planetary-defense-neocp-vetting

** Now has 2 other capabilities. Go to https://spacesift.vercel.app to see **

Machine-learned vetting for the Minor Planet Center's NEO Confirmation Page: for every candidate awaiting confirmation, predict the probability it's a real near-Earth object vs. a false positive (artifact, satellite, star, or main-belt asteroid), so the volunteer follow-up network stops spending scarce telescope time on junk. Only ~50% of NEOCP candidates are real NEOs today; in the Rubin/LSST era that drops to ~8%.

Benchmark to beat (A&A 2025): exclude >80% of non-NEOs while losing ≤1% of real NEOs. Current status and honest numbers: see PIPELINE.md.

How it works

1. Capture (src/capture_neocp.py, runs 2×/day): snapshots the live NEOCP feed, saves each new object's astrometric tracklet while it is still public (MPC deletes them on removal — the dataset cannot be backfilled), and records outcomes from the Previous-NEOCP page. data/capture/ is the accumulating labeled archive.
2. Features (src/features.py): the 26 digest2 sub-scores (13 orbit classes × raw/noid) + tracklet geometry the published benchmark ignored (sky-motion rate, ecliptic latitude, solar elongation, observatory code…).
3. Train/evaluate (src/train.py --with-backfill): LightGBM, time-based split, metrics reported at 99% NEO recall. Includes a guard against provenance leakage between captured and backfilled rows.
4. Score & publish (src/predict.py, src/make_dashboard.py): ranked outputs/daily_vetting.csv + a live dashboard (outputs/dashboard.html) with an extrapolated-position sky map.

Plus an image track (src/fetch_css_bulk.py, src/motion_baseline.py, src/train_cnn.py): real-vs-artifact classification on Catalina Sky Survey 4-frame detection sequences from Zooniverse's Daily Minor Planet project.

Setup

pip install numpy scikit-learn lightgbm pillow          # + torch for the CNN
git clone https://github.com/Smithsonian/digest2        # see PIPELINE.md for build notes

Full runbook, environment gotchas, and current results: PIPELINE.md. Original research plan: Planetary Defense v1 — Project Plan.md.

Data credits: Minor Planet Center (NEOCP, MPECs, observations API), Catalina Sky Survey / Zooniverse Daily Minor Planet. Please cite the MPC when using their data and poll politely.

** Now has 2 other capabilities. Go to https://spacesift.vercel.app to see **