Vacant land and derelict site detection within built areas
Spectral indices, SAR coherence and multi-year inactivity signatures identify vacant and derelict parcels inside built-up areas, supporting brownfield registers and compulsory purchase intelligence with honest caveats about parks and gardens.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at mid-latitudes with two satellites. Provides NDVI, NDBI and bare-soil indices for multi-year inactivity signatures. Free and open archive from 2015.
- Sentinel-1 SAR (C-band): IW mode ground range detected at 10 m, 6-day repeat with both satellites. Coherence computed over 6- or 12-day pairs distinguishes stable bare ground from disturbed surfaces; cloud-transparent. Archive from 2014.
- Planet SuperDove: 3–4 m resolution, 8-band PlanetScope imagery, near-daily revisit over most cities. Resolves individual plot boundaries and detects vehicle presence or absence at parcel scale. Commercial licence required.
- Maxar WorldView-3: 30 cm panchromatic, 1.24 m multispectral, 16 SWIR bands. Confirms surface condition, vegetation type and structural decay detail on priority parcels. Revisit typically 1–4.5 days depending on collection angle tolerance.
What a persistent NDVI anomaly inside a city block is actually saying
The Normalised Difference Vegetation Index is a ratio of near-infrared reflectance to red reflectance. On Sentinel-2, that calculation runs at 10 m. The interesting signal for brownfield detection is not high NDVI per se, it is high NDVI that is spatially incongruous and temporally persistent within a predominantly impervious neighbourhood. A manicured park behaves differently from an abandoned plot: parks show regular mowing cycles as seasonal NDVI dips, maintained edge geometry and consistent spectral texture. Derelict land tends toward rough, heterogeneous vegetation with advancing scrub, variable soil exposure and no evidence of mechanical intervention across multiple growing seasons.
Stacking three or more years of Sentinel-2 composites and computing the coefficient of variation in NDVI across that period adds a useful discriminator. Genuinely abandoned sites show moderate temporal variance driven by weed succession, but the variance pattern differs from cultivated green space. Adding a bare-soil index (NDBI or BSI) to the stack catches sites that cycle between exposed rubble and opportunistic vegetation, a signature common to partially cleared brownfield land. This is not a perfect classifier. Private gardens with infrequent maintenance can produce similar signatures, and the 10 m pixel blurs parcel boundaries in dense terraced streets.
SAR coherence as a vacancy clock
Synthetic aperture radar coherence measures how similar the phase of a surface's radar return is between two acquisitions separated by days or weeks. Bare, undisturbed ground maintains high coherence over a 6-day Sentinel-1 repeat pair. Active construction sites, parked vehicles moving between passes, and growing vegetation all decorrelate. The diagnostic value for vacant land detection lies in inverting that logic: a parcel that shows persistently high coherence across dozens of repeat pairs, with no coherence drops indicating construction activity, vehicle movement or soil disturbance, is likely genuinely inactive.
Multi-year coherence time-series distinguish two categories that optical indices often conflate. A temporarily cleared site awaiting development will show a coherence break at the clearance event, then stabilise. A long-vacant plot shows no such break; its coherence history is flat and uninterrupted. Processing 12-day coherence pairs across a two-year Sentinel-1 archive over a city typically requires on the order of 60 interferometric pairs, which is computationally tractable and produces a per-pixel inactivity score that maps directly onto cadastral parcel boundaries. The honest limit: SAR coherence at C-band is sensitive to soil moisture changes after rainfall, which can mimic disturbance signatures. Wet winters in maritime climates require careful seasonal windowing.
Where VHR imagery earns its licence fee
Sentinel and SAR analysis narrows a city-wide search to a candidate list. Planet SuperDove at 3–4 m then checks each candidate for vehicle presence, boundary condition and surface type at a scale where individual plot boundaries are legible. A parcel showing no vehicle activity across 50 or more daily passes over a six-month period is a strong vacancy signal. This is not the same as confirmation: underground car parks, delivery yards with infrequent use and private storage compounds can all appear quiet from above.
Maxar WorldView-3 at 30 cm panchromatic resolves structural decay detail that lower-resolution sensors cannot. Collapsed roof sections, broken perimeter fencing, vegetation growing through hard standing and the absence of any maintained surface are all visible at this resolution. Its 16-band SWIR capability also separates asbestos-containing roofing materials from standard fibre cement, which matters for remediation cost estimation on brownfield sites. The cost of tasking WorldView-3 means it is practical only on a shortlist of high-priority parcels, not as a city-wide screening layer.
The parks-and-gardens confusion problem, stated plainly
Any honest discussion of this technique must address the false-positive rate. Parks, allotments, private gardens, school playing fields, churchyards and golf courses all produce spectral and coherence signatures that partially overlap with derelict land. At Sentinel-2's 10 m resolution, a small urban garden is often sub-pixel or mixed with adjacent impervious surfaces. The classifier cannot reliably separate a neglected private garden from a genuinely vacant plot without additional context.
The practical mitigation is not purely algorithmic. Masking against existing land-use datasets, OpenStreetMap green-space polygons and local authority parks records before running the spectral analysis removes a substantial fraction of false positives. What remains is a candidate set, not a confirmed register. Ground verification, or at minimum a VHR image review, is required before any parcel enters a formal brownfield register or triggers a compulsory purchase assessment. Satellite analysis is best positioned as a triage tool that directs human attention efficiently, not as an autonomous classifier that produces legal-grade outputs.
From candidate list to brownfield register: the workflow
A practical city-scale workflow runs in three stages. First, a Sentinel-2 multi-year composite identifies spectrally anomalous parcels within a built-up area mask, filtered by existing land-use data to remove known green infrastructure. Second, Sentinel-1 coherence time-series scores each candidate for temporal inactivity. Parcels scoring high on both optical anomaly and coherence stability advance to stage three. Third, Planet SuperDove daily mosaics check for vehicle or construction activity over a defined observation window, typically 90 to 180 days, and WorldView-3 is tasked on the highest-priority shortlist for surface condition confirmation.
Output formats that serve planning authorities and property investors differ. A local authority building a brownfield land register typically needs a GIS polygon layer with confidence scores and supporting evidence images, referenced to cadastral parcel identifiers. A property investor or developer running compulsory purchase intelligence wants the same spatial data enriched with ownership information from land registry sources, remediation liability indicators and planning history. Satellize runs the spectral and SAR analytics on open constellation data and adds commercial tasking on client licence; the Tonga crop-estimation programme uses a structurally similar multi-source stacking approach, though the land-cover targets differ entirely. The analytic architecture transfers; the thresholds and masking layers need calibrating to each city's morphology and climate.
Honest limits, stated as specifications
Resolution sets a hard floor on parcel size. At Sentinel-2's 10 m, plots smaller than roughly 400 square metres are unreliable candidates because mixed pixels dominate. Planet SuperDove at 3–4 m pushes that floor down to approximately 50–100 square metres, depending on surrounding surface contrast. Sub-50-square-metre infill plots in dense urban fabric require WorldView-3 or equivalent VHR data to detect reliably.
Cloud cover is the other structural constraint. Optical sensors cannot see through cloud, and cities in northern Europe or monsoon-affected regions accumulate significant cloud-free data gaps. Sentinel-1 SAR is cloud-transparent and fills that gap for the coherence dimension, but cannot substitute for spectral classification. Multi-year compositing mitigates cloud for the optical layer, at the cost of temporal precision: a site that was vacant three years ago but has since been developed may still appear in a composite as a candidate. Revisit frequency and archive depth together determine how current and how historically resolved the inactivity signature can be.
Typical figures
| Optical spatial resolution | 10 m (Sentinel-2), 3–4 m (Planet SuperDove), 1.24 m multispectral / 30 cm pan (WorldView-3) |
| SAR spatial resolution | 10 m ground range detected, Sentinel-1 IW mode |
| Revisit cadence | 5 days optical (Sentinel-2, two satellites); 6 days SAR (Sentinel-1, two satellites); near-daily (Planet SuperDove); 1–4.5 days (WorldView-3, angle-dependent) |
| Minimum detectable parcel (optical) | ~400 m² at Sentinel-2; ~50–100 m² at Planet SuperDove; ~10 m² at WorldView-3 pan |
| Key spectral bands | Red, NIR (NDVI); SWIR (bare soil index, NDBI); 16-band SWIR (WorldView-3 material discrimination) |
| SAR coherence pair interval | 6-day or 12-day repeat pairs, Sentinel-1 C-band (5.405 GHz) |
| Archive depth | Sentinel-2 from 2015; Sentinel-1 from 2014; Planet SuperDove from approximately 2021 at full constellation density |
| Cloud limitation | Optical sensors fully blocked by cloud; multi-year compositing required in high-cloud-cover climates; SAR unaffected |
| Typical latency (operational product) | Sentinel data available within 1–3 hours of acquisition; Planet typically same-day; VHR tasking 1–5 days from order |
| Delivery formats | GeoTIFF raster layers, GeoJSON or Shapefile polygon outputs, confidence-scored parcel attribute tables, PDF site evidence reports |
Analytics Satellize can run
| City-wide vacancy candidate map | Multi-year Sentinel-2 NDVI and bare-soil index compositing with built-up area masking and green-space exclusion | GeoJSON polygon layer with per-parcel spectral anomaly score, delivered against client cadastral grid |
| Temporal inactivity score per parcel | Sentinel-1 C-band coherence time-series over 12-day repeat pairs, aggregated across 24-month window | Raster and parcel-level attribute table showing mean coherence, coherence variance and disturbance event count |
| Vehicle and construction activity check | Planet SuperDove daily mosaic differencing for vehicle presence and surface change over 90–180 day observation window | Per-parcel activity log with pass count, positive detection count and representative image chips |
| Surface condition report on priority shortlist | WorldView-3 30 cm pan and 1.24 m multispectral visual interpretation and SWIR material classification | PDF site evidence report per parcel with annotated imagery, surface condition rating and SWIR material flags |
| Brownfield register-ready GIS layer | Fusion of spectral, SAR and VHR outputs with confidence scoring and false-positive filtering against land-use reference data | Shapefile or GeoPackage with confidence tier (high / medium / unconfirmed), evidence summary and image attachments, formatted for local authority register ingestion |
| Compulsory purchase intelligence pack | Spatial join of satellite-derived vacancy evidence to cadastral parcel identifiers and planning history layers | Parcel-level report combining satellite evidence, inactivity duration estimate and remediation liability indicators |
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.