MCS Morphology AI-based classification of mesoscale convective system structure
saved

Classifying the morphology of mesoscale convective systems

Mesoscale convective systems (MCSs) organize into distinct structures, and those structures matter: linear MCSs are disproportionately associated with damaging winds and other severe-weather hazards. This project builds an AI tool that classifies linear MCS morphology directly from radar reflectivity imagery, so that storm structure can be linked to hazardous impact across large tracking datasets.

Objective

Train and evaluate a model that assigns one of six morphology categories to each radar snapshot of a tracked MCS. Because the model consumes only tracked-MCS reflectivity fields, it transfers to any tracking dataset that supplies them.

Why manual labels

Supervised training needs a reference set. Trained student labelers review each image, assign a category, and record how confident they are. Consistency across labelers matters more than resolving every complicated feature.

This platform

A browser-based labeling tool. Images are paged through with the keyboard, labels save as you go, motion metadata is shown alongside each frame, and results export to CSV in the same layout as the project spreadsheets.

Dataset

2,525tracked MCSs (2004–2021)
43,655hourly radar snapshots
11,835labeled so far (2004, 2005)
6morphology categories

Each snapshot is a two-panel figure: the SL3D classification mask (convective / stratiform / anvil) beside composite reflectivity, restricted to the tracked system. Per-frame system motion, speed, and precipitation-feature shape come from the accompanying track files.

Demo bundle: this deployment ships 5 tracks (86 frames) from 2004 so the workflow can be exercised end to end. Point the tool at a local copy of the full archive from the Label tab to work at scale.

The six categories

Labeling guidance

For every image decide three things: whether the convective region is linear, which linear type it is, and whether the image is too ambiguous to call.

Procedure

  1. Find the convective region. Work from the areas of relatively high reflectivity — the strongest convection. Do not classify from the shape of the broad weak-reflectivity or stratiform area.
  2. Decide whether the convection is linear. Do the cells form a recognizable, continuous or nearly continuous line?
  3. Locate the stratiform region relative to the convective line and to the direction of motion.
  4. Pick the single best label for the dominant structure. Small irregularities and isolated cells do not outweigh the overall organization.

Decision sequence

Does the convective region form a clear line?
No → Nonlinear
Yes → where is the dominant stratiform region?
  • Behind the line → Trailing Stratiform
  • Ahead of the line → Leading Stratiform
  • Alongside, parallel to the line → Parallel Stratiform
  • Downstream of persistent upstream convection → Back-Building
  • Line adjoins / extends from the main stratiform mass → TL/AS
  • Still unclear → Uncertain

Category definitions

Rules that decide the hard cases

Motion defines leading and trailing.

“Leading”, “trailing”, “upstream” and “downstream” are relative to where the system is moving — never to left/right on the screen. A trailing stratiform region can appear north, south, east or west of the line. The tool prints the motion vector for every frame; use it.

Classify only the tracked MCS.

The panels are masked to the tracked system. Nearby storms should not change the label unless they are clearly part of the tracked feature.

Label the dominant structure.

Real MCSs are imperfect. Do not switch labels because of a few isolated cells, a small gap in the line, a minor bend, or a detached shower.

Do not infer beyond the image.

Use the radar structure and the motion information shown. Not the location, the season, or what is climatologically common there.

Each frame stands alone.

Morphology evolves over an MCS life cycle. Label each time step for what it shows; labels need not be constant along a track.

When to flag Uncertain.

Two categories equally plausible · line weak or poorly defined · motion needed but undeterminable · system in transition · MCS partly outside the frame · structure obscured · missing or poor radar data. Do not force a label.

Before you submit

  • You judged the convective region, not the total precipitation shape.
  • You considered the direction of motion.
  • You classified only the tracked MCS.
  • You chose the dominant morphology.
  • You used Uncertain when the image did not support a confident call.
—
radar frame
N
—
Pick a track on the left to begin.
confidence

Model development progress

Where the classifier stands, what was actually run, and the numbers it produced. Each stage opens into the technical settings behind its result so the work can be reviewed rather than taken on trust.

Labeling progress

Export

Two layouts are produced. The long file is one row per labeled image and carries confidence, flags and notes. The wide file is the track × frame matrix used by the existing project spreadsheets, so it drops straight into the current workflow.

Long format columns

labeler, year, track, seq, filename, valid_time, label_code, label_key,
label_name, uncertain, confidence, note, labeled_at

Wide format

Storm Track, 1, 2, 3, …    values: 0=NL 1=TS 2=PS 3=LS 4=TL/AS 5=BB, U=uncertain, blank=unlabeled

Resume / merge

Load a session JSON or a long CSV exported earlier to continue where you left off. Existing labels for the same frame are overwritten by the imported file.