ADASSX

Over the last twenty years, the discovery of Near-Earth Objects (NEOs) has relied heavily on dedicated survey programs and the subsequent vetting of candidates via the Minor Planet Center’s Near-Earth Object Confirmation Page (NEOCP). This platform plays a central role in the rapid identification of NEOs by publishing short-arc tracklets for immediate follow-up by the astronomical community. Thanks to the rapid follow-up, over 38,000 NEOs have been cataloged to date, with discovery rates exceeding 3,000 per year since 2020.

Candidate selection for NEOCP posting is primarily based on the NEO digest2 score — a probability metric estimating whether a given object is an NEO. Tracklets with a digest2 score above 65 are qualified for posting, as NEOs typically score close to 100, while other populations, such as main-belt asteroids, tend to score much lower. Nevertheless, roughly 6,000 candidates appear on the NEOCP annually, of which approximately 11% remain unconfirmed due to insufficient follow-up. Among those confirmed, only about two-thirds are ultimately classified as NEOs, with the rest largely consisting of main-belt objects.

In this study, we perform a systematic evaluation of 13 digest2-derived orbital classification categories, analyzing them in both their "raw" and "noid" configurations. Our objective is to improve the efficiency of the NEOCP by reducing the prevalence of non-NEOs among posted candidates. We show that by incorporating the full set of digest2-derived parameters—rather than relying solely on the NEO digest2 score—it is possible to filter out up to 20% of non-NEO submissions without significantly impacting true NEO recoverability.

Additionally, we explore the predictive capabilities of several machine learning (ML) classifiers—Gradient Boosting Machines (GBM), Random Forests (RF), Stochastic Gradient Descent (SGD), and Neural Networks (NN)—applied to NEOCP candidate data collected between 2019 and 2024. Using observations from 2019–2023 for training and 2024 for validation, we achieve consistent NEO classification accuracies of 91%–92%, with negligible variation across models.

We advocate for integrating digest2-based feature sets with ML methodologies to improve candidate selection on the NEOCP. This approach promises not only to reduce the burden of false positives on follow-up networks but also to enhance the overall efficiency and completeness of NEO detection efforts.