What Satellite Data Measures When It Maps a Building

Building footprint layers have become the quiet workhorse of property analysis. They join cleanly to parcels, they aggregate neatly to census blocks, and they measure something slightly different from what most workflows assume they measure.

What the layer actually digitizes

A footprint derived from overhead imagery is a roof outline. The sensor sees the top of the structure, the extraction traces the boundary it can see, and the polygon that lands in the table is the roof edge projected onto the ground plane.

For a single-story building photographed near nadir, roof edge and wall base sit close enough that the distinction rarely matters. For anything taller, or anything imaged off-nadir, the two separate. The offset is non-uniform across a scene because it scales with building height, so a citywide layer carries a systematic error that varies building by building.

Which sensor and which vendor a footprint came from therefore matters more than it appears, and a real-estate data guide by Sebastian Holt works through the satellite data options and providers used in real estate, including where aerial collection wins outright.

Research on off-nadir extraction treats the offset as the central problem, and recent methods explicitly predict a roof-to-footprint offset vector rather than assuming that roof and base coincide.

The convention question is worth settling too. OpenStreetMap's building tagging documents the ambiguity directly: contributors tracing from imagery capture roof outlines, ground surveys capture wall lines, and both land in one tag.

Four things that move the edge

Four distinct effects push the traced polygon away from the wall line, and they do not cancel out.

Effect	What it does to the polygon	Where it bites hardest
Roof overhang	Extends the outline beyond the wall line	Residential eaves, agricultural structures
Off-nadir lean	Displaces the roof from the base, scaling with height	Tall buildings, scene edges
Shadow	Obscures one side, biasing the traced edge inward	Low sun angle, dense urban blocks
Adjacency merging	Fuses neighboring structures into one polygon	Row housing, warehouses, attached garages

The first three shift an area estimate, and the fourth changes the count, which is the more damaging of the two in most workflows.

Roof outline against wall line. Overhang widens the polygon, and off-nadir viewing displaces it by an amount that grows with building height. Source: own diagram, based on published off-nadir extraction research.

Of those four, only overhang behaves like a constant, and the rest vary per building.

Why merging matters more than area

Anyone who has joined a footprint layer to a parcel layer has met the merged polygon. Two attached row houses become one feature, the join assigns it to one parcel, and the neighboring parcel comes back with zero buildings.

The area error from overhang is usually a few percent and behaves like noise across a large sample. A merged polygon behaves differently: it removes a row, doubles an area, and misassigns both to one owner. Aggregate that to a census block and the building count is wrong in a way no area correction recovers, because the error is topological rather than metric.

The practical consequence for anyone building tiles or running block-level statistics of the kind documented in census data workflows here is that footprint counts need a sanity check against an independent source before they carry any weight. Parcel counts, address points, and unit counts from the census all serve, and disagreement between them locates the merged features quickly.

One property or a portfolio

The second thing a footprint layer cannot settle is whether satellite imagery was the right source for the job.

For a single property, aerial collection wins on nearly every axis that matters. EagleView publishes 1-inch ground sample distance, and Nearmap captures in the range of roughly 1.7 to 2.8 inches, against 30 centimeters for the best commercial satellite products. Add oblique views, street-level imagery, and an actual site visit, and satellite imagery loses on detail, on roof condition assessment, and frequently on cost as well.

The satellite case is reach. Screening several thousand dispersed assets, detecting construction progress across a region, or maintaining a consistent refresh cycle over a whole market is work aerial programs cannot match on coverage or price. The dividing line is repetition across area rather than resolution.

Before joining a footprint layer to anything

Five checks catch nearly every problem that footprint data creates downstream, and none needs specialist software.

• Check the collection geometry of the source, since off-nadir imagery displaces roofsby an amount that varies with height
• Compare feature counts against parcels or address points, because merged polygons show upas a count mismatch rather than an area one
• Establish whether the vendor traces roof edge or models the wall line, as the two are different products
• Look at the vintage per tile, not per dataset, since national layers are assembled from many collection dates
• Decide whether the task needs one property or a portfolio, because that answer picks the sensorbefore any vendor comparison starts

None of this argues against footprint layers. They remain the cheapest route to structure-level geography at national scale, and for density, growth, and screening work they are entirely adequate.

The useful framing

A footprint layer is a measurement of roofs, published as a measurement of buildings, and used as a measurement of property. Each of those translations is reasonable and each one loses something.

Knowing which translation a given analysis depends on is what separates a defensible number from a plausible one. For counting structures across a region, the layer is fine as delivered. For anything where an individual polygon carries weight, the source geometry and the collection date deserve a look before the join.