Levels of Detail (LOD1-4)

The phrase level of detail (LOD) is borrowed from computer graphics, where objects far from the camera are drawn coarsely and near ones finely to save effort. CityGML adapts the idea into something more principled: a defined set of detail levels, each specifying how completely and accurately the real world is represented, so that producers and users of a city model share a common language for 'how detailed is this?' The value of standardising it is exactly the value of any standard - it lets a buyer specify what they need, a supplier state what they delivered, and two datasets be compared on equal terms. Without it, 'detailed 3D model' is a marketing phrase that could mean almost anything.

Crucially, level of detail in CityGML is not only about geometry - how many surfaces, how fine the shapes - but also about semantic richness and accuracy. A higher LOD generally means more geometric resolution (roofs, then openings, then interiors), more semantic decomposition (a building becomes roof, walls, rooms, installations), and tighter positional accuracy. These tend to rise together but are not the same axis, and a thoughtful specification says something about each: you can have geometrically simple buildings that are semantically well attributed, or richly shaped buildings with almost no attributes attached. Honest practice states the LOD and the attribute completeness separately, because a twin often cares more about the attributes than the shapes.

It also matters that a single city model need not be at one uniform LOD. Real programmes routinely mix levels - plain blocks across the whole city for context, higher detail around a project site or a zone of interest, interiors only for a handful of critical buildings. Treating LOD as a dial you set once, globally, is a beginner's error that inflates cost for no benefit. Treating it as a budget you spend where it earns its keep is the professional move. And a final caution carried from Lesson 2.1: a higher LOD looks more authoritative and real, which makes it more seductive and more dangerous - a detailed model can still be stale, biased or wrong, and the extra detail can lend false confidence to data that does not deserve it. Detail is a tool, not a virtue.

Walk the ladder and the purpose of each rung becomes clear. LOD1 is the block model: each building is its footprint extruded to a single height, a flat-topped prism. It is crude - no roof shape, no facade - but it is cheap to generate at scale (often automatically from a 2D footprint plus a height) and surprisingly useful: for city-wide studies of built volume, density, massing, rough shadow and first-pass energy demand, a sea of correctly sized blocks is often all you need. Much of a city can live happily at LOD1.

LOD2 adds the roof shape and major surfaces: buildings now have pitched or flat or complex roofs, and their walls are distinguished as surfaces. This is the workhorse level for city twins, because roofs matter enormously - for solar-panel suitability, for realistic shadow and wind, for district-scale energy, and for a model that actually looks like the city rather than a warehouse estate. LOD2 across a city, generated from aerial survey, is a common and sensible target. LOD3 goes to the detailed exterior: real window and door openings, balconies, facade articulation, detailed roof superstructures. This is what you need for street-level design, daylight studies through actual openings, facade analysis and convincing close-up visualisation - but it is far more expensive to capture and maintain, so it is usually reserved for areas and buildings that warrant it, not a whole city.

LOD4 adds the interior: rooms, internal walls, doors, stairs, sometimes furniture and building installations - the model now describes the inside of the building, not just its shell. LOD4 is essential for indoor applications (wayfinding, evacuation modelling, facilities and operations) and is where a city model meets building-scale BIM, but it is the costliest by far, intrudes on the occupied, private interior, and is almost never justified city-wide. In practice you build LOD4 for the few buildings where indoor questions genuinely matter. (A fair caution: newer revisions of CityGML have reorganised this scheme and handle interiors differently, so treat LOD1-LOD4 as the durable mental model rather than a frozen specification, and check the version in use.) The throughline is simple: each rung answers questions the one below cannot, and costs far more to earn.

The discipline at the heart of this lesson is fit: choose the level of detail from the question you are trying to answer, not from what is technically possible or what looks impressive in a demo. Start from the use and read the LOD off it. Studying city-wide density, massing options, or a rough shadow envelope across a district? LOD1 blocks are enough, and anything more is wasted money and slower tools. Assessing rooftop solar potential, district energy, or realistic shadow and wind with true roof forms? You need LOD2. Designing at street level, studying daylight through real windows, or analysing a specific facade? That calls for LOD3, around the site that matters. Routing people through a building, modelling an evacuation, or running indoor operations? Only LOD4 will do, for those specific buildings.

Notice the pattern: the detail you need is tightly coupled to the scale and nature of the question. Broad, city-scale, volumetric questions live low on the ladder; narrow, building- or street-scale, surface- or interior-specific questions live high. This is why the mixed-LOD model is almost always the right architecture for a twin: a coarse base everywhere for context and city-wide analysis, with detail concentrated where specific questions demand it. A competent brief does not say 'build me a LOD3 city'; it says 'LOD1 across the municipality, LOD2 where we study solar and energy, LOD3 around these three regeneration sites, LOD4 for the two transport interchanges where we model crowds.' That brief spends detail like a budget.

There is also a quieter accuracy dimension to fit. A model's positional accuracy should match the decision it supports: a rough options study tolerates loose geometry, while anything approaching a real dimensional decision needs survey-grade data from the proper custodians - and, as always, binding decisions about boundaries, structures and infrastructure rest with the authorities, engineers and surveyors, never with the model's convenient geometry. The skill, then, is twofold: read the required LOD off the purpose, and state the accuracy the purpose demands - and resist, firmly, the pull to over-build detail that no one asked for and no one will maintain.

It feels counterintuitive, even irresponsible, to argue for less detail - surely a more detailed model is a better model? In a twin, usually not, and understanding why is a mark of real competence. Every rung up the ladder multiplies four separate burdens. Capture cost: LOD1 can be generated almost automatically; LOD3 and LOD4 demand intensive survey, scanning, modelling and manual work, building by building. Data volume and performance: high-LOD city models are enormous, slow to load, hard to stream to a browser or a phone, and expensive to store and serve (Module 5). Maintenance: this is the killer. A model is only useful if it is current, and the more detail it carries, the more there is to keep up to date as the real city changes - and cities change constantly. A LOD1 block is trivial to update; a LOD3 facade or a LOD4 interior decays into fiction the moment a renovation happens and no one re-surveys it. False confidence: a gorgeously detailed model invites people to trust it more than its underlying data warrants, lending a spurious authority to what may be stale or wrong.

There is a fifth, ethical cost that higher detail carries, especially indoors. LOD4 interiors and fine facade data describe where people live and work in intimate detail, raising real privacy and surveillance questions - the kind Module 8 confronts. Detail that no decision requires but that exposes private space is not neutral thoroughness; it is data you now have to protect, justify and govern, and that a poorly run twin could misuse. The honest default is to capture the least detail that answers the question, particularly where the detail reaches into occupied, private space.

So the mature judgement inverts the naive instinct. The question is never 'how much detail can we capture?' but 'how little detail do we need to answer our questions well, and can we actually keep it current?' A sprawling LOD3 city model that no one can afford to maintain is worse than a clean LOD2 one that stays true, because a twin's whole value rests on being a faithful, current mirror - and a detailed model that has quietly gone stale is a particularly convincing lie. Match detail to purpose, spend it where it earns its keep, and treat every extra rung as a maintenance commitment you are promising to honour for years, not a one-time flourish.