Parking lot occupancy analysis for retail land valuation
Sub-metre commercial satellites count vehicles in surface car parks, producing a footfall proxy that bypasses tenant-reported figures. Temporal stacking across seasons reveals peak-load patterns that ground-truth retail land valuations.
Sensors
- Planet SkySat: Delivers 50 cm pan-sharpened imagery, sufficient to distinguish a saloon car from a van. Tasking latency is typically same-day to next-day. Revisit at a given site is flexible under tasking contracts, making it practical to acquire morning, midday and afternoon passes on the same date.
- Maxar WorldView-3: 30 cm panchromatic, 1.24 m multispectral. The sharpest commercially available optical archive at scale. WorldView-3's SWIR bands also distinguish wet tarmac from dry, which matters when separating genuine low-occupancy days from cloud-shadow artefacts.
- Airbus Pléiades Neo: 30 cm native resolution, four-band optical. A 27-satellite tasking constellation (Pléiades Neo plus legacy Pléiades 1A/1B) gives daily revisit over most mid-latitude retail corridors. Archive stretches back to 2011 for the legacy pair.
- Planet PlanetScope: 3 m resolution, daily global revisit. Individual cars are not reliably distinguishable at this resolution, but the sensor is useful for large surface car parks where total dark-pixel area (vehicles) versus light-pixel area (empty tarmac) can be estimated statistically, and for flagging acquisition dates worth revisiting with a tasked sub-metre asset.
Why car parks tell the truth when tenants do not
Retail landlords and their valuers have long relied on tenant-reported footfall figures: door-count sensors, loyalty-card transactions, or simply the tenant's own trading statements. All of these are controlled by the tenant and disclosed selectively. A car park has no such filter. The number of vehicles present at a given moment is a direct, observable expression of demand, visible to any satellite with sufficient resolution to distinguish a 4.5-metre vehicle from the 2.5-metre gap beside it.
This is not a new insight in principle, but it became practically tractable only when sub-metre commercial constellations achieved reliable, on-demand tasking. Before that, analysts either relied on analyst-driven manual counts from expensive archive imagery or used coarser sensors that could not separate vehicles from background. The combination of 30 to 50 cm resolution and same-day tasking changes the economics enough to make systematic, multi-site monitoring feasible.
What a floating roof gives away
Surface car parks have one feature that makes automated vehicle detection relatively tractable: the background is uniform. Tarmac has a consistent spectral signature in visible and near-infrared bands, and vehicles introduce a mixture of high-reflectance (light-coloured roofs, windscreens) and low-reflectance (dark bodywork, shadow beneath the vehicle) pixels that stand out from it. Published detection pipelines, including work documented in the Remote Sensing journal, typically combine a shadow-detection step with a roof-highlight step, then apply morphological filtering to match the expected aspect ratio of a parked vehicle.
WorldView-3's 30 cm panchromatic band is the reference standard for this task. At that resolution, a standard UK parking bay (2.4 m wide, 4.8 m long) occupies roughly 8 by 16 pixels, enough to apply template matching or a convolutional detector trained on labelled satellite imagery. SkySat at 50 cm gives bays of roughly 5 by 10 pixels, which is workable but begins to lose reliability for compact cars in tight bays. Below 1 m, detection accuracy degrades sharply for anything smaller than a light commercial vehicle.
Honest caveat: covered multi-storey car parks are invisible to optical sensors. Sites where a significant share of parking is structured rather than surface-level will undercount demand. Analysts need a site survey or planning data to flag this before drawing conclusions.
Temporal stacking: one image is anecdote, twelve is evidence
A single acquisition tells you occupancy on one day at one time. That is useful for due diligence on a specific transaction but insufficient for a valuation that must hold across market cycles. The methodology becomes genuinely powerful when acquisitions are stacked across multiple dates, times of day, and seasons.
A practical programme might acquire imagery at three times of day (late morning, midday, mid-afternoon) on one weekday and one weekend day per month across a 12-month period. That yields 72 observations per site per year. From this stack you can extract peak occupancy (the 95th-percentile count), typical trading-day occupancy, seasonal index (December versus August for a UK retail park), and the ratio of peak to average, which is a rough proxy for how efficiently the car park is being used relative to its design capacity.
Seasonal patterns are particularly valuable for distinguishing structurally strong retail assets from those propped up by a single anchor tenant's promotional calendar. A car park that fills only in the four weeks before Christmas is a different asset from one that sustains 70 per cent occupancy from March through October.
Translating vehicle counts into valuation inputs
The output of the detection pipeline is a time-series of vehicle counts per site. Converting that into a valuation input requires a calibration step. The most common approach is to establish a relationship between observed vehicle count and known footfall at a reference site where both ground-truth door-count data and satellite imagery are available, then apply that relationship to sites where only satellite data exists.
The ratio of car park spaces to observed peak occupancy also informs yield assumptions. A retail park where the car park saturates by 11 a.m. on a Saturday is likely turning away demand; one where peak occupancy reaches only 40 per cent may be over-parked relative to its actual draw. Both observations affect the capitalisation rate a prudent valuer should apply.
A further use is competitive benchmarking. If a subject property's car park occupancy is declining relative to a competing centre two kilometres away, that trend is visible in the satellite record before it appears in rental evidence or void rates. Satellize applies this kind of comparative temporal analysis as part of its analytics work; the methodology is the same one used in its Tonga crop-estimation programme, where relative change across sites matters more than any single absolute figure.
Limits the methodology cannot escape
Cloud cover is the unavoidable constraint on optical methods. The UK averages more than 50 per cent cloud cover across the year; a tasked acquisition has a meaningful probability of returning a cloudy scene. Mitigation involves scheduling multiple acquisition attempts per target date and accepting that some months will have gaps. Analysts should report cloud-affected gaps explicitly rather than interpolating through them.
Time-of-day ambiguity matters. A satellite pass at 10:30 local time captures a different trading moment than one at 14:00. Unless the acquisition time is recorded and controlled, comparisons across dates can conflate genuine occupancy change with time-of-day variation. Responsible analysis always logs the UTC acquisition time and corrects for it.
Finally, car parks are not footfall. A site serving a large catchment with high car dependency will show strong vehicle counts even if pedestrian footfall is modest. Sites in dense urban areas with high public-transport use will show weak vehicle counts even if trading is strong. The method works best for out-of-town and edge-of-town retail formats where car is the dominant access mode, which in the UK and much of continental Europe covers the majority of large-format retail land by value.
Typical figures
| Spatial resolution (primary sensors) | 30 cm (WorldView-3 pan, Pléiades Neo); 50 cm (SkySat pan-sharpened) |
| Minimum reliably detectable vehicle | Standard passenger car at 30–50 cm resolution; light commercial vehicles more reliable than compact cars at 50 cm |
| Revisit / tasking cadence | Daily revisit achievable with Pléiades Neo constellation; SkySat same-day to next-day tasking; WorldView-3 typically 1–4 days depending on latitude and cloud |
| Spectral bands used | Panchromatic (primary detection); multispectral visible + NIR (shadow and tarmac discrimination); WorldView-3 SWIR (wet-surface correction) |
| Archive depth | WorldView archive from 2007; Pléiades legacy from 2011; SkySat from approximately 2014; coverage of specific retail sites varies by prior tasking history |
| Cloud cover constraint | Optical only; scenes with >20% cloud over the car park polygon are typically rejected; multiple acquisition attempts per target date recommended in mid-latitude climates |
| Typical detection accuracy | Published studies report 85–95% precision and recall for vehicle detection at 30 cm in well-lit, cloud-free conditions; accuracy degrades at 50 cm for compact vehicles |
| Deliverable latency | 24–72 hours from tasked acquisition to processed vehicle-count output, depending on pipeline configuration |
| Coverage per acquisition | SkySat scene: approximately 16 km²; WorldView-3 strip: up to 110 km × variable; Pléiades Neo: 14 km × 14 km standard swath |
| Delivery formats | GeoTIFF annotated with detected vehicle bounding boxes; CSV time-series of counts per site; GeoPackage or Shapefile for GIS integration |
Analytics Satellize can run
| Vehicle count time-series per site | Template matching or CNN-based object detection on pan-sharpened imagery, with shadow-highlight preprocessing; method class published extensively in Remote Sensing journal literature | CSV and GeoJSON time-series with acquisition timestamp, count, confidence interval, and cloud-flag per observation |
| Peak and typical occupancy indices | Percentile statistics across the temporal stack (95th percentile for peak, median for typical trading day); seasonal decomposition using standard time-series methods | Site-level summary report with peak, median, and seasonal index figures; suitable for insertion into valuation models |
| Occupancy heatmap per car park | Spatial aggregation of detected vehicle centroids into a grid; kernel density estimation to show which bays and zones fill first | GeoTIFF heatmap layer per acquisition date, stackable in GIS; useful for identifying access-point demand and dead zones |
| Competitive benchmarking dashboard | Parallel extraction across a defined set of competing or comparable sites; relative occupancy ratio calculated against a user-defined reference site | Multi-site comparison table and chart pack, updated on each new acquisition cycle |
| Trend and inflection-point detection | Rolling 13-week occupancy trend fitted to the time-series; change-point detection to flag statistically significant shifts in demand level | Alert report when a site's occupancy trend crosses a user-defined threshold; delivered as email summary or feed to client data platform |
| Structured parking adjustment flag | Cross-reference of surface car park capacity (from planning data or site survey) against total site parking provision; flags sites where multi-storey or basement parking materially affects the surface-count interpretation | Per-site metadata flag in the deliverable GeoJSON; narrative note in valuation report |
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.