A building seen from above has no obvious outline. The roof shifts away from its base, adjoining constructions touch, outbuildings attach badly, and each of those situations calls for a decision the image alone does not supply.
This article sets out those difficulties. It extends the article on semantic segmentation.
The uses that structure the field
Six built-environment applications occur in satellite imagery.
Updating a cartographic base that describes the built environment.
Monitoring artificialisation along with the spread of urban sprawl.
Counting the population from the housing stock observed.
Assessing a property stock’s exposure to a given hazard.
Detecting the constructions that have never been declared.
And preparing interventions in the wake of an event.
One important practical consequence follows. The third use is satisfied by a count while the first requires an exact footprint, which separates two projects whose annotation cost is not of the same order of magnitude.
What the footprint must designate
Four conventions occur in satellite imagery for delimiting a building.
The roof outline as it appears directly on the image.
The ground footprint, deduced by correcting the roof offset.
The outline of the built plot, all outbuildings included.
And a simple bounding box, sufficient for a count.
One important observation follows for a satellite imagery project. The first two conventions produce different surfaces on the same building, a gap growing with height, which makes confusing them invisible in a suburban area and considerable in a city centre.
What the roof offset imposes
Four satellite imagery effects accompany an oblique capture.
The top of the building shifts visibly relative to its base.
The facade becomes visible on one of the construction’s sides.
The gap grows with the building’s height and the obliquity of the view.
And two distinct acquisitions no longer superimpose quite exactly.
One practical consequence follows. The last effect hampers any temporal monitoring, one same block appearing offset between two dates with no construction having moved, which produces false change detections in dense areas.
What adjoining construction imposes
Four situations make the separation ambiguous in satellite imagery.
Terraced houses that form an entirely continuous row.
A single block composed of several distinct building bodies.
A recent extension directly attached to the main building.
And an inner courtyard that is entirely surrounded by construction.
One important observation follows. The first situation recurs massively in older fabric, a row of terraced houses being countable as one object or as ten depending on the rule adopted, which makes this convention decisive for any count.
What outbuildings require
Five elements raise a question of scope in satellite imagery.
A garage that stands clearly detached from the main dwelling.
A garden shed of very small surface.
A conservatory or indeed any other lightweight extension.
An agricultural building standing isolated in the middle of a field.
And a construction still under way, therefore incomplete.
One important practical consequence follows for a satellite imagery project. The last element deserves a rule of its own, a building site appearing as a partial building whose annotation depends on what the project seeks to count, existing constructions or construction activity.
What urban density changes
Four satellite imagery effects accompany a dense urban fabric.
The number of objects present per tile becomes very high.
Cast shadows come to mask some of the lowest buildings.
The roofs meet each other with no visible separation whatsoever.
And urban vegetation comes to partly cover the constructions.
One observation follows. The second effect is the most underestimated, a narrow street lined with tall blocks staying in shadow for much of the day, which makes some constructions undetectable depending on the hour of capture.
What the geometry of built form brings
Four regularities ease the work in satellite imagery.
The angles of constructions are predominantly right angles.
Facades stay rectilinear on most constructions.
Alignments most often follow the line of the road network.
And dimensions stay within predictable ranges by type of built form.
One important practical consequence follows. Those four regularities justify specific tracing tools, a building outline gaining from being built by segments and right angles rather than point by point, which improves both throughput and geometric quality.
What the reference requires here
Four sources establish a ground truth in satellite imagery.
The land registry, very precise on the footprint but sometimes behind.
Address bases along with the administrative construction registers.
Field surveys conducted on a deliberately restricted sample.
And very high resolution imagery acquired over the same areas.
One important observation follows for a satellite imagery project. The first source describes the ground footprint while the image shows the roof, which produces a systematic gap any evaluation must take into account before concluding to a system error.
What the type of built form changes
Five contexts raise different difficulties in satellite imagery.
Suburban housing, regular and well separated, by far the simplest to handle.
The old centre, dense and terraced, where separation becomes purely conventional.
The industrial estate, with very large buildings and uniform roofs.
Dispersed rural housing, where outbuildings dominate largely in number.
And informal settlement, with very irregular forms and varied materials.
One important practical consequence follows. The last context defeats most of the field’s geometric regularities, a construction with neither right angles nor alignment demanding a point-by-point trace, which multiplies the time per object in the very areas where counting matters most.
What the measurement must reflect
Five indicators describe such a satellite imagery system’s performance.
The proportion of buildings detected, systematically broken down by size.
The accuracy of the count obtained on reference areas.
The quality of the footprints produced, measured by surface overlap.
The rate of wrong merging or splitting observed between neighbouring buildings.
And the behaviour observed on urban fabrics absent from the training.
One important practical consequence follows. The fourth indicator is specific to this field and rarely produced, a row of houses counted as one object degrading the count without affecting the built surface detected, which makes both measurements necessary.
What change monitoring requires here
Four conditions make a comparison usable in satellite imagery.
Acquisitions made at comparable viewing angles.
A footprint convention strictly unchanged between the dates.
An explicit rule on what constitutes a new construction.
And explicit treatment of buildings left partly masked.
One important observation follows for a satellite imagery project. The third condition deserves settling at scoping, an extension, a rebuild or a change of roof being able to count or not as a change, a distinction belonging to the administrative use targeted rather than to observation.
What vegetation masks
Four situations reduce a construction’s visibility in satellite imagery.
An isolated tree whose crown covers a part of the roof.
A line of street trees masking both the facades and the edges.
A densely planted garden entirely surrounding a low house.
And a leafy season concealing everything the winter had revealed.
One important practical consequence follows for a satellite imagery project. The last situation opens a useful possibility, a winter acquisition in a temperate zone showing constructions the summer conceals, which makes the choice of season a lever for detection rather than a constraint endured.
What height adds where it is available
Four satellite imagery contributions follow from relief information on built form.
Separating a building from its own immediate surroundings becomes clear-cut.
Two adjoining constructions of different heights are finally clearly distinguished.
The roof offset is then corrected by calculation rather than by estimate.
And built volume is deduced directly from the footprint and the elevation.
One important observation follows. The second contribution partly resolves the chapter’s central difficulty, a difference in height supplying an objective criterion of separation where the image alone offered only a convention, which explains the interest of LiDAR on urban counting projects.
What updating a base implies
Four operations compose a reconciliation with an existing cartographic base.
Identifying the buildings that are present simultaneously in both sources.
Spotting those detected on the image but absent from the existing base.
Spotting those recorded in the base but absent from the recent image.
And flagging the footprints whose geometry has appreciably changed.
One important practical consequence follows for a satellite imagery project. The third operation produces the most false flags, a construction masked by vegetation or shadow appearing demolished, which requires treating these cases as verifications to be conducted rather than as observed disappearances.
What the corpus must cover here
Five axes structure a built-environment dataset in satellite imagery.
The types of urban fabric covered, from suburban housing to informal settlement.
The regions handled, whose forms and roofing materials differ.
The construction heights covered, which command the offset of the roofs.
The shadow conditions encountered by hour and season of acquisition.
And the situations of adjoining construction, rare in the periphery and dominant in old centres.
One important observation follows. The second axis is the most demanding to cover, a roof of tiles, of sheet metal or of planted terrace having no common visual signature, which makes a corpus built on one region far narrower than its size suggests.
What this work’s throughput presupposes
Four factors determine the time spent per building in satellite imagery.
The geometry finally adopted, from the simple point to the detailed outline.
The construction’s regularity, which permits or forbids a trace by segments.
The fabric’s density, which multiplies the separation decisions.
And the presence of cast shadows or masking vegetation.
One important practical consequence follows for a satellite imagery project. The third component explains a considerable gap between contexts, an old centre demanding a separation decision at almost every building while a housing estate demands none, which makes the urban fabric as decisive as the number of constructions to handle.
What the annotator must know here
Five kinds of knowledge condition correct satellite imagery work on built form.
Reading a roof according to its material and its shape.
Distinguishing a building from a mineral surface of the same tone.
Spotting the offset existing between the roof and the ground base.
The separation rule adopted for adjoining constructions.
And the exact scope of the outbuildings actually recorded in the reference.
One important observation follows for a satellite imagery project. The second kind avoids a frequent error, a concrete yard, a car park or a terrace presenting from above a tone close to certain roofs, which produces false positives no geometric regularity permits ruling out.
What simplifying the outlines changes
Four satellite imagery effects accompany a regularisation of the delivered footprints.
The delivered outlines become polygons with perfectly clean angles.
The number of points describing each single building falls sharply.
The measured surface departs slightly from that of the initial trace.
And the result obtained comes visually closer to a cartographic base.
One important practical consequence follows for a satellite imagery project. The third effect deserves announcing rather than being discovered, a regularisation modifying surfaces by a few percent, a gap negligible for a map and awkward for a tax calculation or a development right.
What the first batch must establish
Four results justify a satellite imagery trial batch before production.
An average time per building, measured separately on each type of fabric.
The list of adjoining configurations the reference had not foreseen.
The gap observed between the footprints produced and a reference base.
And the disagreement rate between two operators on the same tiles.
One important observation follows for a satellite imagery project. The third result avoids a lasting misunderstanding, a systematic gap revealed on fifty buildings being explained by the footprint convention rather than by an error, which permits correcting it at scoping instead of discovering it at acceptance.
What the delivery format requires here
Four choices condition the client’s reuse of the satellite imagery footprints.
The vector format adopted along with the coordinate system finally employed.
Whether or not a stable identifier is present for each building.
The attributes attached, such as computed surface or construction class.
And the linkage to the identifiers of a pre-existing base.
One important practical consequence follows for a satellite imagery project. The last choice determines the deliverable’s usefulness for an update, unlinked footprints obliging the client to redo the reconciliation themselves, work sometimes longer than the annotation itself.
What the provider brings here
Four contributions distinguish a built-form detection engagement in satellite imagery.
A footprint convention settled and then applied with no variation at all.
Tracing tools fully exploiting the geometric regularity of constructions.
Explicit treatment of adjoining construction, written before production.
And deliberate coverage of the hardest fabrics rather than of suburban housing alone.
One important practical consequence follows. The second contribution appreciably changes the project’s economics, a trace by segments and right angles producing more regular outlines in less time, which improves geometric quality and cost per building at once.
Approaching a building detection project
Five questions scope such a satellite imagery project.
Is counting or delimiting required. The cost differs strongly.
Does the footprint follow the roof or the base. The surfaces differ.
How are adjoining constructions handled. The count depends on it.
Do outbuildings fall within scope. The rule must be written.
And does the reference describe the roof or the ground. The gap is systematic.
Those five answers determine the corpus’s consistency. Asking them before production avoids a count nobody will be able to compare with an existing base.
The question that frames the project
One question determines the project’s cost in satellite imagery.
Is the expected result a count or a geometry.
A count, a population estimate or monitoring of artificialisation, is satisfied by a point or a box per construction, which permits a high throughput and a light reference.
A geometry, a land registry update or a surface calculation, requires a precise outline, a footprint convention and explicit treatment of adjoining construction.
That question belongs to the administrative use and not to technique, it is asked before the first tile, and it separates two projects whose cost per building differs by an order of magnitude.
Three decisions before producing
Three decisions commit a building detection project in satellite imagery.
Choosing between roof outline and ground footprint, the resulting surfaces differing.
Writing the separation rule for adjoining constructions, on which every count depends.
And fixing the scope of outbuildings, garages, sheds and building sites included.
Those three decisions cost one meeting, they precede the first tile, and their absence produces a count nobody will be able to compare with an existing base.
Three checks on a building corpus
Three checks qualify a built-environment dataset in satellite imagery.
The footprint convention adopted, roof outline or ground base.
The written separation rule for adjoining constructions.
And the presence of varied urban fabrics among the annotated tiles.
Those three checks are each asked in one question, they require no cartographic competence, and their absence indicates a corpus whose counts will match no existing base.
What this chapter teaches
One cross-cutting observation deserves closing this examination.
A building is an object whose limits belong to a convention.
Three findings compose it.
The roof outline and the ground footprint produce different surfaces.
A row of terraced houses counts as one object or as ten depending on the rule.
And the land registry describes the ground where the image shows the roof.
That finding extends the preceding chapter, instance detection adding decisions surface segmentation did not raise.
What this chapter leaves to the next
A building sometimes carries on its roof what the project genuinely seeks.
Three questions stay open in satellite imagery.
How to detect installations whose size approaches that of the pixel.
What a regular visual signature brings to detection.
And how to estimate a usable surface rather than a mere count.
Those three questions belong to solar panel detection, which constitutes the subject of the following chapter.
Why the footprint convention belongs in the first meeting
One item is worth settling before any other on this kind of project.
State whether the outline follows the roof or the ground base.
Three reasons follow in satellite imagery.
The two produce different surfaces on every building above one storey.
The difference is invisible on the suburban tiles usually shown as samples.
And the client’s own reference base uses one of the two without saying so.
One important practical consequence follows for a satellite imagery provider. Raising this first signals experience more efficiently than any capability statement, since a client who has been asked the question understands immediately that the provider has delivered against a land registry before, and a client who has never been asked it usually discovers the gap at acceptance.
Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.
- Confusing roof outline with ground footprint.
- Leaving implicit the separation rule for terraced houses.
- Comparing a result with the land registry without correcting the systematic gap.
- Omitting a rule on outbuildings and constructions under way.
- Measuring built surface without measuring the count.
- Neglecting the effect of shadows in dense fabric.
- Monitoring change without allowing for the offset of roofs.
- Tracing an outline point by point rather than by segments.
- Evaluating on a single type of urban fabric.
- Ignoring vegetation that partly masks the roofs.
What to take away
Building detection rests on conventions more than on observation.
Three readings emerge. The roof outline and the ground footprint produce different surfaces on the same building, a gap growing with height, invisible in a suburban area and considerable in a city centre. A row of terraced houses counts as one object or as ten depending on the rule adopted, which makes that convention the decisive parameter of any count. And the land registry describes the ground footprint where the image shows the roof, which produces a systematic gap any evaluation must correct before concluding to a system error.
For isolated installations, the article on solar panels details the approach. For networks, the article on infrastructure mapping sets it out.
To explore delivery arrangements, supported formats and applicable control mechanisms, see our dedicated page on annotation for geospatial. And if you are preparing a building detection project, let us discuss your need.
