Weapons storage site inventory-change detection
Sub-metre optical imagery and SAR coherence analysis can track munitions stocks, armoured vehicle parks, and weapons storage areas with enough fidelity to detect significant inventory changes between passes.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic resolution, 1.24 m multispectral (eight VNIR bands plus eight SWIR bands). The SWIR bands can partially penetrate thin camouflage netting. Revisit at mid-latitudes is roughly 1 to 4.5 days depending on off-nadir scheduling. The practical floor for individual-item identification and shadow-based length estimation.
- Airbus Pléiades Neo: 0.30 m native panchromatic resolution, four multispectral bands at 1.2 m. Stereo and tri-stereo collection modes allow surface-height estimation of stacked materiel. Revisit is approximately 1 day at mid-latitudes with the two-satellite constellation. Suitable for change-detection time series where consistent off-nadir angle is scheduled.
- Planet SkySat: 0.50 m panchromatic, 1.0 m multispectral. The constellation of roughly 20 satellites can be tasked for multiple passes per day over a fixed site. Lower radiometric depth than WorldView-3 but high temporal density is valuable for establishing normal activity patterns before anomalies are flagged.
- Sentinel-1 SAR (IW mode): C-band synthetic aperture radar at 5 m x 20 m (IW mode, range x azimuth). Freely available at 6-day revisit with both satellites. Cannot resolve individual items at the scale of a rifle crate, but detects large metallic objects, vehicle concentrations, and coherence loss caused by ground disturbance or materiel movement. Cloud-independent and works at night, making it the baseline continuity sensor when optical tasking is unavailable.
What a shadow library actually is
The technique has been applied openly by analysts at the Arms Control Association, the Middlebury Institute, and Bellingcat to sites in Russia, North Korea, and across the Middle East. The principle is straightforward: at a known solar elevation angle, a vehicle or munitions container of known height casts a shadow of predictable length. Measure the shadow in a sub-metre image, apply the trigonometry, and you get a height estimate precise enough to distinguish a T-72 from a T-80, or a standard NATO ammunition pallet stack from a single layer.
A shadow library is a catalogue of reference silhouettes and shadow-length ratios compiled from confirmed imagery of known equipment types, then used to classify ambiguous objects in new scenes. The method works best when the sun angle is between roughly 20 and 60 degrees elevation. Very low angles produce shadows too long to fit within a revetment bay; very high angles compress shadows to near zero. Analysts therefore note the acquisition time and date for every scene and flag images where the geometry degrades classification confidence.
The honest limits of counting from orbit
Partial occlusion is the most persistent problem. A revetment wall designed to protect a vehicle from blast also hides it from nadir-looking sensors. Analysts at open-source intelligence organisations routinely report counts as ranges rather than precise figures precisely because some fraction of any storage area is structurally obscured. Off-nadir collection can reduce this, but it introduces perspective distortion that degrades shadow-length measurements.
Camouflage netting is a more variable obstacle. Standard thermal-IR-defeating nets are largely opaque to the VNIR bands used by most commercial sensors. WorldView-3's SWIR bands (1.195 to 2.365 micrometres) offer partial penetration of some net materials, particularly older or degraded netting, but this is not reliable enough to be treated as a detection guarantee. Underground or hardened storage is simply invisible to optical sensors and requires surface-signature inference covered on the underground-facility page in this library.
Revisit creates a second ambiguity: a vehicle absent in one image may have moved temporarily for maintenance rather than departed the site. Statistically meaningful inventory-change conclusions require a time series of at least several weeks, with enough passes to distinguish a one-off movement from a sustained drawdown or build-up.
SAR as the continuity layer
Sentinel-1's C-band radar cannot count individual shells, but it contributes two things that optical sensors cannot. First, it operates regardless of cloud cover and darkness, which matters considerably for sites in northern latitudes or during winter months. Second, coherence analysis between two Sentinel-1 passes separated by 6 or 12 days produces a change map sensitive to surface disturbance at scales larger than roughly a vehicle footprint. A sudden coherence-loss patch over a previously stable storage area is a reliable cue to task a commercial optical sensor.
The practical workflow is therefore layered. Sentinel-1 provides persistent, free, cloud-independent monitoring and generates alerts. Commercial tasking of WorldView-3 or Pléiades Neo is then triggered selectively, controlling cost while preserving the ability to characterise what changed. This is not a theoretical architecture; it mirrors the approach described in published open-source intelligence methodologies.
Quantifying uncertainty in change estimates
A responsible inventory-change report states a count range, not a single figure, and explains the sources of variance. Typical published analyses from open-source groups acknowledge three main contributors: occlusion fraction (what percentage of the site is structurally hidden), classification confidence (how many objects were positively typed versus ambiguously typed), and temporal sampling (how many passes covered the reference period).
Statistical uncertainty compounds quickly. If 20 percent of a vehicle park is occluded and classification confidence for visible objects is 85 percent, the effective uncertainty on a count of 50 vehicles is roughly plus or minus 12 to 15 units before any temporal sampling error is added. Clients should treat satellite-derived inventory figures as indicators of direction and magnitude of change, not as auditable stock records. That is an honest position, and it is also the position taken by every credible open-source analyst who has published on this subject.
Building a site baseline and change threshold
Detection of meaningful change requires a stable baseline. For a site with commercial archive imagery going back to 2014 (WorldView-2 launched in 2009, WorldView-3 in 2014), it is possible to characterise normal occupancy variance across seasons, exercise cycles, and known historical events. Anomalies are then defined statistically relative to that baseline rather than against a single prior image.
The minimum detectable change is site-specific. A large open vehicle park with consistent solar geometry and no netting can support detection of the removal or addition of a single main battle tank at 0.31 m resolution. A partially netted revetment area may require a change of five or more vehicles before the signal clears the noise floor. Satellize structures site assessments to report both the observed change and the detection threshold applicable to that specific site geometry, so clients understand what the data can and cannot confirm.
For clients building a treaty-verification or arms-control monitoring programme, the methodology used here is consistent with published approaches discussed in open literature on national technical means and open-source verification. Satellize's analytics work, including the Tonga crop-estimation programme, applies the same principle of stating explicit confidence intervals rather than false precision.
What a delivered product looks like
A standard site-monitoring report annotates the most recent high-resolution image with object classifications, shadow-derived height estimates, count ranges by category (armoured vehicles, artillery pieces, logistics vehicles, unclassified objects), and a change summary relative to the baseline period. A GIS layer accompanies the report, with polygons for each identified object and attribute fields for classification confidence and occlusion status.
For clients requiring faster turnaround, a change-alert feed can be configured: Sentinel-1 coherence is processed automatically on each new pass, and an alert is issued when a coherence-loss area exceeds a defined threshold. The alert triggers a commercial tasking request, and the optical analysis follows within the sensor's next available pass window. Latency from Sentinel-1 alert to optical delivery is typically two to five days, depending on tasking queue and cloud probability at the target location.
Typical figures
| Best optical resolution (panchromatic) | 0.30 m (Pléiades Neo), 0.31 m (WorldView-3) |
| Multispectral resolution | 1.0 to 1.24 m (commercial VHR); 10 m (Sentinel-2, not suitable for item counting) |
| SAR resolution (Sentinel-1 IW) | 5 m range x 20 m azimuth; 10 m ground-range detected product |
| Optical revisit (commercial tasking) | 1 to 4.5 days at mid-latitudes; multiple passes per day possible with SkySat |
| SAR revisit (Sentinel-1, two-satellite) | 6 days at equator; shorter at higher latitudes due to orbit geometry |
| Minimum detectable individual item | Sub-metre optical required; practical floor for a main battle tank is 0.5 m resolution or better |
| Archive depth (commercial optical) | WorldView-2 from 2009; WorldView-3 from 2014; Pléiades from 2011 |
| Archive depth (Sentinel-1 SAR) | April 2014 (Sentinel-1A launch) to present, freely accessible |
| Spectral bands relevant to netting penetration | WorldView-3 SWIR: 1.195 to 2.365 µm (eight bands); partial penetration of some net materials |
| Delivery formats | Annotated GeoTIFF, GeoJSON/Shapefile object layer, PDF site report, alert feed (webhook or email) |
Analytics Satellize can run
| Baseline occupancy profile | Multi-date archive image stack; manual and semi-automated object detection with shadow-length classification | PDF report with annotated imagery, count ranges by vehicle/equipment category, seasonal variance table |
| Periodic change summary | Image differencing and object re-classification against baseline; occlusion-fraction adjustment | Bi-weekly or monthly GIS layer update with change polygons, confidence scores, and narrative summary |
| SAR coherence-change alert | Sentinel-1 interferometric coherence difference between successive 6-day passes; threshold exceedance detection | Automated alert (webhook or email) with coherence-loss polygon and recommended optical tasking priority |
| Equipment-type classification report | Shadow-library matching against reference silhouettes at known solar elevation; confidence-rated per object | Annotated image with per-object classification, height estimate, and confidence tier (confirmed / probable / possible) |
| Occlusion and uncertainty statement | Geometric analysis of revetment walls and netting coverage relative to sensor look angle; Monte Carlo count-range estimation | Uncertainty annex appended to each site report, stating count range, occlusion fraction, and detection threshold for that site geometry |
| Multi-year trend series | Archive tasking across commercial and open-access imagery; longitudinal count and configuration tracking | Time-series chart and GIS archive layer suitable for treaty-verification or intelligence-assessment workflows |
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.