OpenLiDARViewer v0.4.8

A focused classification release: scans that ship without a usable classification can now derive one, colour helps separate trees from structures, and the result tells you how much to trust it. Still browser-native, local-first, nothing uploaded.

Highlights

• Classify unclassified scans that carried an empty class channel — a file whose every point is ASPRS 0 ("Created, never classified") or 1 ("Unclassified") is functionally unclassified, even though it technically carries a classification field. Those scans — raw photogrammetry exports, many drone clouds — are now eligible for Classify (derive), where before the action treated the empty channel as "already classified" and declined. Clouds with no classification at all still classify as before.
• RGB-assisted vegetation — when a scan carries colour, Classify now folds a vegetation-greenness cue (normalised Excess-Green) into the geometry: a tall, locally-smooth patch that reads strongly green is treated as canopy rather than a building. This is the cue that matters on photogrammetry, where the surface is noisy and shape alone confuses tree crowns with roofs. Colour only breaks the tree-vs-structure tie; it never invents vegetation on the ground, and a scan without colour classifies exactly as before.
• Confidence on the derived classification — every run now reports an overall confidence and per-class confidence, surfaced in the Classes legend caption and the result toast, so a derived classification reads as "trust this a lot" or "rough visual aid" rather than landing unqualified. Confidence drops on sparse, void-ridden, or coarse-grid scans — exactly when a heuristic is least reliable.
• Void-honest heights — the classifier fabricates a bare-earth surface inside data gaps to keep the math finite, but it no longer pretends that surface is real: a tall point whose height rests mostly on hole-filled void is left Unclassified rather than guessed, and the run raises a plain warning naming the gap. You never get a confident class invented over empty space.

Trust you can act on

• Dataset Story — a command-palette action ("Dataset Story") opens one compact card that answers what the scan is, how good it is, the single biggest limiter, what it's best for, what to use with caution, what isn't established, and the one next step — synthesised from signals the app already computes, so nothing is fabricated and nothing overclaims.
• Export Health check — before a hand-off, the "Export health check" action summarises exactly what is about to leave the app (scan scope, source-vs-derived classification with its confidence, CRS, vertical datum, density, terrain-product readiness) and gives a single ready / caution / blocked verdict with an actionable "before you hand this off" list.
• Shared sessions remember your class filter — when you hide classes in the Classes panel (say, "ground only") and export an .olvsession, the saved file now carries that filter, and re-importing restores exactly the classes you left visible. The recipe stays additive: older sessions load unchanged, and a session with no filter behaves as before.

Honesty

• A scan that already carries a real producer classification (any ASPRS class ≥ 2) is left untouched — Classify never overwrites a surveyor's classes. The derived result stays tagged "derived (heuristic)" with its method and a "validate before relying on it" caveat, and the provenance line now records when the RGB cue was used.

Credits

• Data sources & credits. A new credits page — reachable from a Credits link in the header, with a mirror at docs/credits.md — names the providers behind the public sample datasets (USGS 3DEP and the swisstopo / GURS / AHN national programmes) and thanks the open-source projects the viewer is built on: three.js, loaders.gl, proj4, pdf-lib, laz-perf, and the COPC / EPT formats. The README gained an Acknowledgements section, and THIRD_PARTY_NOTICES.md is now complete.

Fixes

• Export units. The Studio raster exports (contour, depth, height) and the shared Width / Depth / Height / Density rows now label values in the scan's real linear unit — feet for foot-based CRSs, metres otherwise — instead of a hardcoded "m". Coordinates are stored in native units, so this corrects feet shown as metres on US-survey-foot scans.

Under the hood

• The classifier gained a "classify the gaps" capability — derive only the unclassified points of a partially-classified cloud while preserving every producer class — with full unit coverage. It is not yet wired to a user action (so producer-classified scans are never silently modified); a deliberate "fill unclassified points" surface will follow.
• npm run test:release runs the full browser-independent gate in one command: typecheck, the main-deferral lint, the production build, the bundle-budget check, the bucket partition check, and all four test buckets.
• Open-format and analysis foundations landed as pure, tested cores ahead of their interactive surfaces: a measurement → GeoJSON/CSV serializer, a two-epoch DTM change-detection engine (with Level-of-Detection thresholding and co-registration honesty flags), an axis-aligned clipping-box region core, and a breadth-first full-cloud sampling planner. Each is unit-tested in isolation; the GPU/worker layers that drive them from the UI follow, so nothing ships half-wired.

Derived classification is a geometry-and-colour heuristic for visual inspection and research, not a survey-grade or producer classification. Treat any output as survey-grade only after validating it against survey-grade data and procedures.

See CHANGELOG.md for the full list.

Deploy

Static files — host anywhere (GitHub Pages, Netlify, any CDN). The deploy zip extracts with web-safe permissions (644 files / 755 directories) and carries index.html plus assets/ at the archive root.