Field hospital and forward triage site detection from imagery
Very-high-resolution optical and thermal sensors can identify field medical facilities by their physical signatures: tent clusters, vehicle parks, generator heat and freshly cut access tracks. Detection is possible but unambiguous identification requires sensor combinations and honest tolerance for residual ambiguity.
Sensors
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral, plus 8 SWIR bands at 3.7 m. SWIR bands distinguish canvas tent material from bare ground and detect surface thermal contrast from generator exhaust panels. Revisit approximately 1 day at mid-latitudes with off-nadir tasking.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral. Stereo and tri-stereo collection in a single pass enables surface-height modelling, which helps distinguish erected structures from flat ground clutter. Revisit 1 to 2 days with constellation of two satellites.
- Planet SkySat: 50 cm panchromatic, 1 m multispectral. Lower spectral depth than WorldView-3 but rapid tasking and video mode useful for vehicle-movement confirmation. Revisit can be daily for priority targets.
- Landsat 8/9 TIRS: Thermal infrared at 100 m native resolution (resampled to 30 m in products). Useful only for coarse contextual thermal mapping of an area of interest, not for resolving individual generator signatures. 16-day repeat per satellite, 8-day combined.
- MWIR commercial sensors (e.g. Satellogic thermal, EROS-B thermal mode): Mid-wave infrared sensors with ground sampling below 5 m are the threshold for detecting generator heat plumes at night with low ambiguity. This class of sensor remains limited in constellation size; tasking latency and cloud cover are real constraints.
What a field medical facility looks like from 500 km up
A field hospital is not a single object. It is a pattern. The canonical signature combines a cluster of large-footprint tents or inflatable shelters (typically 10 m to 20 m per unit), a vehicle park containing ambulances and logistics trucks, one or more generator sets with associated fuel bladders, a cleared landing zone or vehicle access loop, and a freshly graded track connecting the site to the nearest hard road. Each element is individually ambiguous. Together they are not.
At 30 cm to 50 cm resolution, WorldView-3 or Pléiades Neo imagery resolves tent ridge lines, guy-wire shadows and the characteristic rectangular footprint of military medical shelters such as the NATO DRASH or equivalent systems. Red-cross markings on tent roofs are occasionally visible at this resolution, though they are not always present and their absence proves nothing. The access track is often the most reliable early indicator: mechanically graded dirt roads that appear between one collection and the next, leading to a previously empty site, are a strong signal that something organised is being established.
Generator heat as a night-time discriminator
Medical facilities run continuously. Generators do not switch off at night, which makes thermal detection at night a useful complement to daytime optical analysis. A diesel generator set in the 50 kW to 200 kW range produces a surface temperature anomaly detectable by MWIR sensors, but the ground sampling distance must be below roughly 5 m to separate the generator footprint from background clutter. Landsat TIRS at 100 m native resolution cannot do this; it can only flag a warm zone across a broad area, useful for cueing but not for identification.
Published work by UNOSAT and the American Association for the Advancement of Science (AAAS) on humanitarian site characterisation has demonstrated that thermal anomalies co-located with tent clusters and vehicle parks significantly raise confidence in the medical-facility classification. The honest caveat is that any continuously operating industrial or military site produces a similar thermal signature. A forward fuel depot or a command post with heavy electronic equipment looks similar in thermal data alone. Multi-source fusion is not optional; it is the only way to reduce false positives to an operationally useful rate.
The ambiguity problem and where it bites hardest
Unambiguous identification of a field medical facility from imagery is harder than detection. Several site types share the same physical vocabulary: disaster-relief staging areas, military command posts, forward logistics elements and, in some theatres, civilian construction camps. The distinction matters enormously under international humanitarian law, which affords protected status to medical facilities. Misclassification in either direction carries serious consequences.
Resolution is the first constraint. Below 1 m, tent type and vehicle markings become interpretable. Above 3 m, classification rests almost entirely on site geometry and thermal context, which is insufficient for confident identification. Cloud cover is the second constraint: optical sensors are blind through it, and SAR, while cloud-penetrating, does not provide the spectral or thermal information needed for this task. A site that is established during a multi-day cloud event will be missed until the cloud clears. Revisit rate is the third: even with daily tasking, a 24-hour gap is enough for a facility to be established, used and partially struck. Temporal completeness is never guaranteed.
Method: from raw collection to a site classification
The analytical workflow follows the published UNOSAT humanitarian site-characterisation approach, adapted for the defence context. Step one is change detection: comparing a new collection against a baseline to flag any new structures, tracks or vehicle concentrations. Step two is object-based image analysis (OBIA), segmenting the flagged area into objects and classifying them by shape, size, spectral response and shadow geometry. Step three is thermal overlay: if a night-time MWIR collection is available for the same period, generator-scale heat anomalies are mapped and spatially joined to the optical objects. Step four is contextual scoring: proximity to a road network, distance from the forward line of troops, site area and internal layout geometry all contribute to a classification confidence score.
The output is a probability-weighted site classification, not a binary answer. A site scoring high on tent-cluster geometry, vehicle presence, thermal anomaly and road access is classified as probable medical facility. A site with only two of those four indicators is flagged as possible, requiring further collection. This tiered output is deliberate. Analysts who receive a binary yes/no classification from automated systems tend to stop asking questions.
Archive depth and the historical baseline question
One of the more useful aspects of this problem is that the sites of interest are often in areas with years of archived commercial and open-source imagery. WorldView-3 archive coverage for many conflict-affected regions extends back to 2014. Sentinel-2 (10 m multispectral) provides free, open archive coverage from 2015 with 5-day revisit at mid-latitudes, sufficient for detecting the access-track formation and large-footprint tent clusters even if individual tent identification requires tasking a higher-resolution asset.
Archive analysis allows an analyst to establish a pre-conflict baseline for a given grid square, making change detection far more sensitive than it would be from a single collection. A site that appears to be a permanent structure may, on archive review, have been erected within the past week. That temporal context is often more informative than the current image alone.
Practical limits a buyer should understand before tasking
Three limits deserve plain statement. First, protected-site verification under the Geneva Conventions requires corroborating evidence beyond imagery alone. Satellite analysis can raise or lower the probability of a classification; it cannot provide legal certainty. Second, active camouflage, including thermal blankets over generators and vehicle dispersal under tree cover, degrades detection significantly. A well-concealed site may produce no detectable signature at all. Third, the MWIR sensors capable of resolving generator-scale heat anomalies below 5 m ground sampling are not yet available on open constellations. Tasking them requires a commercial arrangement, and availability is not guaranteed on demand.
None of these limits make the capability useless. They make it a component of a broader intelligence picture, not a substitute for one.
Typical figures
| Optical resolution (VHR) | 30 cm to 50 cm panchromatic (WorldView-3, Pléiades Neo, SkySat) |
| Thermal resolution (MWIR) | Below 5 m GSD required for generator-scale detection; Landsat TIRS at 100 m for coarse context only |
| Revisit (optical VHR) | 1 to 2 days with off-nadir tasking; not guaranteed in cloud or conflict airspace |
| Revisit (Sentinel-2 context layer) | 5 days at mid-latitudes, free and open archive from 2015 |
| Spectral bands used | Panchromatic, multispectral (RGB + NIR), SWIR (WorldView-3 bands 1-8), MWIR thermal |
| Minimum detectable structure | Tent footprint approximately 10 m x 5 m at 50 cm resolution; generator heat anomaly requires sub-5 m MWIR |
| Archive depth | WorldView-3 from 2014; Sentinel-2 from 2015; Landsat from 1972 (coarse) |
| Classification output | Probability-weighted site classification (probable / possible / insufficient evidence), not binary |
| Cloud limitation | Optical and MWIR sensors blind through cloud cover; SAR does not provide thermal or spectral discrimination for this task |
| Delivery formats | GeoTIFF change layers, GeoJSON site polygons, PDF site-characterisation report, GIS-ready vector overlays |
Analytics Satellize can run
| Access-track formation alert | Optical change detection against baseline (OBIA on Sentinel-2 and VHR tasking) | Georeferenced alert with before/after image chips and track geometry, delivered within 24 hours of new collection |
| Tent-cluster footprint mapping | Object-based image analysis on VHR panchromatic and multispectral; shadow-geometry modelling for structure height | GeoJSON polygon layer with per-object classification confidence scores |
| Generator thermal anomaly overlay | MWIR night-time collection fused with daytime optical; spatial join to tent-cluster objects following UNOSAT site-characterisation methodology | Thermal anomaly raster co-registered to optical base layer, with flagged co-located objects |
| Site classification report | Multi-indicator scoring (tent geometry, vehicle presence, thermal signal, road access, site area) producing tiered classification | PDF site-characterisation report with confidence tier (probable / possible / insufficient evidence) and supporting imagery |
| Historical baseline and establishment timeline | Archive change detection across Sentinel-2 and available VHR archive; temporal sequencing of site-establishment phases | Time-series GIF and annotated timeline showing site growth from first detectable change to current state |
| Vehicle-park activity assessment | Multi-date vehicle counting and classification from VHR optical; comparison against baseline vehicle counts | Tabular vehicle-count time series with annotated image chips, delivered as GIS layer and spreadsheet |
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.