Emergency shelter material and density classification after displacement
Sub-metre multispectral and panchromatic imagery, combined with texture analysis and supervised classification, can distinguish tarpaulin, canvas, and corrugated-metal shelter types within newly established displacement camps to inform logistics and protection planning.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral (8 bands including coastal blue and near-infrared), plus 8 SWIR bands. Revisit roughly 1 day at mid-latitudes with off-nadir tasking. The SWIR bands are particularly useful for separating synthetic materials from natural fabrics. Tasking latency after a sudden displacement event can be several days depending on cloud and queue.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m multispectral (6 bands). Constellation of 4 satellites gives revisit of roughly 12 hours at moderate off-nadir angles. Radiometric quality is well-suited to supervised classification of rooftop materials.
- Planet SkySat: 0.50 m panchromatic, approximately 1 m multispectral (4 bands: blue, green, red, near-infrared). A constellation of 21 satellites provides flexible tasking. Four spectral bands limit material discrimination compared to WorldView-3, but rapid tasking and competitive latency make it useful for initial camp extent mapping.
- Airbus SPOT 7: 1.5 m panchromatic, 6 m multispectral. Below the reliable individual-shelter resolution threshold for most camp configurations, but useful for overall camp footprint, density zonation, and change detection between collection dates when sub-metre assets are unavailable.
Why shelter material matters to a logistics planner
When a displacement camp forms rapidly, the material composition of its shelters is not an aesthetic question. Tarpaulin degrades under ultraviolet exposure within weeks and offers poor thermal insulation. Canvas breathes but is heavier to distribute and requires different framing. Corrugated metal is durable but dangerous in high winds and conducts heat aggressively in tropical climates. A planner deciding how to allocate a consignment of replacement sheeting, or a protection officer assessing heat-stress risk, needs to know not just how many shelters exist but what they are made of.
Ground surveys take time and are dangerous in some post-displacement contexts. A classified satellite map produced within 48 to 72 hours of tasking can give a rough material breakdown across an entire camp that would take a field team days to replicate. The map is not a replacement for ground truth; it is a way to direct field teams to the right sectors first.
What a floating roof gives away: the spectral and textural signatures
Blue and white polyethylene tarpaulins have strong, narrow reflectance peaks in the visible blue and green bands and fall off sharply in the near-infrared. Corrugated galvanised metal reflects broadly across the visible and near-infrared spectrum with a characteristic specular glint at certain solar and sensor angles. Canvas and woven fabrics sit spectrally between the two, with more diffuse reflectance and a texture signature that differs from the regular parallel corrugations of metal sheeting.
Texture analysis, specifically grey-level co-occurrence matrix (GLCM) features computed from panchromatic imagery at 0.3 to 0.5 m resolution, adds a second discriminating dimension. Corrugated iron produces strong, regular periodicity in GLCM contrast and correlation measures. Tarpaulin, when laid flat, produces low contrast and high homogeneity. Crumpled or folded tarpaulin introduces irregular texture that can be confused with canvas; this is a genuine ambiguity the classifier must handle with care.
WorldView-3's eight SWIR bands (1195 to 2365 nm) provide an additional separability layer. Many synthetic polymer films have diagnostic absorption features in the SWIR that natural fabrics do not share. In practice, SWIR at 3.7 m resolution must be pan-sharpened or fused with the 1.24 m multispectral data before it contributes meaningfully to per-shelter classification.
The resolution floor and why 0.5 m is the practical threshold
A standard UNHCR emergency shelter unit occupies roughly 3.5 by 5 metres of ground. At 0.5 m resolution, that shelter spans approximately 7 by 10 pixels, enough to compute a meaningful texture window and extract a spectral signature that is not dominated by edge mixing with adjacent shelters or pathways. At 1.5 m resolution, the same shelter spans only 2 by 3 pixels. Spectral mixing with the ground, neighbouring structures, and shadows makes per-shelter classification unreliable. Camp-level density and footprint estimates remain feasible at 1.5 m, but material classification requires sub-metre data.
Shadow is a persistent problem. Shelters with vertical walls cast shadows onto adjacent structures, and shadow pixels carry no useful material signature. In dense camps where shelter spacing is less than one metre, shadow fractions can exceed 20 percent of the camp area in morning or late-afternoon imagery. Acquisition near solar noon reduces shadow extent but does not eliminate it.
Classification workflow: from raw image to material map
The standard workflow begins with orthorectification and atmospheric correction to surface reflectance. Pan-sharpening (typically Gram-Schmidt or principal component methods) fuses panchromatic and multispectral bands to the highest available resolution. A training sample of labelled shelter pixels, ideally verified against any available ground photographs or field reports, is then used to train a supervised classifier. Random forest and support vector machine classifiers are both well-documented for rooftop material mapping in the peer-reviewed remote sensing literature, with random forest generally showing better performance on imbalanced class distributions common in mixed camps.
Post-classification, individual shelter objects are delineated using image segmentation. Object-based analysis (OBIA) outperforms pixel-based approaches here because it allows shape, size, and compactness constraints to filter out misclassified ground patches. The output is a vector polygon layer with each shelter attributed by predicted material class, area, and a confidence score derived from the classifier's posterior probability.
Accuracy assessment requires independent ground-truth samples. Published studies on similar urban material mapping tasks at sub-metre resolution report overall accuracies of 80 to 92 percent for three to four material classes under good illumination conditions. Accuracy degrades with cloud shadow, dense canopy overhang, and mixed-material shelters where, for example, a tarpaulin is draped over a metal frame.
Honest limits: what the method cannot resolve
Material classification from above identifies rooftop surfaces only. A corrugated-metal roof on a mud-brick wall looks identical from orbit to a corrugated-metal roof on a timber frame. The structural integrity and thermal performance of the shelter depend on the full construction, not just the roofing material. Satellite classification should be presented to decision-makers as a rooftop-material map, not a shelter-quality map.
Cloud cover is the most disruptive operational constraint. In tropical and monsoonal contexts where many large displacement events occur, persistent cloud can delay usable optical acquisition by days or weeks. There is no optical workaround; SAR can detect camp extent and density under cloud but cannot discriminate shelter materials. Tasking latency from commercial constellations adds further delay: even with priority tasking, a clear-sky acquisition within 24 hours of a sudden displacement event is not guaranteed.
Finally, camp conditions change rapidly. A tarpaulin sheet replaced with metal roofing between two acquisition dates will appear as a change but carries no humanitarian interpretation without field context. Time-series classification is more informative than any single snapshot, but it requires repeated tasking and compounds cost.
From classification map to actionable output
The classification layer becomes operationally useful when it is combined with camp population estimates and shelter density metrics. Coverage area per shelter type, expressed in square metres, can be compared against UNHCR minimum standards (3.5 square metres of covered living space per person in emergency phase) to flag sectors where coverage is below threshold. Material-type proportions by camp sector feed directly into procurement planning for replacement sheeting.
Satellize structures analytics of this kind as GIS-ready vector deliverables with accompanying summary statistics, built on the same supervised-classification pipeline used in its crop-estimation work for the Kingdom of Tonga. The next step for an organisation evaluating this capability is to share a sample scene or a camp boundary polygon so that a pilot classification can be run against archived imagery before any tasking commitment is made.
Typical figures
| Minimum spatial resolution for per-shelter classification | 0.5 m or better (0.30 to 0.31 m panchromatic from Pléiades Neo and WorldView-3) |
| Multispectral bands used | Visible (blue, green, red), near-infrared, red-edge where available; SWIR (WorldView-3, 8 bands, 1195 to 2365 nm) for polymer discrimination |
| Typical tasking revisit | 12 to 24 hours (Pléiades Neo, off-nadir); 1 day (WorldView-3); 1 to 2 days (SkySat) |
| Post-event tasking latency | 24 to 96 hours typical for priority tasking; cloud and queue-dependent |
| Minimum detectable shelter unit | Approximately 3 × 3 m at 0.5 m resolution (36 pixels); smaller units unreliable |
| Classification accuracy (published analogues) | 80 to 92 percent overall accuracy for 3 to 4 material classes under good illumination |
| Archive depth | WorldView constellation from 2007; Pléiades from 2011; SkySat from 2013 |
| Delivery formats | GeoTIFF (classified raster), GeoPackage or Shapefile (object polygons), CSV summary statistics |
| Cloud cover constraint | Optical methods fail above approximately 20 to 30 percent cloud cover over camp area; no optical substitute under persistent cloud |
Analytics Satellize can run
| Shelter material classification map | Supervised random forest or SVM classifier trained on pan-sharpened multispectral imagery with GLCM texture features | Vector polygon layer (GeoPackage) with per-shelter material class, area, and confidence score |
| Camp density and coverage-area statistics | Object-based image analysis (OBIA) segmentation of classified shelter polygons; density calculated per 100 m grid cell | Gridded density raster and sector-level summary table (CSV) |
| Coverage-adequacy assessment against UNHCR minimum standards | Spatial join of classified shelter area with population-sector estimates; comparison against 3.5 m² per-person emergency threshold | Sector-level adequacy report (PDF and GIS layer) flagging below-threshold zones |
| Multi-date material change detection | Bi-temporal classification comparison; change pixels attributed to material upgrade, deterioration, or new construction | Change polygon layer with change type attribute and acquisition-date metadata |
| Procurement planning input | Aggregation of classified tarpaulin area by sector; comparison with expected degradation timeline from material type | Replacement-sheeting demand estimate by camp sector (tabular report) |
| Camp footprint and expansion monitoring | Automated change detection on sequential imagery using normalised difference built-up index or thresholded reflectance change | Camp perimeter polygon with dated expansion layers and area-growth time series chart |
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.