A New Cosmic Ray Rejection Routine for HST WFC3/UVIS via label-free training of deepCR
Zhuo Chen
Oct 8, 2023
2 min read
Figure 1: Top left: WFC3/UVIS calibrated dark frame (HST:16568). Top right: WFC3/UVIS image in F336W of M31 (HST:16801). Bottom left: The result of adding the dark frame to the raw image. Bottom right: The same image with all CRs removed by deepCR, including both from the raw image and dark frame. Masked pixel values are filled with a 3 × 3 median filter.
deepCR is a deep-learning-based cosmic ray (CR) rejection framework originally presented by Zhang & Bloom (2020). The original approach requires a dedicated training set that consists of multiple frames of the same fields, enabling automatic CR labeling through comparison with their median co-adds. Here, we present a novel training approach that circumvents the need for a dedicated training set, but instead utilizes dark frames and the science images requiring CR removal themselves. During training, CRs present in dark frames are added to the science images, which the network is then trained to identify. In turn, the trained deepCR model can then be applied to identify CRs originally present in the science images. Using this approach, we present a new deepCR model trained on a diverse set of Hubble Space Telescope (HST) images taken from resolved galaxies in the local group, which is universally applicable across all WFC3/UVIS filters.
Figure 2: Top left: Fraction of total number of GST stars from one example HST visit as applied with different deepCR masking thresholds compared to the legacy mode. The resulting number of GST stars drops quickly below the threshold of 0.1. Top right: Fraction of GST stars that are added or lost with different deepCR masking thresholds, respectively, which together construct the overall curve on the left figure. Bottom left: UV GST CMD. The red box marks the CMD features of the most importance. Bottom right: Fraction of GST stars that are added or lost within the CMD featured box as a function of deepCR masking thresholds. Dashed lines represent the results when the featured box is dithered to avoid selection edge effects. deepCR masking keeps adding more GST stars within the CMD featured region until the threshold is around 0.08, while the masking starts to lose featured stars prominently below the threshold of 0.1.
