Rotation detection

A storm cell is the wrong unit to send a tornado alert on.

One cell observed on an ordinary afternoon was a 59 km-equivalent object - roughly 3,500 km² - with its rotation confined to the southwest corner. Alerting everyone under its cone that a tornado was possible would have reached hundreds of thousands of people for a hazard occupying a few percent of the footprint. Almost all of them needed a severe thunderstorm alert. A small minority, in the path of the rotation, needed a tornado alert.

Those are two different hazards, two different populations and two different cones. So a rotation - a velocity couplet - is tracked as its own object, with its own identity, its own motion, its own lifecycle and its own cone.

Why a cell is the wrong unit

Cells grow abruptly. A cell’s area can double, or occasionally quintuple, between one two-minute scan and the next, so every couple of minutes somewhere in the country a cone’s footprint and the population under it multiply. That is survivable for a severe alert. It is not acceptable for a tornado alert, because the thing that grew is the storm outline and not the rotation.

One rotation per cell is not one rotation. Take the single strongest rotation inside each polygon and follow it, and it appears to jump up to 62 km in two minutes. Nothing moves like that: it is the strongest of several rotations, flipping between them. One storm was observed holding five distinct rotations at once, three of them over the alert threshold.

How a rotation is found

Azimuthal shear is read off a 500 m grid, at the 0-2 km layer, every two minutes. Inside each cell’s polygon:

1. Discard everything below the detection floor.
2. Separate peaks at 8 km, so one rotation is not split into several and two genuine neighbors are not merged into one.
3. Cap at five per cell. A long squall line genuinely holds several.

Each surviving peak is a candidate, and nothing alerts on the scan it is found.

Tracking one

A rotation gets the same treatment a cell does - a motion fitted from its own recent positions, a limit on how far it can plausibly jump, and a coasting window before it is declared over - with every window tuned tighter, because a rotation is a far smaller thing than a storm.

It can afford to be tighter: the parent cell’s motion is already known, so the rotation is predicted forward and matched within a few kilometers rather than searched for blind.

Below 5 km/h a rotation reports no motion at all. Positions are quantized onto the 500 m grid, so a slow rotation can sit in the same grid cell for several scans and produce an exact zero slope, which would otherwise draw a cone confidently claiming it is going nowhere.

A rotation is linked to a cell, not owned by it

A couplet never inherits its identity from its parent. If it did, it would take on every merge and split failure that stitching exists to paper over. A rotation can outlive its parent, change parents when cells merge or split, and several can share one.

What has to be true before it alerts

emerging → confirmed → [alerting] → coasting → ended

Persistence is the gate that matters. Shear is noisy, and a single-scan maximum is not a rotation: a feature has to be seen in three consecutive scans - six minutes - before it can alert at all, and then hold above the alert threshold for two scans on top of that. This is the tornado-side equivalent of the cells’ hysteresis and it is the single biggest lever on tornado false alarms. Without it, one grid cell flickering across the threshold latches a tornado glyph on a storm that is falling apart.

Clearing is slower than firing, and a rotation that the radar cannot see well enough to trust does not alert at all. Radar quality is checked first, and where it is too poor, the system says so rather than reporting nothing found.

The environment has to support it

Rotation is necessary for a tornado and nowhere near sufficient. A rotation may not alert unless its parent cell has itself reached the Elevated band, which is the check that catches real, measured rotation sitting in an environment that will not produce anything. It is the biggest single lever on false alarms here, and it is not a rotation measurement at all.

Rotation with nothing above it is vetoed

The 3-6 km shear grid is sampled above every low-level detection. Strong shallow rotation with no rotation above it is a gust-front kink, not a mesocyclone, and it is blocked: the strongest non-stacked detections measured sat in the top few percent of all shear while carrying tornado probabilities of 0 and 1.

Depth is used as a veto and never as a score. A mid-level grid that simply failed to download promotes rather than blocks, because a missing file is not evidence against a rotation.

The rotation’s own cone

A rotation’s cone is fitted from rotation tracks rather than storm tracks, and it behaves differently in two ways a reader can see.

It is far tighter at short lead - at ten minutes it covers under half the area of the parent cell’s cone - because the thing being projected is a few kilometers across rather than 25.

It has a rear margin, which the cell cone has no term for at all. A storm outline is carried along and does not reappear behind itself. A rotation dissipates and re-forms, and when a cycling pair hands the fit a bearing pointing at the member that just died, about a third of the misses land behind the origin.

Why it stops at 25 minutes

At equal capture, a rotation cone against its parent cell’s cone:

Past about 25 minutes the rotation’s cone stops being the smaller of the two, and past that it is a bigger audience for a narrower hazard. So tornado alerts are a short-lead product at 10 to 20 minutes, and severe alerts are the long-lead one at 30 to 60.

What a located rotation is worth

It is a precision instrument, not an early one. On two outbreak days the rotation tier fired within a minute or so of the NWS tornado warning, where the cell tier ran 45 to 54 minutes ahead of it:

The rotation fires at about the same time the forecaster does. What it adds is where, and that is enough: it is the difference between telling a whole storm footprint and telling the few kilometers actually in the path.