Counter-UAS Track Triage: Drone or Bird?

Preview mode. Browse milestones and edit starter code here. Sign in to run and submit.

0/8EDA

Python

intermediate~7h8 milestones

The problem

A perimeter air-defense site fuses a short-range radar with a slewable EO/IR camera. Every second it holds hundreds of tracks, and about 92% of them are birds, crewed aircraft, road vehicles, wind turbines and weather. One operator watches the screen. You are building the model that decides which tracks get the operator's attention. Miss a small drone and it reaches the perimeter unobserved. Cue the operator on every gull and they stop looking. The dataset is synthetic and physically motivated: radar cross-section, SNR that falls with range, rotor blade flash (micro-Doppler) that fades into the noise floor, and a camera that only reaches about 5 km and only looks at some tracks. Your job is to engineer kinematics from raw radar plots, fuse the sensors, and defend an operating point in operator minutes.

Your task · Read the Air Picture

Establish how rare drones are, what else is in the sky, and how often the camera looked.

Implement explore_tracks(train_df) returning a dict with:

row_count: number of tracks in the training set
uas_count: number of drone tracks (is_uas == 1)
uas_rate: fraction of tracks that are drones, rounded to 6 places
class_counts: dict mapping each track_class value to its row count
eo_missing_rate: fraction of tracks where eo_confidence is NaN, rounded to 6 places
uas_rate_eo_tasked: drone rate where eo_confidence is present, rounded to 6 places
uas_rate_eo_missing: drone rate where eo_confidence is NaN, rounded to 6 places
median_micro_doppler_by_class: dict mapping each track_class to its median micro_doppler_hz, rounded to 6 places

track_class exists only in the training data. It tells you what each track really was: bird, quadcopter, fixed_wing_uas, crewed_aircraft or clutter. is_uas is 1 for quadcopter and fixed_wing_uas.

Look hard at the last three numbers. The camera looked at fewer than 3 tracks in 10, and the drone rate among tracks it looked at is several times the rate among tracks it did not. That is not a coincidence: the site's cueing logic prefers tracks the radar already finds suspicious. A missing camera score is evidence, so you must not impute it away. Then notice that the median micro-Doppler of a bird is not zero. Rotor blade flash is the best radar clue you have, and it sits on top of a noise floor.

Constraints

▪About 8% of tracks are drones (quadcopters and fixed-wing), so accuracy is useless
▪Birds and quadcopters overlap on speed, altitude, turn rate and radar cross-section
▪The file has raw plots (x0, y0, x1, y1, x2, y2 at 1 s spacing), not a speed or turn rate
▪eo_confidence is missing when the camera was never slewed onto the track, which happens for about 72% of tracks
▪The operator can look at 12% of tracks
▪A drone detected in time is worth 1,000 units; a false cue costs 25 units of operator time
▪Hidden test set is scored server-side, so you cannot tune against it

Graded on

Track statistics and finite-difference kinematics match the reference to 1e-4
PR-AUC >= 0.62 and ROC-AUC >= 0.90 on the hidden test set
Recall >= 0.65 while cueing the operator on no more than 12% of tracks
Net benefit >= 950,000 with cues capped at 50% of tracks
Batch scoring, training included, completes within 3 seconds with PR-AUC >= 0.60

Evaluated server-side against a hidden test set.

Skills you'll demonstrate

Sensor FusionFeature EngineeringClass ImbalancePrecision-Recall CurveThreshold SelectionCost-Sensitive LearningInformative MissingnessGradient Boosting