Pulling unlabeled frames from the warn archive…
Every available frame is labeled. Check back after the next convective day — the archive grows with every warning TIMA scans.
Can't reach the labeler API. Check that tima-api is up and this page is served from the same origin (or pass ?api=…).
Panel 1 — Reflectivity (dBZ): the storm's precipitation structure. Warm colors = heavy precip. Look for shape: a hook or pendant appendage on the storm's flank is the classic tornadic footprint; a long diagonal line is a squall; blotchy greens are ordinary rain.
Panels 2 & 3 — Rotation (azimuthal shear): red = cyclonic spin, blue = anticyclonic, white = none. Panel 2 is near the ground (0–2 km); Panel 3 is mid-storm (3–6 km). The deeper the red toward the top of the scale, the stronger the spin. The + marks the point being scored — judge what's at and around it.
A tight, intense red couplet in the 0–2 km panel at/near the +, usually echoed in the 3–6 km panel (deep rotation), and ideally co-located with a hook/appendage or notch in reflectivity. Compact and strong near the ground is the tell — tornadoes live in Panel 2.
Clear rotation, but not the tornado picture: spin that's broad or smeared rather than tight; strong in the 3–6 km panel but weak near the ground (elevated mesocyclone); or a decent couplet with no supporting reflectivity structure. This is the model's hardest case — your call here is the most valuable label in the whole system.
Little or no meaningful echo, or plain rain with no organized rotation — shear panels mostly white/speckle with no focused red maximum near the +.
Genuinely ambiguous, broken image, or the + sits at the edge of the action. Skipping is always correct when unsure — skipped frames recirculate, and an honest skip beats a guessed label every time.
Your handle is attached to every label you make — it's how your calls get credited (and how we build the leaderboard later).