A point cloud is the native product of LiDAR (and dense photogrammetry): millions to billions of 3D points, each carrying coordinates plus attributes — intensity, return number, timestamp, and after processing, a classification code. Classification turns an undifferentiated cloud into structured data: ground here, vegetation there, buildings, water, wires — the step on which every derived product depends.
Anatomy of the data
The LAS format (and its compressed twin LAZ) is the industry standard, storing per-point attributes and standardised classification codes maintained by ASPRS: ground, low/medium/high vegetation, building, water, rail, road, wire classes and more. Projects add bespoke classes — conductor versus shield wire, pylon components — as specifications demand.
Point density is the resolution analogue: 2 points/m² supports terrain modelling; 30+ supports building reconstruction; hundreds enable engineering-detail extraction.
How classification is actually done
Automated routines do the heavy lifting: morphological and TIN-based ground filters, geometric rules and increasingly deep learning for structures. But automation degrades exactly where products matter most — steep terrain, dense canopy, complex urban fabric, wires — so production workflows finish with skilled manual editing.
The realistic split on demanding specifications is substantial manual refinement, which is why per-tile QC dashboards, class-accuracy sampling and documented editing standards distinguish production vendors from script-runners.
What classified clouds unlock
From ground points: DTMs, contours, hydro-enforced surfaces. From vegetation: canopy height, biomass, fuel loads, encroachment against clearance envelopes. From buildings: footprints and 3D models. From wires: catenary models for powerline engineering.
Every one of these inherits the classification’s errors — a misclassified berm becomes a phantom flood barrier — which is why accuracy reporting per class, not just overall, is the professional norm.