Every deliverable PRISM·lidarcloud ships is backed by a validation chain: cross-validation against the industry-reference workflow, surveyed ground control, independent national references, public benchmarks, and bit-level reproducibility. This page shows the methodology and the measured results.
The engine is a self-contained scientific pipeline — no third-party GIS licenses anywhere in the chain. Each stage is versioned, and every run records the exact configuration that produced it.
Before any classification, the cloud is profiled — point density, canopy structure, ground visibility, terrain roughness. Processing parameters adapt to the regime (coastal mangrove behaves nothing like an open ag field), including automatic adaptation across two orders of magnitude in point density, from sparse aerial collections to dense drone surveys.
A multi-stage adaptive ground filter separates bare earth from vegetation, buildings, and noise — engineered specifically to keep the terrain features generic tools smooth away: bank tops, erosion scarps, ditch inverts, grade breaks. Outlier scrubbing is density-aware, so legitimate sparse returns under canopy survive.
Bare-earth, surface, and canopy-height models at 25 cm. The ground surface uses continuity-preserving interpolation that recovers the convex breaks — crest lines, scarp lips — that straight-line interpolation rounds off. Every DEM ships with a per-cell uncertainty raster, so you can see exactly where the model is well-supported and where canopy occlusion forced interpolation.
The delivered cloud carries ground, vegetation height tiers, buildings, and water — assigned with shape- and regime-gated rules tuned to avoid the classic false positives (mangrove canopy is not a building; a dry field is not a lake).
Your surveyed checkpoints produce an accuracy assessment following ASPRS Positional Accuracy Standards (both the 2014 edition and the 2023 second edition conventions). An independent cross-check against the USGS 3DEP national elevation model runs automatically. Deliverables export with FGDC metadata, and the accuracy report is a client-ready PDF.
Calibration spans five full-scale sites — coastal mangrove, dense creek-cut forest, and agricultural research plots, 66 to 342 million points each — processed end-to-end and compared against the industry-reference commercial workflow, surveyed control, and national reference data.
| What was measured | Result | Reference |
|---|---|---|
| Bare-earth DEM agreement, four terrain regimes | within 2–5 cm mean differencemangrove coast, creek-cut forest, two ag sites | industry-reference commercial workflow |
| Coastal erosion-scarp crest elevation | within ~1 cm of surveyed truththe reference workflow rounds the same crest off by ~26 cm | field survey |
| Reproducibility of every raster & cloud | byte-identical (MD5-verified)same input ⇒ same output, independent of machine load or core count | repeated end-to-end runs |
| Agreement with the national elevation model | σ ≈ 0.4–0.5 m on dense-forest terrainsurface-to-surface check dominated by the reference's own uncertainty and collection-epoch differences — not PRISM checkpoint accuracy | USGS 3DEP |
| Public benchmark — forest & rural ground filtering | median ≈5% total error (2–11% per sample)terrain-matched configuration, documented; in the range of the best published filters of its class | ISPRS ground-filter test (independent labels) |
| Public benchmark — noise-contaminated metropolis | 70.2 ground-IoU, untuned — the best non-learned resultthe benchmark authors "carefully chose the fine-tuned parameters of the four filters for each test data, through the trial-and-error strategy" (Qin et al. 2023, ISPRS J. Photogramm. Remote Sens. 202:246–261); each still scored ≤ 39.9 and the commercial reference collapsed to 0.0. PRISM ran production defaults — no tuning, no training. Neural networks trained on the benchmark: 76.6–89.1. | OpenGF Test II (independent labels) |
| Building removal from the bare-earth DEM | all reference buildings correctly removedroof points classified, ground interpolated from true surrounding terrain | surveyed ag site with known structures |
What we don't claim: benchmark scores we haven't earned yet. Our public-benchmark program is ongoing — dense-urban scenes at sparse aerial density are an active engineering focus, and we publish results when they are real. Accuracy on your site, with your checkpoints, is the test that matters — which is why the trial is free.
The engine's development discipline: no change ships unless it is quality-up, or provably neutral — verified on full-scale calibration sites before promotion, every time.
Coastal mangrove, dense creek-cut forest, and agricultural plots — chosen because each terrain breaks a different assumption. Every engine change re-validates against all of them.
Hillshade review from above — and from below. Underside relief exposes classification artifacts that hide in a top-down view. If it isn't clean from both sides, it doesn't ship.
Candidate changes must reproduce the locked production output byte-for-byte wherever they claim no effect. Drift of a single byte fails the gate.
The per-cell uncertainty raster is calibrated against surveyed checkpoints, so "how good is the DEM here?" gets a numeric answer — not just a color ramp.
Every processed site can be compared against the USGS national elevation model in one click — an outside reference we don't control.
Accuracy statements follow ASPRS Positional Accuracy Standards (Ed. 1 2014 and Ed. 2 2023 conventions), with FGDC metadata on exports. Built to support a licensed surveyor's review — never to replace it.
Ellipsoidal heights natively; NAVD88 orthometric via GEOID18 with rigorous per-cell undulation — not a single site-wide constant. Vertical adjustment from your control points with the full before/after disclosure in the report.
State Plane coordinate systems with county-correct zone selection (Florida's zone boundaries follow counties, not latitude — we resolve them the way a surveyor would), US survey feet handled exactly.
DXF (3D breakline-ready polylines, points, TIN faces) and LandXML 1.2 surfaces that rebuild correctly in Civil 3D-class software — northing/easting order and all the ingest foot-guns handled.
Every output carries the exact engine and module versions that produced it. Two years from now, you can state precisely how a deliverable was made — and regenerate it.
Run your hardest site through it yourself — free trial, full engine, next to your current tool's output and your checkpoints if you have them. Keep whichever is better.
Start your free trial Launch PRISM·lidarcloud