Artillery firing-position and impact-crater density mapping
Sub-metre optical and SAR imagery can map shell-crater density, orientation, and gun-pit scrapes across contested terrain, supporting probabilistic back-azimuth estimation of firing positions and bombardment-intensity assessment.
Sensors
- Maxar WorldView-3: 31 cm panchromatic resolution at nadir; enables measurement of crater diameter and ejecta orientation for individual impacts. Tasked commercially; revisit over a fixed point is typically 1–4.5 days depending on latitude and cloud.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral. Stereo and tri-stereo modes allow shallow depth-of-field modelling that distinguishes fresh craters from older depressions. Revisit 1–2 days with the two-satellite constellation.
- Planet SkySat: 50 cm resolution, video mode available. Lower dynamic range than WorldView-3 but useful for rapid area coverage and change detection between passes. Revisit on tasked targets can be same-day.
- Capella Space SAR (X-band): Spotlight mode delivers roughly 50 cm resolution regardless of cloud cover or time of day. SAR backscatter detects freshly disturbed soil and gun-pit earthworks when optical is denied. Does not yield ejecta colour contrast, so crater count confidence is lower than optical.
What a crater actually tells you
A shell impact on soft ground produces a roughly circular pit with an asymmetric ejecta fan. The fan's long axis points back towards the firing direction. For high-angle fire (mortars, howitzers in high-quadrant elevation), the fan is nearly symmetric and back-azimuth precision degrades. For low-angle direct or semi-direct fire, the asymmetry is pronounced and the back-azimuth can be estimated to within a few degrees from a single well-preserved crater, assuming the ground surface is reasonably flat.
Crater diameter is weakly correlated with shell calibre and charge. A 152 mm artillery round detonating in agricultural soil typically produces a crater 3–6 m across; a 122 mm round, 2–4 m. Those ranges overlap substantially, and soil type, fusing, angle of impact, and whether the round was airburst all shift the result. Calibre attribution from satellite imagery alone is probabilistic at best. Analysts who claim precise calibre identification from crater geometry without ground-truth are overstating the method.
The geometry of back-azimuth estimation
Back-azimuth estimation from crater fields is a well-documented forensic technique. The New York Times Visual Investigations team and Bellingcat applied it systematically to imagery of the Mariupol and Kharkiv regions during the 2022 Ukraine conflict, identifying probable firing positions by intersecting back-azimuths from multiple craters across a field. The logic is straightforward: if twenty craters in a 500 m grid all have ejecta fans pointing within a 10-degree arc, the firing battery lies somewhere along that arc.
Precision improves with crater count and degrades with terrain complexity. Urban craters are harder to read: hard surfaces reduce ejecta asymmetry, and secondary fragmentation obscures the primary fan. Wooded terrain hides craters from nadir sensors entirely. The honest working range for this method is open or semi-open terrain with at least five to ten clearly resolved craters and a sensor with better than 50 cm ground sample distance. Below that resolution, ejecta fans are not reliably measurable.
Gun-pit scrapes and position indicators
Artillery firing positions leave their own signatures. A self-propelled gun or towed howitzer in a prepared position creates a shallow scrape or berm, typically 5–15 m across, often with a small spoil heap to one side and vehicle track marks approaching from the rear. At 30–50 cm resolution these are detectable in open terrain. At 1 m resolution they are ambiguous with other agricultural or construction features.
Repeated occupation of the same position, common when a battery fires multiple missions from a prepared site, produces cumulative soil disturbance that becomes more legible over time. Change-detection between two passes separated by days or weeks can isolate freshly disturbed ground from background agricultural activity. SAR coherence analysis (comparing phase coherence between Sentinel-1 passes at 6–12 m resolution) can flag disturbed soil at the field level, though it cannot resolve individual gun pits at that resolution.
Density mapping and bombardment-intensity assessment
Beyond individual craters, the spatial density of impacts across an area quantifies bombardment intensity. Crater counting over a defined polygon, divided by area, gives a craters-per-hectare figure that can be compared across time periods or geographic zones. This is directly useful for post-conflict damage assessment, humanitarian demining prioritisation, and legal documentation of area bombardment.
Automated crater detection using convolutional neural networks trained on labelled WorldView and Pléiades imagery has been published in the remote-sensing literature. Detection rates in open terrain at 30–50 cm resolution typically reach 80–90 % with false-positive rates of 5–15 %, depending on training data quality and terrain type. Urban environments and forested areas reduce detection rates sharply. Manual analyst review remains necessary for any product that will be used in legal or accountability contexts.
Archive depth matters here. Maxar's commercial archive extends back to the late 1990s for some areas; Pléiades data is available from 2012. Establishing a pre-conflict baseline, confirming that a feature is a crater rather than a pre-existing pond or agricultural pit, requires at least one clean pre-event image of the same area.
Honest limits of the method
Cloud cover is the primary operational constraint. Optical sensors cannot see through it. Persistent cloud over a target area can delay collection by days or weeks in continental climates. SAR partially compensates but with reduced crater-detection confidence. Night collection is possible with SAR and with WorldView-3's low-light capability, but resolution and contrast are lower.
The method produces probabilities, not certainties. A back-azimuth estimate places the firing position somewhere along a line; it does not identify a grid square. Intersecting azimuths from multiple crater clusters narrows the position, but terrain masking, gun displacement between missions, and measurement error in ejecta orientation all introduce uncertainty. Presenting this analysis to a legal tribunal or a command authority requires explicit confidence intervals and a clear statement of assumptions.
What a structured analytic workflow looks like
A practical programme for a government client typically combines three layers. First, a wide-area change-detection pass using Sentinel-1 SAR coherence or Planet's daily optical mosaic to flag newly disturbed terrain at low resolution. Second, targeted tasking of WorldView-3 or Pléiades Neo over flagged areas to collect sub-metre imagery. Third, automated crater detection followed by manual analyst review to produce a georeferenced crater database with per-crater attributes: diameter estimate, ejecta orientation, confidence score, and pre-event baseline confirmation.
Output formats that are actually useful include GeoJSON or shapefile layers for GIS ingestion, a density raster (craters per hectare) clipped to the area of interest, and a back-azimuth vector layer showing probable firing arcs. A written assessment should accompany any geospatial product, stating the sensor, collection date, resolution, detection method, and confidence level for each finding. Satellize builds these workflows on open and commercial constellations depending on client licence, applying the same structured uncertainty framework it uses for analytics work such as the Tonga crop-estimation programme.
Typical figures
| Best available optical resolution | 30 cm (WorldView-3 panchromatic, Pléiades Neo panchromatic) |
| Minimum detectable crater diameter | ~2 m at 30 cm GSD; ~5 m at 1 m GSD (open terrain, good contrast) |
| SAR resolution (spotlight mode) | ~50 cm (Capella Space X-band); 3–6 m (Sentinel-1 IW, usable for coherence change only) |
| Optical revisit (tasked) | 1–4.5 days (WorldView-3); 1–2 days (Pléiades Neo two-satellite pair) |
| SAR revisit (Sentinel-1) | 6 days (single satellite); 3 days (two-satellite constellation, where active) |
| Archive depth | Maxar: late 1990s for some areas; Pléiades: 2012 onwards; Sentinel-1: 2014 onwards |
| Cloud penetration | SAR only; optical sensors blocked by cloud and smoke |
| Crater-detection accuracy (published, open terrain) | 80–90 % detection rate, 5–15 % false positives (CNN methods, 30–50 cm imagery) |
| Delivery formats | GeoJSON, Shapefile, GeoTIFF density raster, PDF assessment report |
| Back-azimuth angular precision | ±3–10° per crater (open terrain, low-angle fire); degrades significantly for high-angle or urban impacts |
Analytics Satellize can run
| Crater density raster | Automated CNN-based crater detection on sub-metre optical imagery, validated against manual count | GeoTIFF raster (craters per hectare) with confidence layer, clipped to area of interest |
| Per-crater attribute database | Object detection and ellipse fitting on panchromatic imagery; manual review for legal-grade products | GeoJSON or Shapefile with diameter estimate, ejecta orientation, collection date, confidence score, pre-event baseline flag |
| Back-azimuth firing-arc vectors | Ejecta-fan orientation measurement; statistical clustering of azimuths across crater field | GIS vector layer of probable firing arcs with angular uncertainty bounds; written assessment of intersection confidence |
| Gun-pit and scrape detection | Change detection between pre- and post-event optical passes; SAR coherence-loss flagging for initial cueing | Georeferenced point layer of probable firing positions with supporting imagery chips |
| Temporal bombardment timeline | Multi-date crater count differencing across archive and tasked imagery | Time-series chart of cumulative crater density by polygon; GIS animation of impact accumulation |
| Wide-area disturbance alert | Sentinel-1 SAR coherence change detection at 6–12 m resolution for rapid cueing before optical tasking | Alert polygon feed (GeoJSON) flagging newly incoherent terrain, updated on each Sentinel-1 pass |
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.