Closing the Loop on Contrail Avoidance with Satellite Verification

Contrails---the thin ice clouds that aircraft leave behind---cause a large share of aviation's warming, and rerouting the few flights that produce them could avoid much of it. But an avoided contrail only counts if a satellite can confirm it never formed, and that check is non-trivial: contrails are one to two pixels wide, cover 0.18\% of pixels, and closely resemble natural cirrus. We build a small diffusion model (8.4M parameters, one GPU) that detects them, and run a controlled study of effective components. It reaches 0.476 PR-AUC against 0.414 for a DeepLabV3+ baseline and $0.119$ for an adapted MedSegDiff; doubling the CNN's input resolution reaches parity (0.499, p=0.07). Three lessons travel beyond contrails. Check input resolution before you build an architecture. Simple flips and rotations more than double accuracy, outweighing every architectural choice we measured. And pretraining the model on contrail shapes is detrimental: it learns that thin strokes are everywhere and paints them onto empty scenes, collapsing precision to 1\% while recall metrics still score the degraded model as excellent---a failure no threshold or guidance heuristic repairs. Code available at: https://anonymous.4open.science/r/contrail_detection-362B