Speaker
Description
Over the past two decades, nearly all Near-Earth Objects (NEOs) have been discovered mostly by dedicated surveys through the use of preselected candidates posted to the Minor Planet Center's (MPC) Near-Earth Object Confirmation Page (NEOCP). Rapid follow-up observations from the astronomical community typically allow for the designation of new NEOs within a few days. As of today, more than 37,000 NEOs have been discovered, with an average discovery rate of approximately 3,000 per year since 2020. Candidate preselection is based on the NEO digest2 score [1,2], with only short-arc tracklets achieving a score of NEO digest2 $>$ 65 being posted to NEOCP. Previous studies [1] have shown that NEOs generally achieve digest2 scores close to 100, while other orbital classes exhibit significantly lower scores. For instance, typical main-belt objects have scores near zero. The frequency of NEOs sharply drops as a function of decreasing digest2 NEO score on NEOCP. Annually, about 6,000 candidates are posted to NEOCP, with roughly 10\% remaining unconfirmed due to a lack of follow-up observations [3]. Among confirmed candidates, two-thirds are designated as NEOs, while the remainder are predominantly main-belt objects.
We present a detailed analysis of 13 distinct digest2 orbital categories , evaluated in two modes—'raw' and 'noid'—for NEOCP candidates since 2019. Our aim is to reduce the proportion of non-NEOs on NEOCP. By leveraging all derived digest2 parameters, rather than relying solely on the NEO digest2 score, we demonstrate the potential to eliminate up to 20\% of non-NEO candidates.
Furthermore, we applied four machine learning (ML) methods that we already studied on simulated short-arc LSST tracklets [4]: Gradient Boosting Machine (GBM), Random Forest (RF) classifier, Stochastic Gradient Descent (SGD) classifier, and Neural Network (NN)—to the derived digest2 scores for NEOCP candidates from 2019 onward. Using NEOCP candidates observed between 2019 and 2023 as the training dataset, and 2024 candidates as the validation dataset, we achieved NEO prediction accuracies of 91\%-92\%, with model performance differing by at most one percentage point.
We propose that the implementation of derived digest2 filters and ML methods could significantly reduce the number of non-NEO candidates on NEOCP. This would provide more efficient use of follow-up observation time and decrease the fraction of unconfirmed NEOCP candidates, most of which are likely NEOs and therefore increase the number of NEO discoveries.
References:
[1] S. Keys, P. Veres, M. J. Payne, M. J. Holman, R. Jedicke, G. V. Williams, T. Spahr, D. J. Asher, C. Hergenrother, The digest2
neo classification code, Publications of the Astronomical Society of the Pacific 131 (2019) 1–22.
[2] P. Veres, R. Cloete, R.Weryk, A. Loeb, M. J. Payne, Improvement of Digest2 NEO Classification Code-utilizing the Astrometry
Data Exchange Standard, 135 (2023) 104505.
[3] P. Veres, M. J. Payne, M. J. Holman, D. Farnocchia, G. V. Williams, S. Keys, I. Boardman, Unconfirmed Near-Earth Objects,
156 (2018) 5.
[4] R. Cloete, P. Veres, A. Loeb, Machine learning methods for automated interstellar object classification with LSST, 691 (2024)
A338.
2