Very-high-resolution optical change detection
Sub-metre commercial optical imagery from Maxar, Airbus and Planet can resolve individual vehicles, construction equipment and structural damage. This page explains how change detection works, where it breaks down, and what the resolution-versus-revisit trade-off means in practice.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic resolution, 1.24 m multispectral (8 bands including SWIR), single-pass swath of 13.1 km. Revisit roughly 1 day at mid-latitudes when tasked, though cloud and competing tasking demands routinely stretch that. The SWIR bands help discriminate roofing materials and camouflage netting that fool visible-band analysis.
- Airbus Pléiades Neo: 0.30 m native panchromatic, 1.2 m multispectral (6 bands). Two satellites in the constellation give a same-day stereo collect over a target, which matters when you need a 3-D surface model as well as a 2-D change map. Swath is 14 km.
- Planet SkySat: 0.50 m panchromatic, 1.0 m multispectral (4 bands). The 21-satellite constellation can task a site multiple times per day in some orbits, trading some spatial resolution for dramatically higher revisit than WorldView or Pléiades. Useful for tracking fast-moving construction or equipment repositioning.
- BlackSky Global: 0.90 m multispectral. The constellation is optimised for rapid revisit over a small set of high-priority sites rather than broad area coverage. Latency from collect to delivery can be under 90 minutes, which suits time-sensitive monitoring of specific facilities.
What sub-metre resolution actually resolves
At 30 m resolution, a Landsat pixel covers roughly the footprint of a small office block. At 0.31 m, a WorldView-3 pixel is smaller than a sheet of A4 paper. That difference is not cosmetic. It means the imagery can resolve individual shipping containers, parked aircraft by type, construction equipment by class, and damage to specific structural bays of a building. An analyst can count vehicles in a motor pool. A change-detection algorithm can flag a new earthwork berm that appeared between two collects.
The practical floor for automated detection is roughly the size of the target relative to the ground sampling distance. Objects smaller than about three pixels across in any dimension are unreliable for automated classification, though a trained analyst can often infer more. At 0.30 m to 0.50 m GSD, that puts the reliable automated detection threshold at roughly one to two metres for compact objects, and somewhat smaller for linear features such as fences, trenches or road markings.
Three methods, and where each one fails
Image differencing is the oldest approach: subtract a baseline image from a new collect, pixel by pixel, and threshold the residual. It is fast and interpretable, but it is brutally sensitive to misregistration. A one-pixel shift between two WorldView-3 images at 0.31 m introduces a 31 cm positional error, which in a dense urban scene produces a fringe of false positives along every roofline and kerb. Orthorectification using a high-quality digital elevation model reduces this, but steep terrain and tall buildings still cause residual errors.
Object detection using convolutional neural networks trained on labelled VHR imagery can identify specific target classes, such as military vehicles, excavators or shipping containers, without requiring a registered baseline image. The published literature shows strong performance on open training sets such as DOTA and DIOR, but models trained on one geography or sensor often degrade significantly when applied to another. A model trained on WorldView-3 imagery of European industrial sites does not transfer without retraining to SkySat imagery of tropical port facilities.
Analyst interpretation remains the method of last resort and, for high-stakes decisions, the method of first resort. A skilled imagery analyst reading context, shadow geometry, access roads and surrounding activity can draw inferences that no current automated pipeline reliably replicates. The honest position is that automation accelerates curation and flags candidates; the analyst decides.
The resolution-revisit trade-off is a real constraint, not a marketing footnote
WorldView-3 and Pléiades Neo offer the finest spatial detail commercially available, but each satellite has a single sensor and a fixed orbital period. WorldView-3 has a published revisit of approximately 1 day at nadir for a given latitude, but that figure assumes the satellite is tasked to your site and the sky is clear. In practice, a busy constellation serving many clients, combined with persistent cloud cover over tropical or maritime targets, can stretch effective revisit to several days or longer.
SkySat's 21-satellite fleet can deliver multiple collects per day over a priority site, but at 0.50 m rather than 0.30 m, and with a narrower swath of roughly 6.6 km. For tracking a known facility, that trade is often acceptable. For searching a broad area for unknown activity, the swath limitation makes daily coverage expensive in tasking terms.
The honest implication: if your requirement is to detect a change that might occur on any given day across a 500 km² area, VHR optical tasking alone is unlikely to catch it reliably. SAR-based change detection, which is covered separately, does not share the cloud or solar-elevation constraints and can provide broader-area daily coverage, though at coarser resolution for most commercial systems.
Cloud and sun angle: the physics you cannot task around
Optical sensors record reflected sunlight. Two physical constraints follow from that. First, solar elevation must be sufficient to illuminate the scene without excessive shadowing. At high latitudes in winter, usable collection windows shrink to a few hours around local noon, and steep shadows from buildings or terrain obscure the very features an analyst needs to see. Second, cloud cover renders a collect useless regardless of resolution. Persistent cloud over a target during a critical monitoring window is not an edge case in many operational environments; it is the norm.
Tasking providers mitigate this through cloud-cover thresholds in collection orders, typically set at 10 to 20 per cent cloud cover for acceptance. But a 15 per cent cloud-cover collect can still have the cloud positioned directly over the site of interest. For time-critical monitoring in cloudy regions, the practical solution is to combine VHR optical tasking with a SAR layer that provides a weather-independent baseline, using the optical imagery when it arrives to confirm and characterise what the SAR flagged.
Urban scenes are spectrally hard
Automated change detection performs worst in dense urban environments, which is precisely where many clients most want it to work. The problem is spectral heterogeneity. A single urban block contains concrete, glass, asphalt, metal roofing, vegetation, painted surfaces and shadow, all varying with sun angle, atmospheric conditions and seasonal vegetation state. A legitimate change, such as a new building, produces a signal that competes with illegitimate changes caused by a different illumination angle between the baseline and the new collect, or a parked vehicle that was not there last time.
Normalised difference indices, principal component analysis and machine-learning classifiers all help, but none eliminates the problem. Published studies using WorldView imagery in urban change detection consistently report false positive rates that require significant analyst review to reduce to operationally useful levels. Clients should expect that a change-detection pipeline over a complex urban area will generate a candidate list, not a finished intelligence product, without analyst curation.
Where Satellize sits in this workflow
Satellize brokers commercial tasking on client licence and runs the change-detection analytics on top of the collected imagery. The workflow is not proprietary in its methods: image differencing, object-detection models and analyst review are well-documented techniques. The value is in integrating tasking, processing and curation into a single monitored output, and in being honest with clients about what the imagery can and cannot resolve on a given day.
For clients who need a reference point on what analytic outputs look like before committing to a tasking programme, Satellize's Overhead column publishes regular open-source analysis using publicly available imagery. That is a reasonable starting point for scoping a requirement.
Typical figures
| Best available panchromatic resolution | 0.30 m (Pléiades Neo), 0.31 m (WorldView-3) |
| Best available multispectral resolution | 1.2 m (Pléiades Neo), 1.24 m (WorldView-3) |
| Revisit frequency (tasked, clear sky) | ~1 day for WorldView-3 and Pléiades Neo; multiple times per day for SkySat over priority sites |
| Effective revisit (cloud and tasking contention) | 2 to 7+ days in practice for many tropical and maritime targets |
| Swath width | 13.1 km (WorldView-3), 14 km (Pléiades Neo), ~6.6 km (SkySat) |
| Spectral bands | Panchromatic plus 4 to 8 multispectral bands (VNIR); WorldView-3 adds 8 SWIR bands |
| Minimum reliably detectable object (automated) | ~1 to 2 m for compact objects; smaller for linear features |
| Delivery latency | 90 minutes to 48 hours depending on provider and processing tier |
| Archive depth | WorldView archive from 2009; Pléiades from 2012; SkySat from 2014 |
| Typical delivery formats | GeoTIFF (orthorectified), NITF, cloud-optimised GeoTIFF; metadata in XML or JSON |
Analytics Satellize can run
| Construction activity change map | Image differencing on co-registered orthorectified collects, thresholded by normalised change magnitude | GIS polygon layer of changed areas with date-stamped before/after chip pairs, delivered as GeoPackage or shapefile |
| Military equipment count and classification | Convolutional neural network object detection trained on labelled VHR imagery of vehicle classes | Structured report with bounding-box coordinates, equipment type, confidence score and collect timestamp |
| Infrastructure damage assessment | Analyst-curated change detection combining pixel differencing with object-level interpretation of structural features | Damage grading layer (intact, moderate, severe, destroyed) per building or structure polygon, compatible with QGIS and ArcGIS |
| Facility monitoring alert | Automated change candidate generation followed by analyst triage on a defined site polygon | Email or API alert within agreed latency window, with annotated image chip and analyst note on nature of change |
| 3-D surface model differencing | Stereo photogrammetry from same-day stereo pairs (Pléiades Neo), differenced against baseline DSM to detect volumetric change | Raster DSM difference layer showing height gain or loss in metres, with uncertainty estimate |
| Land-use transition mapping | Multi-date supervised classification of VHR multispectral imagery, change matrix computed between epochs | Classified raster and transition statistics table showing area converted between land-use categories |
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.