Classification guide

How to use the Classify tool.

Build a queue, inspect the valid-time radar image and its evolution, assign the two required labels, add useful tags, and save your work.

Quick start

A complete classification in six steps.

Primary storm mode and data quality are required. Meteorological/context and subjective/reviewer tags are optional and nonexclusive.

1

Enter your classifier ID

Use consistent initials or an assigned identifier. SVRCLIP appends a browser-session suffix so two people entering the same text do not overwrite one another.

2

Define the queue

Optionally filter by event ID, hazard, year, catalog morphology, state, or maximum reflectivity. Leave filters blank to use every event with a mapped preview.

3

Choose queue order

Sequential order is the default. Enable Randomize queue order to create a stable shuffle for that run. Clicking Start / apply again creates a new shuffle.

4

Inspect the radar

Classify the storm associated with the report at the valid time. Use the −1 hour to +1 hour animation as context for organization, evolution, artifacts, and timing.

5

Assign labels and tags

Select one primary storm mode and one data-quality label. Add any meteorological/context or subjective/reviewer tags that materially describe the case.

6

Save and continue

Save & next writes the selection and advances. Skip and Next move without saving. Previous returns to the prior queue item.

Primary storm mode

Label the report-associated echo.

Choose the morphology most directly associated with the severe report at valid time—not every echo visible within the 136 × 136 image.

  1. 1 Discrete cellular
  2. 2 Cellular cluster
  3. 3 Linear / QLCS
  4. 4 Embedded / hybrid
  5. 5 Tropical cyclone
  6. 6 Other / ambiguous
  7. 7 No associated echo

Read the full class definitions →

Data quality

Record usability separately.

A difficult or unusual morphology is not automatically poor quality. Use quality to describe whether the radar evidence can support a classification.

  • Good
  • Partial sequence
  • Missing valid time
  • Artifact / noise
  • Domain edge
  • Unusable other
Missing valid time: use this when the report-time frame is unavailable. Partial sequence means the valid frame is usable but one or more surrounding frames are missing.

Optional annotations

Add context without changing the primary class.

Tags are searchable in Explore. Multiple tags may be applied to the same event.

Meteorological/context tags

Use these for physical structures or evolution visible in the still image or sequence. “In image” tags may describe features elsewhere in the domain, even when they are not the report-associated echo.

Bowing segmentEmbedded cellMergerCell splitConvective initiationUpscale growthTransition to linearTransition to cellularRapid intensificationRapid weakeningBack-buildingStationaryMCS in imageQLCS in imageDiscrete cells in imageCellular cluster in imageTropical in imageEmbedded / hybrid in image

Subjective/reviewer tags

Use these as review cues for unusual, uncertain, or especially useful examples. They are intentionally separate from meteorological labels and quality control.

MessyWeirdInterestingWhat is happening?Take a lookCoolWowResearch worthyCase studyLow confidenceI need an adult

Navigation

Move efficiently.

KeyAction
1–7Select a primary storm mode
SSkip the current event without saving
← / →Move to the previous or next event

Keyboard shortcuts are ignored while an input or menu has focus.

Review and export

Revise without losing provenance.

  • Returning to an event loads your saved label and tags.
  • Save again to replace your own classification for that event.
  • Clear my label removes your saved classification and its tags.
  • Export my CSV downloads all classifications stored under your effective classifier ID.
Avoid anchoring bias: the catalog label stays hidden until you deliberately reveal it. Classify independently before viewing it.
Random queues: the shuffle is stable while you move through the current queue. Reapplying the filters starts a new randomized run, so previously skipped or labeled events may appear at a different position.