Newly Identified Planetary Nebula Candidates in M31 using Deep-Learning Search of Hubble Space Telescope Images

Abstract

The Hubble Space Telescope has imaged approximately a quarter of a billion stars in M31's disc, yet only slightly over a thousand planetary nebulae (PNe) are known, mostly from ground-based surveys. We show that deep learning can find them in broad-band images alone. In what is the first such search, we scan the Panchromatic Hubble Andromeda Treasury (PHAT) and the Panchromatic Hubble Andromeda Southern Treasury (PHAST) mosaics, nearly one square degree of the disk, where a PN is a faint blue point among crowded stars. Because the existing catalogs are incomplete, we train four Faster R-CNN detectors with partial supervision, so that uncataloged PNe are not taught as background, and combine their rankings into one candidate list, an ensemble. On spatially separate test fields, this list places 73% of the known PNe in its top 200, 4 to 6 times the purity of classical color-magnitude selection. Newer detection transformers and segmentation networks trained on the same data do not surpass our detectors. The ranking also follows classifications that the detectors never saw: emission-line objects that the literature calls PNe enter the list at 47%, against 0.6% for those called H II regions, which makes them candidates for reclassification. Among the top 400 candidates, 175 appear in no published catalog, 12 of them flagged as cosmic-ray residuals. One of them, a point source with [O III] emission in nine independent exposures, meets the PN criteria of the three major PN studies of M31, so by those criteria it is a PN. Most of the others lack strong [O III], as expected if narrow-band surveys have found the bright PNe, although some may be PNe with weak [O III]. These results support the near completeness of ground-based surveys for bright PNe and extend it from the PHAT area to PHAST. Finally, we identify evaluation pitfalls for detectors of rare objects and recommend simple remedies.