Retail unit vacancy detection from street-frontage facade change
Very-high-resolution optical satellites resolve individual shopfronts at 0.3–0.5 m, allowing change-detection algorithms to flag boarded, papered or darkened facades as vacancy signals weeks before a ground survey would catch them.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral (8 bands including coastal blue and SWIR), revisit roughly 1–4.5 days depending on latitude and tasking priority. The panchromatic channel resolves individual window bays on standard retail units wider than about 3 m.
- Airbus Pléiades Neo: 0.30 m panchromatic, 0.80 m multispectral (6 bands), tasking revisit 1–3 days at mid-latitudes. Radiometric consistency across the constellation of two satellites supports reliable reflectance comparisons between acquisition dates.
- Planet SkySat: 0.50 m pan-sharpened, revisit on demand with tasking (multiple times per day over priority areas). Slightly coarser than WorldView-3 or Pléiades Neo, but the high tasking cadence is useful for tracking rapid re-occupation or boarding events.
- Airbus Pléiades (classic): 0.50 m panchromatic, 2 m multispectral, revisit 1–3 days. Adequate for wider-frontage units on retail parks; marginal for narrow high-street units below roughly 4 m facade width.
What a boarded window actually looks like to a satellite
An active shopfront reflects a mixture of interior illumination, window-display materials and glazing. Plywood boarding, brown paper or whitewash each produce a distinct reflectance signature: plywood sits in the 0.35–0.55 reflectance range in the red and near-infrared bands, paper is spectrally flat and bright, and whitewash is uniformly high across the visible. All three suppress the slight specular glint that clean glazing produces under direct solar illumination. That glint suppression is detectable in multispectral imagery at 0.3–0.5 m resolution when the facade occupies at least four to six pixels across its width.
Texture is equally informative. Active window displays produce high local variance in pixel values because of colour blocks, lettering and product arrangements. A boarded or papered unit flattens that variance sharply. A simple Grey-Level Co-occurrence Matrix computed on the panchromatic band quantifies this difference. The combination of reflectance shift and texture collapse is what makes the signal specific: either alone produces false positives, but together they are a reliable discriminator in controlled conditions.
The change-detection workflow and its honest limits
The core method is bi-temporal change detection: a baseline image acquired when occupancy status is known is compared pixel-by-pixel against a later acquisition. Before comparison, both images must be orthorectified to sub-pixel accuracy, atmospherically corrected to surface reflectance, and co-registered to better than 0.5 pixels. At 0.3 m ground sampling distance, a co-registration error of one pixel translates to 0.3 m of spatial offset, which is enough to shift a facade edge into the adjacent pavement or interior, corrupting the signal entirely. This is the step most often underestimated in practice.
Cloud is the principal operational constraint. Unlike synthetic aperture radar, optical sensors at these resolutions cannot penetrate cloud cover. In northern European climates, usable clear-sky acquisitions over a given high street may arrive only once every two to six weeks in winter, depending on the constellation and tasking budget. That revisit gap is acceptable for monthly vacancy-rate reporting but too slow for same-week event detection.
Seasonal window displays are the most persistent source of false positives. A department store covering its windows in opaque seasonal graphics for a promotional period produces a reflectance and texture signature almost identical to boarding. Mitigation requires either a minimum persistence threshold (flagging only units that remain changed across two or more acquisitions separated by at least two weeks) or integration of ancillary signals such as business-registry data or footfall proxies. Neither approach eliminates the ambiguity entirely.
Minimum detectable unit and resolution floor
At 0.3 m panchromatic resolution, a facade must span at least 1.5 m in width to place five pixels across it, which is the practical minimum for texture analysis. Most UK high-street units are 4–8 m wide, comfortably above this floor. Narrow units in Victorian arcade formats, sometimes as little as 2 m wide, sit at the margin. Retail parks with large-format units of 10 m or more are the easiest targets and can be detected reliably even at SkySat's 0.5 m resolution.
Viewing angle matters more than resolution alone. A sensor acquiring at off-nadir angles above roughly 25–30 degrees will see the upper facade rather than the full frontage, and facade-level texture analysis degrades accordingly. Tasking requests for this application should specify near-nadir collection geometry where the constellation schedule permits.
From pixel flags to vacancy rate: the aggregation step
Individual unit flags are noisy. Aggregated across a defined retail zone, say a 400 m high-street segment or a named retail park, the noise averages out and the vacancy rate estimate becomes meaningful. Published academic work on VHR change detection in urban commercial zones (see the MDPI Remote Sensing journal archive) suggests that unit-level precision of around 75–85% and recall of 70–80% is achievable under good conditions, rising toward 90% when persistence filtering is applied. Those figures assume well-maintained baselines and clear-sky acquisitions; they degrade in cluttered urban canyons where shadow from adjacent buildings obscures the lower facade for part of the day.
The output that matters to a property valuer or lease negotiator is not a pixel map but a vacancy rate expressed as a percentage of units, updated on a defined cadence. A GIS layer showing unit-level flags, a summary table by zone, and a time-series chart of vacancy rate over twelve to twenty-four months are the three deliverables that translate the remote-sensing output into a form a property professional can use directly in a valuation model.
Where satellite data fits in the evidence chain
Satellite-derived vacancy rates are not a replacement for a ground survey. They are a screening tool that makes ground surveys cheaper by concentrating surveyor time on zones where the satellite signal has already identified elevated vacancy. They also provide temporal coverage that ground surveys cannot: a monthly satellite-derived rate across a portfolio of fifty retail parks is operationally and financially feasible in a way that fifty monthly ground surveys are not.
For lease negotiation, the most valuable output is often not the current vacancy rate but the trend. A high street that has moved from 8% to 22% vacancy over eighteen months, as measured from archive imagery, is a materially different negotiating context than one that has held at 12% throughout the same period. Archive depth matters here: WorldView-3 and Pléiades imagery archives extend back to 2014 and 2012 respectively, providing multi-year baselines without any new tasking cost.
Satellize structures this kind of longitudinal analysis as a standing analytics layer, updated on each usable acquisition, with outputs delivered as a GIS feed or structured report.
Typical figures
| Best available spatial resolution (pan) | 0.30 m (WorldView-3, Pléiades Neo) |
| Minimum facade width for reliable detection | ~1.5 m at 0.3 m GSD; ~4 m recommended for texture analysis |
| Tasking revisit (commercial VHR) | 1–4.5 days depending on sensor and latitude; cloud-limited to 1–6 usable acquisitions per month in temperate climates |
| Spectral bands used | Panchromatic (0.45–0.80 µm) for texture; multispectral visible and NIR for reflectance change; SWIR (WorldView-3) for material discrimination |
| Archive depth | WorldView-3 from 2014; Pléiades from 2012; SkySat from 2014 |
| Co-registration requirement | Sub-pixel (<0.5 pixel) for reliable bi-temporal comparison at 0.3 m GSD |
| Reported unit-level detection accuracy (published range) | 75–90% precision and recall under clear-sky, well-registered conditions |
| Cloud penetration | None. Optical only; SAR cannot resolve individual shopfronts at current commercial resolutions |
| Typical latency from acquisition to delivered output | 24–72 hours after clear-sky acquisition, depending on processing pipeline |
| Delivery formats | GeoJSON or Shapefile unit-level flags, CSV vacancy-rate summary, PDF zone report, time-series chart |
Analytics Satellize can run
| Unit-level vacancy flag layer | Bi-temporal reflectance and GLCM texture change detection on panchromatic and multispectral bands | GeoJSON or Shapefile with per-unit binary flag and confidence score, updated each usable acquisition |
| Zone vacancy rate time series | Spatial aggregation of unit flags across defined retail zones; persistence filtering to suppress seasonal display false positives | CSV and chart showing monthly vacancy rate per zone over the analysis period |
| Baseline vacancy assessment from archive | Retrospective change detection across archived VHR imagery (2012–present) to establish multi-year trend | PDF report with annotated imagery, zone-level vacancy history and trend classification |
| High-vacancy zone alert | Threshold trigger on zone vacancy rate exceeding a client-defined level across two consecutive acquisitions | Email or API alert with zone identifier, current rate, rate at prior period and annotated image chip |
| Re-occupation event detection | Reverse change detection flagging units that transition from inactive to active reflectance and texture state | GeoJSON update to the standing vacancy layer with re-occupation date and before/after image chips |
| Portfolio vacancy dashboard | Aggregation of zone-level outputs across a client-defined portfolio of retail assets | Structured GIS feed or web-accessible dashboard showing vacancy rate by asset, sortable by rate and trend direction |
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.