SVRCLIPNIU Meteorology

Storm morphology

Labels & machine learning.

A clearer home for morphology definitions, labeled subsets, baseline workflows, and known limitations.

Primary classes

Seven report-associated storm-mode classes.

Choose one primary mode at the report valid time; use the sequence as context rather than labeling the entire two-hour evolution.

0

Discrete cellular

Separated convective cells without a continuous convective line.

View web-consensus events →
1

Cellular cluster

Multiple interacting or closely spaced cells without dominant linear organization.

View web-consensus events →
2

Linear / QLCS

A continuous or nearly continuous convective line, including broken or bowing configurations.

View web-consensus events →
3

Embedded / hybrid

A dominant cell embedded within a line or cluster, or a genuine cellular-linear transition.

View web-consensus events →
4

Tropical cyclone

Convection or rainbands associated with a tropical cyclone or remnant circulation.

View web-consensus events →
5

Other / ambiguous

Usable radar data whose report-associated morphology does not clearly fit another class.

View web-consensus events →
6

No associated echo

Usable radar data without a plausible convective echo associated with the report.

View web-consensus events →

Required quality

Keep usability separate

Choose one: Good, Partial sequence, Missing valid time, Artifact / noise, Domain edge, or Unusable other. Quality labels are not storm-mode consensus votes.

Optional tags

Describe context without changing mode

Meteorological/context and subjective/reviewer tags are nonexclusive. “In image” tags describe anything visible anywhere in the 136×136 sequence, not necessarily the echo associated with the report.

Meteorological/context tags

Nonexclusive physical context

  • bowing_segment — Bowing segment
  • embedded_cell — Embedded cell
  • merger — Merger
  • cell_split — Cell split
  • convective_initiation — Convective initiation
  • upscale_growth — Upscale growth
  • transition_to_linear — Transition to linear
  • transition_to_cellular — Transition to cellular
  • rapid_intensification — Rapid intensification
  • rapid_weakening — Rapid weakening
  • back_building — Back-building
  • stationary — Stationary
  • mcs_in_image — MCS in image
  • qlcs_in_image — QLCS in image
  • discrete_cells_in_image — Discrete cells in image
  • cellular_cluster_in_image — Cellular cluster in image
  • tropical_in_image — Tropical in image
  • embedded_hybrid_in_image — Embedded / hybrid in image

Subjective/reviewer tags

Nonexclusive review cues

  • messy — Messy
  • weird — Weird
  • interesting — Interesting
  • what_is_happening — What is happening?
  • take_a_look — Take a look
  • cool — Cool
  • wow — Wow
  • research_worthy — Research worthy
  • case_study — Case study
  • low_confidence — Low confidence
  • i_need_an_adult — I need an adult

Original split

Time-separated benchmark

Training1996–2011
Validation2012–2013
Testing2014–2017

Recommended use

Keep provenance explicit

For new benchmark work, preserve whether each label is human-generated, consensus-derived, or model-generated. Avoid mixing inferred labels into a nominally human-labeled test set.

Python examples ↗