Urban road network extraction from high-resolution imagery
Automated extraction of road centrelines, carriageway widths and junction geometry from sub-metre optical imagery closes the gap between what governments think their road network looks like and what is actually on the ground.
Sensors
- Maxar WorldView Legion: 30 cm panchromatic, 1.2 m multispectral. Up to 15 revisits per day over any target city, which matters when you need to track road construction progress rather than just a snapshot.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral in four bands plus a red-edge channel. Tri-stereo tasking allows a digital surface model to be derived alongside the orthoimage, useful for estimating overpass geometry and distinguishing elevated roads from ground-level lanes.
- Planet SkySat: 50 cm panchromatic, 1 m multispectral. Lower radiometric dynamic range than WorldView or Pléiades Neo, but frequent tasking and competitive archive depth make it practical for change detection across a road construction programme spanning months.
- ESRI Wayback / historical basemap mosaics: Not a satellite sensor but a dated archive of commercial imagery mosaics. Useful for establishing what the road network looked like at a prior legal or administrative moment, giving a baseline against which new extractions can be compared.
What the register does not know about its own roads
In most cities outside the wealthiest OECD countries, the official road register was last comprehensively surveyed by a ground team years or decades ago. New peri-urban subdivisions, informal lane consolidations and post-disaster diversions accumulate faster than any cadastral agency can resurvey. The practical consequence is that emergency routing software, utility permitting databases and transport-demand models all run on stale geometry. A road that was widened from four metres to eight metres five years ago is still recorded as four metres. A junction that was signalised and reconfigured is still a simple cross in the GIS.
Sub-metre optical imagery from commercial constellations makes it possible to extract a current road network across an entire city in days rather than years. The extraction is not perfect, and the limits matter, but even an imperfect satellite-derived network is usually more accurate than a register that has not been touched since the last ground survey.
How the extraction pipeline works
The standard approach combines a convolutional neural network (CNN) trained on labelled road pixels with post-processing steps that convert a raster probability map into clean vector geometry. The CNN stage, drawing on architectures such as D-LinkNet or DeepRoadMapper that have been published and benchmarked on datasets like the Massachusetts Roads Dataset and the DeepGlobe Road Extraction Challenge, produces a per-pixel probability of road surface. Thresholding that map and applying morphological thinning yields a binary skeleton. Centreline vectorisation then traces connected paths through the skeleton, with junction nodes placed where branches meet.
Carriageway width is estimated by measuring the perpendicular extent of the road probability mask at regular intervals along each centreline segment. At 30 cm resolution, a standard two-lane road of roughly seven metres is about 23 pixels wide, which is enough to distinguish it from a single-lane track. Width estimates carry an uncertainty of roughly one to two pixels, meaning an absolute error of 30 to 60 cm under good conditions. That is adequate for permitting and planning; it is not adequate for precision engineering design.
Junction geometry, including turning radii and channelisation islands, can be extracted from the same imagery when the network is paved and unoccluded. Unpaved junctions in informal areas are harder: the boundary between carriageway and verge is spectrally ambiguous, and the CNN tends to under-segment the junction area.
Where the method breaks down
Tree canopy is the most consistent source of error. In cities with mature street trees, the canopy obscures the road surface beneath. A CNN trained on open-sky imagery will simply not see the road under a dense avenue of trees. The gap can be partially bridged by using near-infrared bands to detect the canopy and then interpolating the centreline through the occlusion, but the carriageway width estimate under the canopy is unreliable. Stereo-derived surface models help here: a road-level surface running beneath a canopy layer is detectable in a DSM even when the nadir image shows only leaves.
Dense informal grids present a different problem. Where plot boundaries, footpaths, drainage channels and narrow vehicle lanes all run at similar widths and similar spectral signatures, the CNN cannot reliably distinguish a two-metre footpath from a three-metre vehicle lane. Published benchmarks on informal settlement imagery show F1 scores for road extraction dropping from around 0.85 in formal grid areas to 0.60 to 0.70 in dense informal fabric. That is an honest range: some informal networks extract cleanly, others do not. Ground truthing a sample of the output is not optional in these areas.
Cloud cover is a straightforward constraint. Tropical cities in the wet season can go weeks without a usable acquisition from any optical sensor. Radar imagery from Sentinel-1 can detect the linear structure of roads through cloud, but at 10 m resolution it cannot resolve carriageway width or junction geometry. It is useful for confirming that a major road exists; it is not useful for the detailed extraction this page is about.
Connecting extracted geometry to real administrative decisions
The output of an extraction pipeline is only as useful as its integration with downstream systems. For emergency routing, the extracted centreline network needs to be attributed with passability estimates, which requires combining the geometry with carriageway width, surface-type classification (paved versus unpaved, derived from spectral reflectance) and any known flood-risk overlay. For utility permitting, the critical output is the road reserve boundary, which requires knowing not just the centreline but the full carriageway extent and the building setback on each side.
Transport modelling needs accurate junction counts and turn restrictions. Satellite extraction can provide junction locations and approximate geometry, but turn restrictions are not visible from above. A satellite-derived network is a starting point for a transport model, not a finished input. Planners who treat it as finished will get wrong answers. The honest use case is to replace a badly outdated register with a geometrically current one, then apply field verification selectively to the junctions and segments that matter most for the model.
Archive depth and change detection
One underused application is comparing extractions from different dates to detect road construction, widening or closure. Maxar's archive extends back to 1999 for some areas; Pléiades archive coverage for many cities starts around 2012. By running the same extraction pipeline on imagery from two or more dates and computing the symmetric difference between the resulting networks, it is possible to identify new road segments, widened carriageways and removed roads. This is directly relevant to land administration: a road that appears in a satellite-derived network but not in the official register is a candidate for formal gazetting, and the imagery provides dated evidence of when it was constructed.
Satellize runs road-network extraction as part of its urban analytics suite, using commercial tasking on client licence alongside open-data sources. The Tonga crop-estimation programme is the company's most publicly discussed analytics engagement, but the same pipeline architecture, CNN inference followed by topological vectorisation, transfers directly to road extraction in any city where sub-metre tasking is feasible.
Accuracy expectations a buyer should insist on
Published benchmarks on the DeepGlobe and SpaceNet Road Detection datasets report completeness (recall) and correctness (precision) for road extraction typically in the 0.75 to 0.90 range for formal urban areas at 30 cm resolution. These numbers are for well-lit, low-canopy, paved networks. They are not representative of every city. Any credible delivery should include an accuracy assessment on a held-out sample of the target city, not just a citation of benchmark performance on a different dataset in a different geography.
Minimum detectable road width is approximately two to three metres at 30 cm resolution, corresponding to roughly seven to ten pixels across the carriageway. Below that width, the road probability mask becomes unreliable and centreline extraction degrades. Footpaths and very narrow lanes in informal settlements fall below this threshold and should not be expected in the output without manual digitising.
Typical figures
| Best available spatial resolution | 30 cm panchromatic (WorldView Legion, Pléiades Neo) |
| Multispectral resolution | 1.2 m (WorldView Legion, Pléiades Neo); 1 m (SkySat) |
| Revisit frequency | Up to 15 times per day (WorldView Legion); 1 to 2 times per day (Pléiades Neo constellation); flexible tasking (SkySat) |
| Minimum detectable road width | Approximately 2 to 3 m at 30 cm resolution; informal footpaths below this threshold are not reliably extracted |
| Carriageway width accuracy | ±30 to 60 cm under clear, unoccluded conditions; unreliable under tree canopy |
| Spectral bands used | Panchromatic for geometry; RGB and near-infrared for surface-type classification and canopy detection |
| Archive depth | Maxar: from 1999 for some areas; Pléiades: from approximately 2012; SkySat: from approximately 2017 |
| Typical processing latency | 24 to 72 hours from tasked acquisition to vectorised network delivery, depending on area and cloud conditions |
| Delivery formats | GeoPackage, Shapefile, GeoJSON; attributed with carriageway width, surface type and confidence score per segment |
| Cloud cover constraint | Optical sensors require clear sky; wet-season tropical cities may face multi-week gaps in usable acquisitions |
Analytics Satellize can run
| Road centreline network | CNN-based road segmentation (D-LinkNet or equivalent architecture) followed by morphological thinning and vectorisation | GIS layer of attributed centrelines with segment IDs, surface type and confidence score |
| Carriageway width profile | Perpendicular measurement of road probability mask at 10 m intervals along each centreline segment | Width attribute table joined to centreline GIS layer; flagged segments where canopy occlusion makes width unreliable |
| Junction geometry extraction | Topological node detection at branch intersections in the vectorised skeleton; approximate turning-radius estimation from mask boundary curvature | Point layer of junction nodes with approach-count attribute and approximate geometry polygons for major junctions |
| Road surface classification | Spectral reflectance analysis in RGB and NIR bands to distinguish paved asphalt, concrete, gravel and unpaved earth surfaces | Surface-type attribute on each centreline segment; accuracy assessment report on held-out sample |
| Network change detection | Symmetric difference between extractions from two dated acquisitions; new, removed and widened segments flagged | Change layer with date-stamped new and modified segments; summary table of total new road length by administrative zone |
| Register gap analysis | Spatial comparison of satellite-derived network against official road register geometry; unmatched segments identified | Gap report with candidate unregistered roads, supporting imagery thumbnails and area statistics by district |
| Accuracy assessment | Completeness and correctness scoring against manually digitised reference sample drawn from the target city, not a generic benchmark dataset | Accuracy report with per-class F1 scores, confusion matrix and list of systematic error sources in the target geography |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.