Identification and Localization of Cometary Activity in Solar System Objects with Machine Learning

TL;DR

The chapter surveys how cometary activity detection in Solar System objects has evolved from classical, artifact-susceptible methods to modern machine-learning approaches. It highlights per-image PSF and coma-based diagnostics, and then details CNN-based detectors like the Tails network that leverage difference imaging and active learning to achieve low error rates on all-sky survey data. Key contributions include documenting the training regimes, performance metrics, and real-world AI-assisted discoveries (e.g., C/2020 T2 Palomar), and outlining future paths that combine CNNs with orbital/PSF features for Rubin LSST-scale data. Collectively, the work demonstrates ML’s potential to reliably identify and localize extended cometary activity across heterogeneous surveys, accelerating discovery in the upcoming era of large, time-domain astronomy.

Abstract

In this chapter, we will discuss the use of Machine Learning methods for the identification and localization of cometary activity for Solar System objects in ground and in space-based wide-field all-sky surveys. We will begin the chapter by discussing the challenges of identifying known and unknown active, extended Solar System objects in the presence of stellar-type sources and the application of classical pre-ML identification techniques and their limitations. We will then transition to the discussion of implementing ML techniques to address the challenge of extended object identification. We will finish with prospective future methods and the application to future surveys such as the Vera C. Rubin Observatory.

Figures (9)

• Figure 1: Discovery images of C/2022 E3 (ZTF) taken by ZTF in 2022 March. The top panel shows one of the individual detections of the comet (taken in a series of 5). The top right panel shows a composite stack of the first and fifth images stacked on the background stars. The bottom panel shows a composite stack of all five images co-added together. The exposure time, filter, and cardinal directions are indicated. Adapted from Figure 1 of Bolin2024E3 and reproduced with the permission of the authors and MNRAS. The bottom panel shows a co-added composite stack of all five 30 s exposures of images of comet P/2022 P2 (ZTF) taken by ZTF in 2022 August Bolin2024TS.
• Figure 2: Comparision of the PSF extent, psfextent, and PSF quality, psfquality, properties of asteroid and comet detections from Pan-STARRS taken before 2013 July. psfextent refers to the extendedness of detection PSFs measured as the full width at half maximum of the PSF in arcseconds compared to the expected PSF of point sources. Higher values of psfextent correspond to sources being more extended. psfquality refers to the reliability of PSF measurements with a higher number being more favorable. Top panel: 1-D histograms of PSF extent of asteroid detections (grey) and comet detections (blue). Bottom panel: 2-D distribution of psfextent vs psf quality. The psfextent and psfquality for asteroids is in grey, comets in blue/orange, and Main Belt comets/active asteroids in red and green. Adapted from Figure 5 of Hsieh2015 and reproduced with the permission of the authors and Icarus. We refer readers to the asteroid and comet detection gallery seen in Fig. 18 of Denneau2013 for examples of asteroidal and extended comet detections.
• Figure 3: Azimuth scheme for the detection of tails of comets in individual CCD images. The azimuth ring is divided into 18 sections. The section at 10 o'clock encompassing the tail of the detected comet, 133P, measures a higher degree of flux compared to the other sections. The radius of the azimuthal sections was chosen to avoid background stars. Adapted from Figure 2 of Sonnett2011 and reproduced with the permission of the authors and Icarus.
• Figure 4: Secular brightness evolution of active asteroid (6478) Gault in 2017-2019. The left and right panels show the evolution in the comet's brightness (red and green dots) compared to brightness models for a bare nucleus (red and green dashed lines). The brightness evolution of Gault in 2017 follows a typical pattern for a bare nucleus. The brightness pattern significantly deviates and brightens compared to a bare nucleus in 2018-2019. Adapted from Figure 2 of Ye2019Gault and reproduced with the permission of the authors and AAS Journals.
• Figure 5: Illustration of the Tails network architecture. Tails uses a combination of bidirectional feature pyramid networks (BiFPN), which are fed by the last five layers of an EfficientDet network. The output of the BiFPN is fed into a head network consisting of several convolutional layers and a fully connected layer and activation function to derive a comet probability score between 0 and 1. Adapted from Fig. 3 of Duev2021 and reproduced with the permission of the authors and AAS Journals.