VL3D
Geo-Forest LiDAR MVP
Semantic mapping · external validation · forest intelligence
Commercial product MVP

Turn airborne LiDAR into operational land-cover and forest intelligence.

A reproducible pipeline for point-wise semantic segmentation, GIS-ready outputs and forest analytics, with strong internal performance and demonstrated improvement on a separate geographic holdout.

86.49% best internal mIoU-run OA72.63% best internal mIoU+10.35 pp external mIoU vs baselinePilot-ready workflow
1

Airborne LiDAR

XYZ, RGB/CIR/NIR, intensity and reference classifications.

2

Label-free local geometry

Density, height and surface context at several spatial scales.

3

Robust segmentation

Supervised training with robust normalization and class balancing.

4

GIS and forest outputs

Maps, LAZ, grids, indicators, metrics and review priorities.

The business problem

Manual LiDAR interpretation does not scale.

Large surveys contain millions or billions of points. Teams need consistent maps and a prioritized review process, not another raw data archive.

Slow manual mapping

Reviewing point clouds by hand consumes specialist time and delays projects.

Difficult boundaries

Low vegetation, medium vegetation, roofs and tree crowns require local 3D context.

Underused data

Organizations often own LiDAR but lack a repeatable pipeline that converts it into operational layers.

How the product works

A practical 3D workflow from raw points to decisions.

Raw tiles

Airborne LiDAR, color, infrared and intensity.

Prepared blocks

Clean, aligned and reproducible train/validation/test data.

Local context

Label-free geometry at 2.5 m, 5 m and 10 m.

Point model

Robust normalization, focal loss and balanced training.

GIS delivery

Semantic layers, forest indicators and technical reports.

The deployable core combines supervised segmentation with physically meaningful features calculated without labels.
Model selection

Two reference models serve different objectives.

Internal mIoU leader

Local-context model

Highest internal macro-F1 and mean IoU in the current comparison.

86.49%OA
82.60%Macro-F1
72.63%mIoU
Deployment candidate

Robust local-context model

20k balanced blocks and robust normalization for stronger geographic transfer.

86.71%Internal OA
82.48%Internal macro-F1
72.50%Internal mIoU
The robust candidate keeps essentially the same internal quality while performing substantially better on geographically separate data.
Internal benchmark

Strong gains across all six semantic classes.

MetricOriginal baselineBest internal mIoU runDelta
Overall accuracy83.03%86.49%+3.45 pp
Macro-F174.82%82.60%+7.78 pp
Mean IoU63.49%72.63%+9.14 pp

What changed

The model receives local height, density, color and surface context rather than interpreting each point in isolation.

6/6 IoUs improvedFull test splitLabel-free feature cache
Robust internal class performance

Strong terrain, trees, buildings and water separation.

Ground
76.49%
Low vegetation
43.91%
Medium vegetation
52.71%
High vegetation
94.53%
Building
69.51%
Water
97.86%
Low vegetation remains the hardest internal class and is a priority for future backbone and data improvements.
Why robust normalization matters

The main deployment challenge was domain shift.

Spectral shift

Mean NIR changed from 0.5050 internally to 0.2955 on the external development set.

Terrain shift

Mean block height range changed from 16.99 to 61.58, with a much larger external tail.

Effective correction

Robust coordinate/spectral normalization plus label-free geometry and 20k balanced blocks.

The solution preserved strong internal performance and materially improved external geographic generalization.
External development benchmark

Robustness improved on a separate 32-tile campaign.

ModelOAMacro-F1mIoU
Strong TW baseline72.60%63.99%49.45%
Robust local-context80.83%73.42%60.41%
+10.96 pp

Mean IoU over the strong baseline

Overall accuracy improved by 8.23 pp and macro-F1 by 9.43 pp, with 6/6 class IoUs improved.

This set was later used to diagnose domain shift and develop the robust-normalization strategy, so it is reported as an external development benchmark.
Fresh disjoint external holdout

Improvement carries to unseen, non-overlapping tiles.

ModelOAMacro-F1mIoUmIoU excl. water
Strong TW baseline77.33%51.38%40.62%48.33%
Robust local-context83.27%61.80%50.98%60.62%
+10.35 pp

External mean IoU

OA improved by 5.94 pp and macro-F1 by 10.42 pp. All six class IoUs improved against the strong baseline.

Water represents only 0.064% of reliable points and buildings 0.80%; another representative holdout is required for strong absolute external claims in those classes.
Fresh-holdout class gains

The largest transfer gain appears in buildings.

ClassStrong baseline IoURobust IoUDelta
Ground73.43%79.97%+6.54 pp
Low vegetation21.23%27.70%+6.47 pp
Medium vegetation47.28%56.46%+9.18 pp
High vegetation81.93%88.47%+6.54 pp
Building17.80%50.50%+32.70 pp
Water*2.09%2.77%+0.68 pp
Forest intelligence layer

Segmentation becomes practical forest information.

Vegetation ratios

Low, medium, high and combined vegetation by tile or grid cell.

Canopy cover proxy

High and medium-high vegetation ratios from classified points.

Height proxy

Vegetation z95 relative to local ground, or a conservative fallback.

Priority flags

Canopy gaps, sparse cells and unusual vegetation/building patterns.

These layers support prioritization and monitoring workflows and can be validated against customer reference data during a pilot.
What the buyer gets

GIS-native deliverables and a low-risk deployment path.

Operational outputs

Classified LAZ, semantic maps, grids, forest layers and review-priority maps.

Evidence and QA

Global metrics, class metrics, confusion matrices, error analysis and reproducible reports.

Integration

Delivery to QGIS, ArcGIS, PostGIS, GeoServer, web viewers or customer APIs.

1. Discover

Choose one area and one measurable operational problem.

2. Pilot

Process historical data and validate representative zones.

3. Scale

Run tiled inference over a larger territory.

4. Operate

Monitor quality, update models and add future analytics.

Final message

A commercially credible LiDAR intelligence MVP with demonstrated geographic transfer.

The system combines label-free local geometry, robust normalization and supervised learning to produce GIS-ready land-cover and forest intelligence. It preserves strong internal quality and improves a strong baseline on a fresh, tile-disjoint external holdout.

Paid pilot readyExternal validation completedGIS-ready deliverablesScalable platform
Next priorities: a representative water/building holdout, a stronger efficient 3D backbone, calibrated uncertainty and production-scale tiled inference.