clouds to prevent over-subtraction. The generated cirrus maps can serve as masks or priors for these algorithms.
• Survey Planning: The results are relevant for upcoming deep surveys like Euclid and the Vera C. Rubin Observatory (LSST), where accurate background subtraction is crucial for detecting ultra-faint structures.
• Methodology: The successful application of ensemble learning to a small, manually annotated dataset provides a robust framework for other astronomical segmentation tasks where ground truth is expensive to produce.

In conclusion, this paper not only expands the known census of Galactic cirrus but also uncovers a significant systematic bias in current deep imaging data processing, providing both the data (catalogs) and the methodological insights (ensemble CNNs) necessary to mitigate these effects in future astronomical research.