Scientific ML: Contrail segmentation

The agent gets nine-channel infrared satellite frames with expert-labeled contrail masks and builds a model that predicts a binary contrail mask per frame. Training runs offline; predictions go through a sealed leaderboard gateway.

Persistent contrails are a meaningful share of aviation's warming effect, and mitigating them, by rerouting the small fraction of flights that produce them, first requires seeing them. In satellite imagery a contrail is a thin, faint, linear cloud that looks a great deal like every other thin, faint, linear cloud, which is what makes detection a real modeling problem rather than a thresholding exercise.

The metric is a single global Dice coefficient over the entire evaluation split, so performance on rare positives matters as much as bulk accuracy. Scored submissions are quota-limited and return public scores only; the reward is the floor-normalized private score of the best-public submission.

The graded runs converged on the same recipe independently: a U-Net with a ResNet-34 encoder on a three-channel ash false-color composite of the infrared bands. Each trained at increasing resolution and finished with test-time augmentation and a fold ensemble. What separated the top of the band from the bottom was execution rather than ideas. The strongest run bought its edge with longer training at full resolution and a five-fold ensemble. The weakest ran the same ladder from a weaker base model. It spent its final submissions re-weighting the ensemble for gains in the third decimal.

The discards rhyme too. Every post-processing idea any run tried failed to improve local validation and was dropped: minimum-area filtering, probability smoothing, out-of-fold calibration. The one run that widened the input to all nine channels measured it worse than the composite and reverted.