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

TL;DR

This work automates the detection of Galactic optical cirrus filaments in deep HSC-SSP DR3 images using an ensemble of CNN-based segmenters, building on prior Stripe 82 results. It demonstrates that deeper data reveal roughly 4.5 times more filaments while exposing sky background over-subtraction as a limiting factor for faint features. The authors create high-quality cirrus maps and two filament catalogs, highlight regional variations, and quantify how over-subtraction biases sky estimates around filaments. The approach yields a practical framework and data products to improve sky subtraction and cirrus-aware analyses in current and future deep optical surveys.

Abstract

The existence of Galactic optical cirrus poses a challenge for observing faint objects within our Galaxy and dim extragalactic structures. To investigate individual cirrus filaments in the Hyper Suprime-Cam Subaru Strategic Program public data release 3 (HSC-SSP DR3) we use a technique based on convolutional neural networks and ensemble learning. This approach allows us to distinguish cirrus filaments from foreground and background objects across the entire HSC-SSP, using optical images in the $g$, $r$, and $i$ wavebands. A comparison with previous work using deep Sloan Digital Sky Survey Stripe~82 (SDSS Stripe~82) data reveals that the cirrus clouds identified in this study are highly consistent in location within the overlapping survey region. However, in the deeper HSC-SSP dataset, we were able to detect $4.5$ times more cirrus clouds. Our study indicates that the sky background in HSC-SSP coadd images is over-subtracted, as evidenced by the surface brightness distribution in cirrus filaments and surrounding regions. Objects with surface brightness of $m = 29~\mbox{mag~arcsec}^{-2}$ near large filaments can be dimmed by over-subtraction of $0.5$ magnitude in the $r$ band. This suggests that cirrus clouds should be taken into account in algorithms for estimating the sky background. For practical use, we provide a catalog of filaments and a framework that allows one to train neural network models for segmenting cirri in HSC-SSP coadd images.

Figures (21)

• Figure 1: "global-sky" combined image of the Field f1053 in the $r$ band. Contamination by sky over-subtraction is clearly seen in areas around cirrus filaments (darker areas).
• Figure 2: Schematic map of the HSC-SSP Fall region. The step of the coordinate grid along the DEC is $1^{\circ}$, and on the RA it is $10^{\circ}$. The SDSS Stripe 82 region is bounded by the dashed blue line. The HSC-SSP Fall region is bounded by the red solid line and divided into 3 regions. A hatched region is called the Fall plus region, a solid gray region is called the Intersection region, and a region filled with zigzag lines is called the Fall minus region.
• Figure 3: Diagram of the Intersection dataset preparation process. The top row of the diagram demonstrates the creation of a combined mask by the union of a neural mask and a source mask. The direct sum sign denotes the creation of a cleaned image in the $r$ band from an image in the same band with a combined mask. The dotted arrow demonstrates the creation of a cirrus map using manually selected and corrected contours in DS9.
• Figure 4: The encoder-decoder architecture used for creating cirrus maps.
• Figure 5: Diagram of the prediction of the cirrus map by the ensemble of nine best $4$-channel models. In the left part of the diagram, the data used for model training is shown. It consists of ground-truth cirrus map created using manually selected and corrected contours in DS9 (manual cirrus map) and images from which the input tensor is formed (cleaned $r$-band image and images in $g$, $r$, $i$ bands). Rounded rectangles in the central part of the diagram denote trained models. Gray rounded rectangles denote the best nine models selected for the ensemble. The arrow in the right part of the diagram represents the generation of a cirrus map through direct pixel-by-pixel majority voting by ensemble models.