A road is not an object, it is a continuity. It crosses the scene, disappears beneath the trees, reappears further on, and that linear nature imposes decisions no surface detection ever had to take.
This article sets out those difficulties. It extends the article on multispectral imagery.
The uses that structure the field
Six network applications occur in satellite imagery.
Updating an already existing road base.
Mapping territories still poorly covered by public data.
Monitoring the progress of a large linear worksite.
Inventorying all the existing overhead electricity networks.
Assessing accessibility in the aftermath of a disaster.
And checking the apparent condition of the carriageways.
One important practical consequence follows. The fifth use requires a short delay, an accessibility map delivered after access has been restored having no operational usefulness at all.
What linear geometry imposes
Five decisions structure the annotation of a network in satellite imagery.
A centreline or else a surface footprint.
The treatment given to intersections and interchanges.
The continuity rule applied beneath a masking obstacle.
The level of detail adopted for curves and bends.
And the road hierarchy adopted, from motorway down to track.
One important observation follows for a satellite imagery project. The first decision separates two incompatible corpora, a centreline serving route calculation and a footprint serving calculation of sealed surface, which forbids converting one into the other afterwards.
What continuity requires
Five satellite imagery situations interrupt a network’s visibility.
A tree canopy entirely covering a country road.
A tunnel, which is by construction wholly invisible.
A shadow cast by a block in an urban environment.
A bridge crossing an obstacle without standing out from the carriageway.
And a carriageway whose tone approaches that of the surrounding ground.
One practical consequence follows. Those five situations require an explicit convention on what an annotator may infer, a road’s continuity staying probable without being observed, which places this field between observation and deduction.
What the road hierarchy presupposes
Five levels compose a common classification in satellite imagery.
Express routes, which are both wide and well contrasted.
Main roads, which are almost always surfaced.
Secondary roads, both narrower and more winding.
Local access ways along with ordinary streets.
And unsurfaced tracks, sitting at the limit of the detectable.
One observation follows. The last level concentrates most of the disagreement between operators, a dirt track being poorly distinguished from a field boundary or a machinery pass, which makes its inclusion in the reference a heavier decision than it appears.
What non-road networks add
Five linear objects occur in satellite imagery beyond roads alone.
A railway, which is very regular and generally well visible.
An overhead power line, identifiable only by its pylons.
A canal or else a fully managed watercourse.
An airfield runway, both wide and strongly contrasted.
And a buried pipeline, visible only through the trench above it.
One important practical consequence follows for a satellite imagery project. The second object is not annotated like the others, a cable staying invisible at most resolutions, which requires annotating the pylons then deducing the route rather than following the line itself.
What the reference requires here
Four sources establish a road ground truth in satellite imagery.
Public road bases, very broad but unevenly up to date.
Collaborative data, rich but of very variable quality.
Field surveys conducted on a deliberately restricted sample.
And movement traces recorded by the vehicles themselves.
One important observation follows. The last source holds a particular interest, a trace proving a way is passable, information the image never supplies and which separates a usable road from a merely visible alignment.
What the graph adds to the geometry
Five properties separate a network from a mere set of lines in satellite imagery.
The segments meet each other at clearly identifiable nodes.
A direction of travel can attach to each of the segments.
An explicit hierarchy orders the ways among themselves.
Grade separations are then distinguished from true intersections.
And the whole’s connectedness becomes a directly measurable property.
One important observation follows for a satellite imagery project. The fourth property escapes the image entirely, a bridge and a crossroads resembling each other seen from above, which requires either a resolution permitting the level difference to be seen or an outside source to settle it.
What the measurement must reflect
Five indicators describe such a satellite imagery system’s performance.
The proportion of length detected, systematically broken down by road level.
The continuity of the network produced, measured by its number of breaks.
The accuracy of the position of the centreline traced.
The correctness of the hierarchical classification assigned.
And the effective connectivity of the graph obtained.
One important practical consequence follows. The last indicator is specific to this field and rarely measured, a break of a few metres cutting a whole route, which makes a network detected at ninety-nine percent sometimes unusable for route calculation.
What updating a base implies
Four satellite imagery operations compose a reconciliation with a road base.
Identifying the segments present simultaneously in both sources.
Spotting the new ways still absent from the base.
Spotting recorded segments not found again on the image.
And flagging observed gaps in position or in classification.
One important practical consequence follows. The third operation produces mostly false flags, a road masked by canopy staying rightly recorded, which requires treating it as a verification to be conducted rather than as an observed disappearance.
What consistency between tiles requires here
Four satellite imagery defects appear at the junction of two neighbouring tiles.
A segment that stops dead at the very edge of the tile.
A lateral offset of the centreline between two neighbouring tiles.
A different classification of the same segment either side of the junction.
And an intersection annotated twice with distinct geometries.
One important observation follows for a satellite imagery project. The first defect is the most serious on this precise subject, a break at the junction producing exactly the kind of cut that makes a graph unusable, which requires a control on the reconstituted mosaic rather than tile by tile.
What the corpus must cover here
Five axes structure a network dataset built in satellite imagery.
The types of territory covered, from dense urban to the most isolated rural.
The road levels actually present, from express route down to dirt track.
The masking situations encountered, whether tree canopy or urban shadow.
The intersection configurations, from simple crossroads to full interchange.
And the regions handled, whose surfacings and alignments differ.
One important practical consequence follows. The third axis demands a deliberate effort, a corpus made of well-cleared tiles teaching the system to follow what is visible without ever teaching it to bridge what is hidden, which produces precisely the breaks that make the network unusable.
What this work’s throughput presupposes
Four factors determine the time spent per satellite imagery tile.
The density of the network actually present in the scene.
The number of intersections that must be modelled correctly.
The frequency of maskings requiring an inference from the operator.
And the level of detail adopted for tracing the curves.
One important practical consequence follows. The second factor explains a considerable gap between territories, a motorway interchange demanding more time than a kilometre of straight road, which makes a per-kilometre rate misleading without an indication of the expected node density.
What the annotator must know here
Five kinds of knowledge condition correct work on this subject.
The inference rule beneath an obstacle, applied without the least interpretation.
The distinction between a carriageway and a linear parcel boundary.
The modelling precisely expected for intersections and grade separations.
The hierarchical classification criteria actually adopted.
And the threshold below which a track does not enter the reference.
One important observation follows for a satellite imagery project. The second kind avoids a frequent error in rural settings, a cut hedge, a ditch or a crop boundary forming a line as clean as a way, which produces invented segments whose presence degrades the graph as much as an omission does.
What the delivery format requires here
Four choices condition the client’s reuse of the satellite imagery network.
The structure finally delivered, topological graph or simple collection of lines.
The attributes attached, such as hierarchical level and presumed surfacing.
The explicit marking of segments obtained by inference rather than by observation.
And the linkage to the identifiers of a pre-existing road base.
One important practical consequence follows for a satellite imagery project. The third choice distinguishes a serious engagement, a segment restored beneath a tree canopy not holding the same status as a segment seen, which lets the client decide for themselves what credit they extend to each portion of the network.
What pre-annotation permits here
Three uses appreciably lighten satellite imagery work on this subject.
A route proposal bearing on the wide and well contrasted ways.
An automatic flagging of gaps between the image and an existing base.
And a systematic detection of breaks within the network produced.
Two caveats accompany it in satellite imagery.
An automatic proposal stops exactly where the canopy begins, precisely where inference was needed.
And the topology of intersections stays entirely to be reworked by hand.
What the first batch must establish
Four results justify a satellite imagery trial batch before production.
The disagreement rate between two operators on the masked segments.
The proportion of the length obtained by inference rather than by mere observation.
The number of breaks observed after the full reassembly of the mosaic.
And the average time per intersection, compared with the average time per kilometre.
One important observation follows for a satellite imagery project. The third result directly measures what makes the deliverable valuable, a trial on a few contiguous tiles revealing whether the arrangement produces a connected graph or a collection of fragments, a question no tile-by-tile control permits settling.
What poorly mapped territories change
Four satellite imagery particularities separate an initial mapping from an update.
No existing base then serves as a starting point or as a control.
Surfacings and uses differ markedly from the usual references.
The road hierarchy maps badly onto an imported nomenclature.
And the ground truth very often reduces to movement traces.
One important practical consequence follows for a satellite imagery project. The third particularity deserves its own scoping time, a main track serving an entire region belonging to no level of a classification designed for a dense network, which requires adapting the reference to the territory rather than the reverse.
What this subject asks of the client
Four pieces of information come from the client rather than from satellite imagery.
The exact sense of each one of the levels in their road hierarchy.
The fate given to segments that are private or whose access is restricted.
The distinction expected between a carriageway and a parking area.
And the network’s final use, which entirely commands the deliverable’s structure.
One important observation follows for a satellite imagery project. The second piece escapes observation entirely, a private way serving a holding resembling a public way in every respect, which makes this distinction impossible to produce without a tenure source or an arbitrary rule accepted as such.
What temporality adds here
Four satellite imagery contributions follow from repeated acquisitions over a network.
A new way is dated precisely between two successive satellite passes.
A linear worksite is followed throughout the whole of its progress.
A segment masked during the summer sometimes shows in winter.
And a lasting break is finally distinguished from a passing masking.
One important practical consequence follows for a satellite imagery project. The third contribution resolves by method what resolution did not resolve, a country road invisible beneath leafy canopy appearing on a winter acquisition, which makes a second date sometimes more useful than a finer image.
Three errors proper to this subject
Three satellite imagery defects appear only on linear objects.
A single break that makes an otherwise accurate network unusable.
An intersection modelled as a crossroads when it is in fact a bridge.
And a segment wholly invented from a hedge or a straight ditch.
Those three defects escape a control by sample of tiles, they are detected only on the reconstituted graph, and their common point is to bear on the relation between objects rather than on the objects themselves.
What the provider brings here
Four contributions distinguish a network engagement conducted on satellite imagery.
An inference rule beneath obstacles written well before production.
A connectivity control conducted on the fully reconstituted mosaic.
Deliberate coverage of the most difficult masking situations.
And a deliverable structured as a graph rather than as a mere collection of lines.
One important practical consequence follows. The last contribution determines the possible use of the result, a mere set of unconnected segments obliging the client to reconstitute the whole topology themselves, specialised work that often exceeds what they expected to do.
Approaching an infrastructure project
Five questions scope such a satellite imagery project.
Is a centreline or a footprint required. The corpora are incompatible.
What happens beneath a masking obstacle. The inference must be settled.
Do unsurfaced tracks fall within scope. They are contentious.
Is connectivity measured. Proportion alone misleads.
And does passability matter. The image does not show it.
Those five answers determine the result’s usefulness. Asking them before production avoids a network that is visually complete and unfit for route calculation.
The question that frames the project
One question determines the nature of the deliverable in satellite imagery.
Will the network serve for travelling or for measuring.
Travelling, route calculation or accessibility, requires a connected graph, correctly modelled intersections and continuity restored beneath the obstacles.
Measuring, sealed surface or land footprint, requires exact carriageway outlines and dispenses entirely with the notion of connection.
That question belongs to the use and not to technique, it is asked before the first tile, and it separates two corpora no automatic conversion carries from one to the other.
Three decisions before producing
Three decisions commit a network project in satellite imagery.
Choosing between centreline and surface footprint, the corpora being incompatible.
Writing the continuity rule beneath an obstacle, which permits or forbids inference.
And fixing the inclusion threshold for unsurfaced tracks, the main source of disagreement.
Those three decisions cost one meeting, they precede the first tile, and their absence produces a network that is visually complete and unfit for route calculation.
Three checks on a network corpus
Three checks qualify a network dataset in satellite imagery.
The geometry finally adopted, centreline or carriageway footprint.
The connectivity measurement, quite distinct from the mere proportion detected.
And the actual presence of masking situations 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 resulting network will break at the most useful places.
What this chapter teaches
One cross-cutting observation deserves closing this examination.
A network is worth what its continuity is worth rather than its completeness.
Three findings compose it.
A centreline and a surface footprint constitute two incompatible corpora.
A break of a few metres cuts a whole route.
And a movement trace proves a passability the image never shows.
That finding extends the preceding chapters, the value of an annotation depending here on a global property rather than on the correctness of each object.
What this chapter leaves to the next
A network cut by an earthquake is mapped against the clock.
Three questions stay open in satellite imagery.
How to work when delay outweighs exhaustiveness.
What a before-and-after comparison presupposes by way of preparation.
And how to grade damage whose severity is decided rather than observed.
Those three questions belong to natural disasters, which constitute the subject of the following chapter.
Why connectivity belongs in the acceptance terms
One clause is worth writing into the contract on this kind of project.
Acceptance should test the graph, not a sample of tiles.
Three reasons follow in satellite imagery.
A network can pass every tile-level check and still route nowhere.
The client discovers this only when they first try to use it.
And by then the batches have been signed off one by one.
One important practical consequence follows for a satellite imagery provider. Proposing this test protects the provider more than the client, since a connectivity criterion agreed in advance turns a vague dissatisfaction at the end into a measurable condition met or not met, and it rules out the far worse outcome of delivering work that is accurate everywhere and useful nowhere.
Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.
- Confusing centreline with surface footprint in the reference.
- Measuring a proportion of length without measuring connectivity.
- Leaving implicit the continuity rule beneath an obstacle.
- Including unsurfaced tracks without defining their detection threshold.
- Annotating a power line as a visible linear object.
- Confusing a visible alignment with a passable way.
- Neglecting the treatment of intersections and interchanges.
- Comparing with a collaborative base without assessing its coverage.
- Delivering a network without checking consistency at tile boundaries.
- Treating a bridge as a break in the network.
What to take away
Network mapping is judged on a property annotation by object does not produce.
Three readings emerge. A centreline and a surface footprint constitute two incompatible corpora, one serving route calculation and the other calculation of sealed surface, which forbids converting one into the other afterwards and requires settling at scoping. A break of a few metres cuts a whole route, which makes a network detected at ninety-nine percent sometimes unusable and makes connectivity a more revealing indicator than the proportion detected. And a movement trace proves a passability the image never shows, a distinction that separates a usable road from a merely visible alignment.
For emergencies, the article on natural disasters details the approach. For the third dimension, the article on airborne LiDAR 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 network mapping project, let us discuss your need.
