Retail and industrial activity proxied by parking-lot occupancy
Counting passenger vehicles in car parks from high-resolution optical imagery is a published, peer-reviewed method for estimating retail footfall and factory shift activity before official data are released.
Sensors
- Planet SuperDove (PlanetScope): 3 m native resolution, 8 spectral bands including red-edge and NIR. Daily global revisit makes it the standard choice for time-series occupancy indices. Individual vehicles are detectable but not reliably separated at this resolution without pansharpening or shadow analysis; the method works best as a density signal rather than a discrete count.
- Planet SkySat: 0.5 m panchromatic, 1 m multispectral. Sufficient to resolve individual passenger vehicles (typical car footprint roughly 4 m × 2 m) and apply shadow-based segmentation. Revisit is tasked rather than systematic, so it suits event-driven sampling rather than daily monitoring.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral, 3.7 m SWIR. The SWIR bands add a weak thermal proxy useful for distinguishing recently parked (engine-warm) vehicles in some conditions, though this is not a primary detection method. Archive depth extends to 2014.
- Airbus Pléiades Neo: 0.3 m panchromatic. Comparable to WorldView-3 for discrete vehicle segmentation. Stereo tasking allows canopy-height modelling to flag covered structures that will be opaque to the method.
What a car park actually tells you
A car park is a proxy, not a direct measure. The chain of inference runs: satellite counts vehicles, vehicles imply visits, visits imply revenue or output. Each link introduces uncertainty. But the proxy has genuine value because it is observable before retailers report quarterly sales, before factories release production figures, and before government statistical agencies publish footfall indices.
Published work in the alternative-data literature, including studies using Planet imagery over US big-box retailers, has shown statistically significant correlations between satellite-derived occupancy ratios and same-store sales reported weeks later. The signal is cleaner for large surface car parks with a single tenant than for mixed-use sites where vehicle origin is ambiguous. Industrial sites, particularly automotive assembly plants, offer a different but equally readable signal: shift changes produce sharp occupancy transitions that a twice-daily pass can detect.
From raw pixels to a vehicle count
The detection pipeline has three stages. First, pansharpening fuses a high-resolution panchromatic band with lower-resolution multispectral bands to produce a colour image at or near panchromatic resolution. WorldView-3 at 0.31 m panchromatic is the ceiling of what commercial systems currently offer; SkySat at 0.5 m is the practical workhorse for most published studies.
Second, shadow-based segmentation exploits the fact that a parked vehicle at a known sun angle casts a shadow of predictable length and direction. At 0.5 m resolution, the shadow of a standard passenger car (roof height roughly 1.4 m) is detectable as a dark elongated feature. Algorithms trained on labelled imagery can distinguish vehicle shadows from building shadows, trees and other artefacts with published precision figures typically in the 85 to 95 per cent range on clear-image test sets, though real-world performance degrades with low sun angles, wet asphalt (reduced shadow contrast) and partial cloud.
Third, the raw count is converted to an occupancy ratio: detected vehicles divided by the marked or estimated capacity of the lot. This normalisation matters because absolute counts are hard to compare across sites of different sizes. The ratio is then indexed against a baseline period to produce the leading indicator that analysts actually use.
Where the method breaks down
Optical sensors see surfaces, not interiors. Multi-storey car parks are opaque. A retail centre that migrated its parking underground or into a deck structure in the past decade will show an empty surface lot that tells you nothing about actual footfall. Before deploying this method, a site audit using stereo imagery or existing mapping data is necessary to classify each target as surface-only, partially covered or fully covered.
Cloud is the other structural limit. At equatorial and mid-latitude retail corridors, cloud cover can obscure a site for several consecutive days. A daily-revisit constellation like PlanetScope improves the odds of a usable pass, but it does not eliminate the problem. Published occupancy studies typically report effective observation rates of around 60 to 80 per cent of calendar days for temperate sites, lower in monsoon climates.
Electric vehicles are beginning to distort the industrial shift-pattern signal. Conventional analysis assumed that a full car park during off-hours indicated overnight charging of electric forklifts or other equipment, a recognisable artefact. As the passenger vehicle fleet electrifies, workers at some sites are leaving cars parked during the day to charge rather than driving away at shift end. This is a small but growing source of false positives in shift-detection models, and it is worth flagging honestly to any client using this method for manufacturing output estimation.
Spectral shortcuts and their limits
Some practitioners apply a simple spectral filter before shadow segmentation: asphalt has a distinctive reflectance profile in the SWIR bands available on WorldView-3, and vehicles sitting on asphalt create a local spectral anomaly. This shortcut speeds processing but introduces errors at sites with concrete paving, gravel or light-coloured asphalt, all of which shift the background reflectance enough to confuse the filter.
NIR bands help distinguish vegetation from vehicles at the lot perimeter, reducing false positives from tree canopy. The red-edge band on Planet SuperDove is useful for this boundary classification even though the 3 m resolution is too coarse for discrete vehicle detection. A practical pipeline often combines SuperDove for daily boundary and density estimation with periodic SkySat or WorldView-3 tasking for precise counting on key dates such as Black Friday, quarterly earnings windows or announced factory restart dates.
Building an occupancy index that investors and operators can use
A single observation is a data point. A time series is an index. The analytic product that has genuine decision value is an occupancy ratio tracked over months or years, seasonally adjusted and benchmarked against a pre-defined baseline. Seasonal adjustment matters: a UK retail park will show predictably lower occupancy on wet January Tuesdays than on dry December Saturdays. Without adjustment, the signal is noise.
Satellize runs occupancy analytics on client-specified site lists, combining open-constellation passes for daily density context with commercial tasking for precision counts. The approach is similar in structure to the crop-estimation programme Satellize operates for the Kingdom of Tonga: define the target geometry, establish a baseline, then track deviation from it. For retail and industrial clients, the deliverable is typically a weekly index feed with flagged anomalies, not a raw image archive.
Latency is a practical consideration. A SkySat pass processed through a cloud-hosted pipeline can produce a vehicle count within two to four hours of acquisition. PlanetScope daily mosaics are typically available within 24 hours. Neither figure is fast enough for intraday trading signals, but both are well ahead of monthly retail sales releases or quarterly production reports.
What to ask before commissioning a parking-lot study
Four questions determine whether the method will work for a given site list. What fraction of parking capacity is surface-level and unobstructed? What is the typical cloud-cover frequency at the target locations? Is the tenant mix homogeneous enough that vehicle counts map cleanly to a single economic variable? And has the site undergone any physical changes, such as new deck construction or lot expansion, that would break the baseline?
If the answers are favourable, parking-lot occupancy analysis is one of the more defensible remote-sensing proxies in the alternative-data toolkit. The physics are simple, the imagery is commercially available, and the academic literature provides a reasonable validation base. If the answers are mixed, the honest recommendation is to combine the method with complementary sources rather than rely on it alone.
Typical figures
| Spatial resolution (panchromatic) | 0.31 m (WorldView-3), 0.5 m (SkySat), 3 m (PlanetScope SuperDove) |
| Minimum detectable target | Standard passenger car (approx. 4 m × 2 m footprint) detectable at 0.5 m resolution; density-only signal at 3 m |
| Revisit rate | Daily (PlanetScope); tasked within 1–4 days (SkySat, WorldView-3 depending on latitude and priority) |
| Spectral bands used | Panchromatic (shadow detection), RGB + NIR (vegetation masking), SWIR (asphalt discrimination, WorldView-3 only) |
| Effective observation rate | Approximately 60–80% of calendar days for temperate sites; lower in monsoon or persistently overcast climates |
| Processing latency | 2–4 hours post-acquisition (SkySat, tasked); ~24 hours (PlanetScope daily mosaic) |
| Archive depth | WorldView-3 from 2014; PlanetScope from approximately 2016; SkySat from approximately 2017 |
| Coverage | Global for all three systems; SkySat and WorldView-3 require tasking requests for specific sites |
| Covered/multi-storey car parks | Not detectable by optical sensors; site audit required before deployment |
| Delivery formats | GeoTIFF (imagery), GeoJSON or Shapefile (vehicle detections), CSV or JSON time-series index feed |
Analytics Satellize can run
| Weekly occupancy ratio index | Shadow-based vehicle segmentation on SkySat or WorldView-3 imagery, normalised to lot capacity, seasonally adjusted against a rolling baseline | Weekly CSV or JSON feed with per-site occupancy ratios and deviation flags |
| Daily density heatmap | PlanetScope 3 m multispectral density estimation using spectral anomaly filtering and NIR vegetation masking | Daily GeoTIFF layer per site, suitable for GIS ingestion |
| Event-driven precision count | Tasked SkySat or WorldView-3 acquisition on client-specified dates (earnings windows, trading events, factory restart announcements), processed through pansharpened segmentation pipeline | Single-date vehicle count report with confidence interval and image thumbnail |
| Industrial shift-pattern detection | Twice-daily occupancy comparison (morning and afternoon passes where available) to identify shift-change transitions at manufacturing or warehouse sites | Shift-activity log with timestamps and occupancy transitions, delivered as structured data feed |
| Site coverage audit | Stereo imagery or existing elevation data used to classify each target site as surface-only, partially covered or fully covered before method deployment | Site classification GeoJSON with coverage fraction and method-suitability score |
| Multi-site retail footfall index | Aggregated occupancy ratios across a defined portfolio of retail sites, weighted by lot capacity, benchmarked against a sector baseline | Monthly index report with trend charts and anomaly annotations |
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.