You can now reconstruct a scene. Chapter 15 gave you a surface through the measured points, and chapter 16 gave you solids with named parts, fitted roof planes and a ridge. That is a correct answer to the question those chapters asked, and it is rarely the answer a survey was commissioned for. Think about who pays for the survey. The energy office wants to know which roof faces receive enough sun for photovoltaics. The insurer wants the construction year and the volume of every building in a flood plain. The facility manager wants to click on a wall and see what it is made of. Each request names a kind of object, attaches a property to it, and needs the answer to hold for one identified thing in the world, this roof, on this house, under this register entry. Your mesh provides none of the three answers.
Why not? Consider what a mesh holds. It carries vertex coordinates and the triangles that join them, and nothing else. Which of those triangles form a roof? Where does one house of a terraced row end and the next begin? Which register entry does each house correspond to? A triangle list has no field to hold any of it. So what you have is the raw material of a product rather than the product itself. That product is what this chapter is about, and I will call it a model.
The gap between the two is institutional as much as it is technical. What counts as one building, which attributes it carries, and how its surfaces are typed are decisions taken by people. They are useful only when every producer and every consumer takes them the same way, so they are written into published data models rather than into any one pipeline. I would put the practical consequence more strongly than the literature usually does: on a semantic deliverable the specification fixes more of what your product is worth than the reconstruction algorithm does. So write a good specification against a modest pipeline and you end up with a better dataset than a buyer who writes none against an excellent pipeline. Is that a claim about small projects only? National mapping agencies maintain semantic building models of entire countries on exactly that basis, about fifty-eight million buildings in Germany and close to eleven million in the Netherlands. Each is regenerated on a fixed cycle from national airborne LiDAR of the kind you have been processing since chapter 1.
Each of those five questions marks a place where a delivered dataset fails in practice. Here figure 17.1 lays out the route that answers them, from reconstructed geometry through validity and semantics to a served, maintained product. The same building travels with you the whole way, an Amsterdam canal house of 1758 that the national building register identifies as NL.IMBAG.Pand.0363100012179545. The Dutch national model publishes it at three levels of detail (figure 17.8). We reconstruct it again ourselves with an open engine on the national airborne survey (figure 17.6), the two ridge heights agreeing to 24 mm. So here are your questions. What must a model state that geometry does not? When is a solid valid, and what does an invalid one quietly do to a volume? How much of the building does each level of detail contain? Which standard and which encoding does your consumer need? And what separates a maintained model from one that is merely called a twin? Let's start with the five ingredients, since the rest of the chapter supplies them one at a time.
Five ingredients separate the delivered product from the computed geometry, and reconstruction supplies only the first. Let's see what that costs you on the canal house. Reconstructed at LoD2.2 in figure 17.6, it comes out as a solid of 190 faces, and 28 of them lie on the roof. Nothing in the file records which 28. So when somebody asks you for the south-facing roof area of that house, you have to decide from face orientation which faces are roof faces. Then you decide which of those form one roof plane rather than several. Both decisions are yours rather than the data's, and both move when the grade moves, the same house carrying 45 faces at LoD1.3 and 18 at LoD1.2. You get closer with the fitted model of chapter 16 than with the raw mesh of chapter 15, its roof planes and its ridge being already explicit. It still records none of the four further ingredients below.
You have just read the opening section. Both options unlock it immediately.
Read all 20 chapters online here, and download the complete PDF.
Everything in the book, plus the companion video course as it ships.
By Abderrazzaq Kharroubi, geomatics engineer at the University of Liège: a decade of LiDAR fieldwork, 2,000+ students taught. One payment, lifetime access, 30-day refund.
Anything you highlight or note stays with you after you buy. Compare all tiers or sign in if you already bought.