Skip to main content

Point-cloud classification

The point-cloud classification model takes a raw LAS or LAZ file and labels every point with a semantic class: ground, vegetation, building, water, noise. Classified points drive volumetric measurement, building extraction for the twin viewer, vegetation-encroachment alerts, and ground-versus-surface change detection.

Two models are available depending on the input.

Rule-based classifier (Live)​

The default classifier uses geometric features (relative height, planarity, point density) plus per-point neighbourhood statistics. No ML; deterministic and reproducible.

  • Best for clean LiDAR with consistent point density.
  • Outputs the ASPRS standard classes (1: unclassified, 2: ground, 3: low veg, 4: medium veg, 5: high veg, 6: building, 9: water).
  • Fast: ~30 to 60 seconds per million points on CPU.

RandLA-Net outdoor_v19 (Preview tier)​

A neural model trained on UK DEFRA LiDAR. Tier: Public corpus (Preview).

  • Better than the rule-based classifier on edge cases (sparse buildings, low vegetation overlapping ground, urban clutter).
  • Current accuracy: mIoU 0.296 on a held-out test set. We are honest about the preview-tier status; expect a v2 once enough customer-corrected data accumulates.
  • Runs on the local AI worker (CPU); ~30 to 60 seconds per million points.

Run a classification​

  1. Open the LiDAR survey.
  2. Click the Classification tab.
  3. Pick a model: Rule-based or RandLA-Net (preview).
  4. The estimated credit cost appears (1 credit per ~2 million points). Confirm.

The job runs asynchronously. When complete, the survey viewer colour-codes points by class. The classified LAS is also saved back to the survey for download or re-use.

Visualise classified points​

In the survey viewer:

  • Toggle classes: use the class legend to show / hide specific classes.
  • Filter by classification confidence: only available on the neural model.
  • Export classified LAS: click the download button on the survey toolbar.

Use classified points downstream​

Once classified, ground and building points feed three downstream features:

  • Change detection: the ground class diffs between two surveys produce a DTM-delta raster.
  • Volume calculation: stockpile or earthworks volumes from selected polygons.

Provide feedback on misclassifications​

If RandLA-Net mislabels points, you can correct them and the corrections are logged for fine-tuning:

  1. In the viewer, select a region with a misclassification (lasso or polygon select).
  2. Click Reclassify selection.
  3. Pick the correct class.
  4. Confirm.

Reclassification feedback drives the next model version. After enough corrections accumulate in your tenant, your organisation gets a fine-tuned model that improves on the public-corpus baseline.

Limitations​

  • The RandLA-Net model was trained on UK DEFRA LiDAR (1 m point spacing, mostly rural / mixed terrain). Dense urban LiDAR or different sensor geometries may classify less well.
  • Water classification depends on having returns from the water surface; many LiDAR sensors get no return from water and the classifier cannot recover what is not there.
  • Neither model classifies fine asset types (pipe vs cable vs sign). Those require a feature-class lookup, not point-cloud classification.

Roadmap​

We are working on raising mIoU from 0.296 to 0.50+ with a height-above-ground feature engineered from the ground class itself, and on a outdoor_v2_prod model trained on a mixed corpus. Per-customer fine-tunes follow once enough reclassification feedback lands.

What next?​