Road access and proximity analysis for land valuation
Satellite imagery enables road network extraction and travel-time modelling for parcels where official datasets are absent or years out of date, turning accessibility into a quantifiable input for land valuation.
Sensors
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral. At this resolution individual lane markings and unpaved track edges are resolvable, making it the primary source for road centreline extraction in high-value parcels where precision matters.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral, with tri-stereo capability. Stereo acquisition adds a surface-model layer that helps distinguish road embankments from drainage channels, a common source of error in flat-terrain segmentation.
- Planet SuperDove: 3–4 m multispectral, daily revisit across most latitudes. Coarser than WorldView-3 but sufficient for rural track detection over large areas, and the daily cadence lets analysts track seasonal track appearance (dry-season access routes that vanish in the wet).
- Sentinel-2 MSI: 10 m visible and near-infrared bands, 5-day revisit at the equator. Too coarse for individual track centrelines but useful for generating the land-cover inputs (bare soil, vegetation, water) that constrain travel-speed assumptions in isochrone models.
Why official road data fails peri-urban valuation
In mature urban markets, road datasets from national mapping agencies are good enough. Valuers use them without thinking. The problem emerges at the fringe: peri-urban and rural areas where cadastral surveys were last updated a decade or more ago, where informal tracks have become de facto arterials, and where new sealed roads cut travel times that no database yet reflects.
A parcel that official data places 45 minutes from a market town may, on the ground, sit 12 minutes away via a recently graded district road. The valuation gap that creates is not a rounding error. It is the difference between agricultural and residential pricing. Satellite imagery reads the current ground truth, not the archived one.
What a convolutional network finds in a road image
Road extraction from high-resolution optical imagery is now a mature applied-research field. Convolutional segmentation models, trained on labelled datasets such as the Massachusetts Roads Dataset and the DeepGlobe Road Extraction Benchmark, assign each pixel a probability of belonging to a road surface. Post-processing skeletonises those probability masks into centreline vectors with associated width estimates.
At WorldView-3 resolution (31 cm pan), the method reliably detects sealed roads and well-used unpaved tracks down to roughly 2–3 metres wide. Faint footpaths and single-vehicle tracks through dense canopy remain genuinely difficult: canopy occlusion hides the surface, and the spectral signature of compacted earth blends with bare agricultural soil. That limit is real and should be priced into any commission. Pléiades Neo stereo adds a normalised surface model that partially compensates, because a road embankment has a characteristic cross-sectional shape even when the surface is obscured.
Planet SuperDove's 3–4 m pixels cannot resolve individual tracks but do detect the linear bare-soil signatures of unpaved roads against vegetated backgrounds, particularly in the red-edge band (band 6, approximately 702 nm). This makes SuperDove useful for large-area screening before a targeted WorldView-3 or Pléiades tasking order.
From centrelines to travel-time surfaces
A road centreline layer is only the first step. What a valuer actually needs is a travel-time surface: a raster where each cell records the minimum travel time to a specified destination (a town centre, a port, a school) across the road network and off-road terrain.
The standard method is least-cost path analysis on a friction surface. Each road segment is assigned a travel speed based on surface type (sealed, graded gravel, unimproved track) and gradient derived from a digital elevation model such as the Copernicus DEM at 30 m or SRTM. Off-road cells receive much higher friction values. The result is an isochrone surface, commonly expressed in 5- or 10-minute bands, that can be intersected with any parcel boundary to produce a single accessibility score per parcel.
Speed assumptions matter enormously. A sealed road in good condition might carry 80 km/h; a wet-season track in clay soil might fall to 15 km/h or become impassable. Sentinel-2's 5-day revisit is useful here: multi-temporal composites reveal which tracks show bare-soil exposure only in the dry season, flagging them as seasonally constrained. That seasonal qualifier is a legitimate valuation input in agricultural and rural markets.
Quantifying the accessibility premium
Once each parcel in a study area carries a travel-time score, the figure can enter a hedonic regression alongside other parcel attributes. The published academic literature on hedonic land pricing consistently finds that travel time to urban centres is among the strongest predictors of rural and peri-urban land values, with effect sizes varying by market but often in the range of 1–4% per additional minute of travel time to a primary centre. Those figures come from published studies; local calibration against transaction data is always required.
The satellite-derived accessibility score is most valuable precisely where transaction data is thin, because it provides a continuous, spatially consistent variable across parcels that have never transacted. Interpolating value across a sparse transaction dataset without a spatial accessibility variable produces large errors; with one, the model has a physical anchor.
A practical limit: the method produces a network-based travel time, not a perceived accessibility score. Road quality perceptions, seasonal flooding, and local knowledge of dangerous sections are not captured by spectral data. Ground-truth validation on a sample of roads is good practice before committing the output to a formal valuation model.
Archive depth and change detection
One underused capability is longitudinal comparison. Planet's archive extends to 2016 for most areas; Sentinel-2 to mid-2015; WorldView-3 to late 2014. That span is long enough to reconstruct when a road was sealed, when a new junction was cut, or when a track was upgraded to a district road. For a valuer assessing whether an access improvement is priced into current comparables, the archive answers the question directly.
Satellize runs road-change detection as part of its broader land analytics work. The Tonga crop-estimation programme demonstrated the team's approach to multi-temporal feature extraction in data-sparse environments, a methodology that transfers directly to road network monitoring in markets where ground surveys are infrequent.
Change detection at Planet's 3–4 m resolution can flag new linear bare-soil features within a few weeks of their appearance. A WorldView-3 or Pléiades follow-up task then confirms whether the feature is a road, a pipeline trench, or a drainage channel. Combining both sensors keeps tasking costs proportionate to the area under analysis.
What this method cannot do
Cloud cover is the persistent enemy of optical road extraction. In tropical markets with long wet seasons, consistent cloud-free imagery may require compositing across several months, during which the road network may have changed. SAR (Synthetic Aperture Radar) from Sentinel-1 penetrates cloud but at 5–20 m resolution it detects only major roads and performs poorly on unpaved tracks. There is no clean workaround; the honest answer is that wet-season road mapping in persistently cloudy regions carries higher uncertainty and longer lead times.
The method also does not assess road legal status. A track visible in imagery may be a private farm road, a right-of-way under dispute, or a squatter route with no formal access rights. Satellite data establishes physical existence, not legal accessibility. Title search and local legal due diligence remain necessary. The satellite output is an input to valuation, not a substitute for it.
Typical figures
| Best spatial resolution (road extraction) | 30–31 cm (Maxar WorldView-3 pan; Airbus Pléiades Neo pan) |
| Minimum detectable road width | ~2–3 m on WorldView-3; ~4–6 m on Planet SuperDove; ~10–15 m on Sentinel-2 |
| Revisit frequency | Daily (Planet SuperDove); 5-day (Sentinel-2 at equator); tasked on demand (WorldView-3, Pléiades Neo) |
| Spectral bands used | Panchromatic, RGB, NIR, red-edge (~702 nm on SuperDove); SWIR optional for surface-type discrimination |
| Elevation data for friction surface | Copernicus DEM GLO-30 (30 m) or SRTM (30 m); Pléiades Neo stereo DSM where commissioned |
| Archive depth | Sentinel-2 from mid-2015; Planet from 2016; WorldView-3 from late 2014 |
| Travel-time surface resolution | Typically 10–30 m raster cells, constrained by DEM and land-cover input resolution |
| Delivery formats | GeoJSON / Shapefile centrelines; GeoTIFF isochrone rasters; CSV parcel-level accessibility scores; PDF valuation input report |
| Cloud limitation | Optical only; persistent cloud requires multi-month compositing or SAR fallback with reduced road-type discrimination |
Analytics Satellize can run
| Road centreline vector layer | Convolutional semantic segmentation (U-Net class architecture) applied to high-resolution optical imagery, followed by morphological skeletonisation | GeoJSON or Shapefile of extracted road centrelines with surface-type classification (sealed / gravel / unimproved track) |
| Travel-time isochrone surface | Least-cost path analysis on a friction raster combining road speed assignments, off-road terrain friction and Copernicus DEM slope | GeoTIFF raster in 5- or 10-minute isochrone bands, clipped to study-area extent |
| Parcel-level accessibility score | Zonal statistics intersecting isochrone raster with parcel boundary polygons to extract minimum and mean travel time per parcel | CSV or GeoPackage attribute table joinable to client cadastral data |
| Seasonal access classification | Multi-temporal Sentinel-2 compositing to identify tracks with bare-soil exposure only in dry months, flagging wet-season impassability | Attributed road layer with seasonal-access flag; summary table by parcel |
| Road network change report | Bi-temporal or multi-temporal change detection comparing Planet or Sentinel-2 composites across user-defined date ranges to identify new or upgraded road segments | PDF report with imagery chips and GIS layer of changed segments, timestamped to within the archive's revisit cadence |
| Accessibility premium input for hedonic model | Regression-ready feature engineering: travel-time score, distance to nearest sealed road, road density within 1 km buffer, all derived from centreline and isochrone outputs | Structured CSV with one row per parcel, ready for integration into client valuation or AVM model |
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.