Cosmic Ray Detection and Rejection for CSST

TL;DR

This work addresses cosmic ray contamination in CSST single-exposure surveys by developing a single-image CR processing pipeline. It retrains a DeepCR detector per detector for high-accuracy CR identification and introduces a morphology-sensitive inpainting stage using UNet++ to repair CRs on stars and galaxies, with adaptive median filtering handling background regions. The approach yields recall $=97.90 \pm 0.18\%$ and precision $=98.67 \pm 0.05\%$ for CR detection and substantially improves photometric accuracy, increasing detectable sources by $13.6\%$ and achieving magnitude-distribution fidelity close to ground truth (galaxy SSIM $=0.914$, star SSIM $=0.749$). The method offers robust CR mitigation for CSST and has potential for transfer to other exposure-limited space missions, enabling reliable science even when multi-frame stacking is impossible.

Abstract

As a space telescope, the China Space Station Survey Telescope (CSST) will face significant challenges from cosmic ray (CR) contamination. These CRs will severely degrade image quality and further influence scientific analysis. Due to the CSST's sky survey strategy, traditional multi-frame stacking methods become invalid. The limited revisits prompted us to develop an effective single-image CR processing method for CSST. We retrained the DeepCR model based on CSST simulated images and achieved 97.90+-0.18% recall and 98.67+-0.05% precision on CR detection. Moreover, this paper puts forward an innovative morphology-sensitive inpainting method, which focuses more on areas with higher scientific value. We trained a UNet++ model especially on contaminated stellar/galactic areas, alongside adaptive median filtering for background regions. This method achieves effective for CRs with different intensities and different distances from centers of scientific targets. By this approach, the photometric errors of CR-corrected targets could be restricted to the level comparable to those of uncontaminated sources. Also, it increases the detection rate by 13.6% compared to CR masking. This method will provide a robust CR mitigation for next-generation space telescopes.

Figures (7)

• Figure 1: The examples of the training dataset for CR rejection. The top row displays a star while the bottom row shows one galaxy. The left column features true images without CRs, the middle column shows CR samples (after shifting), and the right column displays combined images with CR interference.
• Figure 2: Recall (top) and precision (bottom) of CR detection across 18 detectors. Data points and numbers represent median values calculated from 1,221 experimental trials, with asymmetric error bars indicating the 25th–75th percentile range. Detector indices (6–25) are shown on the x-axis, with missing detectors 10 and 21 identified as slitless spectroscopic instruments.
• Figure 3: Recall distribution versus CR intensity. Each panel corresponds to one detector, showing median recall rates from 1,221 exposure images. The horizontal axis displays CR intensity on a logarithmic scale. The vertical axis shows recall rate (uniform range: 0.75-1.0) for cross-detector comparison. The red dashed line marks the $5 e^{-}$ readout noise (RON) threshold.
• Figure 4: CR intensity distribution from simulated images. The left y-axis displays frequency distributions: brown bars represent all CRs, while purple bars represent track-like CRs. The right y-axis shows the ratio of track-like to all CR counts (black line).
• Figure 5: Photometric accuracy between CR inpainting methods for individual sources. Left: original images of sources. Middle: magnitude differences versus CR distances (pixels) for CR-contaminated images (blue), corrected images (red) and median-filter corrected images (green). Solid lines show binned means, shaded regions indicate ±1 standard deviation, and raw data points display individual measurements. Right: magnitude differences versus CR intensity for the same three cases. Vertical scales adapt to magnitude ranges.