Public transit stop activity estimation from very-high-resolution revisit
Sub-hourly VHR passes over bus stops, ferry terminals and rail stations let analysts count vehicle queues and crowd proxies where fare-gate data simply do not exist. The method is indirect but spatially universal.
Sensors
- Planet SkySat: 0.50 m native resolution, up to 12 passes per day over a target when tasked across the full constellation of 21 satellites. The intra-day cadence is the key property for counting peak and off-peak vehicle occupancy at the same stop.
- Maxar WorldView Legion: 30 cm panchromatic resolution, up to 15 revisits per day over mid-latitude cities. At 30 cm, individual pedestrians are not resolved but crowd density can be estimated from shadow geometry and pixel-brightness variance, consistent with published methods.
- Airbus Pleiades Neo: 30 cm panchromatic, 1.2 m multispectral, daily revisit in constellation of four. Stereo tasking adds a surface-model dimension useful for distinguishing double-deck buses from single-deck, which affects capacity estimates.
- Airbus OneAtlas archive and tasking: Provides access to historical SPOT and Pleiades imagery for baseline comparisons, alongside new-tasking workflows. Useful for establishing pre-project activity levels before an intervention such as a new route or terminal.
What a satellite can and cannot see at a bus stop
A 30 cm image resolves a saloon car unambiguously and a bus with enough detail to distinguish it from a heavy goods vehicle. A pedestrian at ground level occupies roughly two to four pixels at that resolution, which is not enough to read a face but is enough to count discrete objects in an open apron. Published crowd-density estimation studies, including work using WorldView imagery over public squares, have demonstrated object-detection approaches that achieve reasonable counts where crowds are not tightly packed and shadows do not merge.
The honest ceiling is low-density crowds in open settings. A ferry terminal with a broad concrete forecourt and clear sightlines is a good candidate. A covered bus station with a canopy blocking the roof view is not. Rail platforms beneath station roofs are largely invisible to nadir-pointing sensors. The method therefore works best at surface-level stops, open-air terminals and park-and-ride facilities, not at underground or covered infrastructure.
Revisit cadence is the whole argument
A single image tells you what was happening at one moment. Transit demand is a time-series problem. Morning peak, midday trough and evening peak can differ by a factor of three or more at the same stop. Sub-hourly revisit, which SkySat and WorldView Legion can provide over a tasked city, allows analysts to construct an intra-day activity curve rather than a single snapshot.
In practice, cloud cover and satellite geometry constrain usable passes. Over tropical cities, afternoon convective cloud regularly eliminates one or two passes per day. A realistic campaign over a six-week period might yield eight to fourteen usable observations per stop per day on average, not the theoretical maximum. That is still a richer time series than most transit authorities hold for informal stops, which often have no instrumentation at all.
The comparison that matters is not against automated passenger counters at well-instrumented metro stations. It is against the blank cells in a network spreadsheet where a planner has no data whatsoever. Satellite-derived proxies fill those cells with something defensible.
From pixel counts to activity indices
The analytic pipeline has three stages. First, vehicle detection: classifying pixels into bus, minibus, motorcycle-taxi, private car and bicycle using object detection models trained on VHR imagery. Bus count at a stop is the most direct proxy for service frequency and, indirectly, demand. Second, feeder-vehicle density: the number of motorcycle-taxis or shared taxis idling near a stop correlates with interchange demand. Drivers congregate where passengers are. Third, open-apron crowd estimation: where a forecourt is visible, pedestrian density is estimated from object detection or, where crowds are dense enough to merge, from texture and brightness variance methods documented in the remote-sensing literature.
None of these outputs is a passenger count in the sense a turnstile produces. They are relative indices. Stop A consistently shows three times the bus dwell and twice the feeder-vehicle density of Stop B. That ranking is actionable for route-planning even if the absolute numbers carry a margin of error that a transit engineer would find uncomfortable in a fare-revenue model.
Calibration against ground truth, where any exists, sharpens the relationship between index and actual boardings. Even a one-day manual count at a sample of stops gives the regression anchor needed to convert relative indices into plausible boarding ranges.
Cloud, shadow and the geometry of failure
Optical VHR is the only sensor class with the resolution to count individual vehicles at a bus stop. Synthetic aperture radar resolves vehicles on motorways at 1 m resolution (Sentinel-1 is 5 x 20 m in IW mode, which is too coarse; commercial SAR at 0.3 to 1 m from ICEYE or Capella is closer but still struggles to classify vehicle type in cluttered urban scenes). There is no radar substitute for this application at present.
Shadow is a compounding problem. A bus casts a shadow that can obscure adjacent pedestrians. Low sun angles, common at high latitudes in winter or in early-morning passes, increase shadow extent. Analysts should flag passes with solar elevation below roughly 30 degrees as lower confidence. Seasonal scheduling of tasking campaigns to favour high-sun conditions is straightforward and worth doing.
Where this fits in a transit planning workflow
Transit authorities in high-income cities spend considerably on automatic passenger counting systems, GPS tracking and smart-card analytics. The satellite method is not a replacement for any of that. It is a tool for the planning phase in cities where that infrastructure has not been built, or for rapid assessment of informal route networks that operate outside the formal fare system entirely.
A government commissioning a new bus rapid transit corridor needs to know which existing informal stops carry the highest interchange demand before it designs terminal locations. A development bank funding a ferry network expansion needs a baseline activity estimate at candidate piers before it can model economic returns. These are the settings where a six-week VHR tasking campaign, producing a ranked activity index across fifty candidate stops, has a clear decision value.
Satellize has applied similar multi-temporal VHR analytics to agricultural settings, including the Kingdom of Tonga crop-estimation programme, and the same pipeline principles transfer to urban object-counting tasks. Analysts interested in a scoping conversation for a specific city or corridor can request a sample analysis over a defined area of interest.
Typical figures
| Spatial resolution (panchromatic) | 0.30 m (WorldView Legion, Pleiades Neo) to 0.50 m (SkySat) |
| Intra-day revisit (tasked) | Up to 12 passes/day (SkySat constellation); up to 15 passes/day (WorldView Legion over mid-latitude cities) |
| Minimum detectable vehicle | Saloon car reliably at 0.30 m; motorcycle at 0.30 m with lower confidence; pedestrian as discrete object in open areas at 0.30 m |
| Cloud sensitivity | Optical only; tropical convective cloud typically reduces usable passes by 30-50% in wet season |
| Spectral bands | Panchromatic plus 4-band multispectral (blue, green, red, NIR) for surface classification |
| Typical tasking latency | 24-72 hours from order to first collect; historical archive available for baseline |
| Archive depth | Pleiades/SPOT archive from 2011; WorldView archive from 2008; SkySat from 2016 |
| Coverage per collect | SkySat strip: 6.6 km wide; Pleiades Neo: 14 km swath; WorldView Legion: 13.1 km swath |
| Delivery format | GeoTIFF orthoimage; GeoPackage or Shapefile for detected objects; CSV activity-index time series |
Analytics Satellize can run
| Intra-day bus-count time series per stop | Object detection on VHR panchromatic imagery using convolutional neural network classifiers trained on labelled vehicle datasets | CSV and GIS point layer with hourly bus-count index per stop across campaign period |
| Feeder-vehicle density map | Motorcycle-taxi and shared-taxi detection and spatial clustering around stop centroids | Heatmap GIS layer showing feeder-vehicle concentration by stop and time-of-day band |
| Open-apron crowd-density index | Pedestrian object detection and pixel-variance texture analysis in forecourt polygons, following published VHR crowd-estimation methods | Per-stop crowd-density score (low/medium/high) per pass, aggregated into daily peak and off-peak values |
| Stop activity ranking | Composite index combining bus count, feeder density and crowd proxy, normalised across the stop network | Ranked table of stops by activity level, suitable for corridor prioritisation or terminal siting decisions |
| Baseline vs. post-intervention comparison | Change detection between pre- and post-intervention tasking campaigns using the same object-detection pipeline | Before/after report with statistical significance flags and confidence intervals on index change |
| Shadow and cloud quality flags | Automated solar-elevation and cloud-fraction masking applied per pass before analysis | Per-image quality flag appended to all analytic outputs, allowing clients to filter low-confidence observations |
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.