Informal settlement mapping for pre-disaster exposure assessment
Corrugated metal and tarpaulin roofs have a spectral and textural signature that formal construction lacks. Mapping that signature before a disaster, then overlaying hazard zones, turns a satellite image into a population-exposure register.
Sensors
- Maxar WorldView-3: 30 cm panchromatic, 1.24 m multispectral (8 bands including SWIR), revisit roughly 1 day at mid-latitudes. The SWIR bands distinguish bare metal roofing from painted concrete at wavelengths formal sensors miss. The sub-metre resolution resolves individual roof panels and lane widths, which are the primary morphological discriminators.
- Airbus Pleiades Neo: 30 cm native resolution, 4-band plus panchromatic, daily revisit across a four-satellite constellation. Stereo and tri-stereo tasking produces surface models that help separate single-storey informal structures from multi-storey formal blocks where plan-view texture alone is ambiguous.
- Planet SkySat: 50 cm resolution, up to 12 tasked passes per day over a target. Change detection between monthly or quarterly SkySat collections catches the rapid settlement expansion that makes a static baseline map obsolete within one rainy season.
- ESA Sentinel-2 MSI: 10 m resolution in visible and near-infrared, 5-day global revisit at the equator (2–3 days at higher latitudes with both satellites). Too coarse to resolve individual roofs, but the 13-band archive back to 2015 supports settlement-growth time series and provides a free, cloud-diverse baseline for change flagging before expensive commercial tasking is committed.
What a corrugated roof gives away
Informal roofing materials, principally galvanised corrugated iron and polyethylene tarpaulin, have a spectral profile that diverges from fired clay tile or painted concrete in the near-infrared and shortwave-infrared portions of the spectrum. Galvanised steel reflects strongly and relatively uniformly across visible wavelengths, producing high brightness with low colour saturation. Tarpaulin, often blue or silver, adds a narrow spectral peak that separates it from both soil and vegetation. WorldView-3's eight SWIR bands (1195 nm to 2365 nm) were designed partly for mineralogical discrimination, but they work equally well on roofing: corrugated metal shows a flat SWIR response while asphalt shingles absorb, and the contrast is detectable even in moderate cloud shadow.
Spatial texture matters as much as spectral signature. Informal settlements typically have irregular roof orientations, highly variable roof sizes (often 10 to 60 square metres), and lane widths below 2 metres that are invisible to Sentinel-2 but clearly resolved at 30 cm. Morphological indices derived from object-based image analysis, including building footprint compactness, nearest-neighbour distance, and road network connectivity, consistently outperform pixel-level spectral classifiers in the peer-reviewed literature on settlement mapping. The combination of SWIR spectral features and sub-metre morphological indices is the current state of practice.
Where the method struggles
The spectral distinction between informal and formal dense housing collapses in several common situations. Old formal housing in South and Southeast Asian cities frequently uses corrugated metal roofing for cost reasons, producing spectral signatures identical to informal settlements. Newly built informal structures in some African cities now use fired brick and concrete roofing as materials become cheaper, making them spectrally formal. In both cases, the classifier must fall back on morphological cues alone, and accuracy typically drops from the 85 to 92 per cent range reported in favourable conditions to 70 to 80 per cent.
Cloud cover is the operational ceiling. Tropical cities, which concentrate the highest informal-settlement populations, routinely experience 60 to 80 per cent cloud cover during monsoon seasons. A single clear-sky acquisition over Lagos or Dhaka in August may require weeks of tasking attempts. Optical methods cannot penetrate cloud; a mapping campaign must be planned around seasonal windows or accept partial coverage. Rapid settlement change compounds this: Planet SkySat studies of expanding peri-urban areas in sub-Saharan Africa have documented 15 to 25 per cent areal growth in a single year, meaning a baseline older than twelve months may misrepresent exposure significantly.
Vertical ambiguity is a subtler problem. A single-storey informal structure and a three-storey formal block can have similar roof-level spectral signatures. Without a surface model from stereo tasking or external LiDAR, population estimates based on footprint area alone will undercount formal dense areas and overcount informal ones at similar densities. Stereo collection adds cost and requires two passes, which doubles cloud-interference probability.
From classified pixels to an exposure register
Settlement classification is only the first step. The analytic product that matters for disaster preparedness is a spatially explicit count of people inside each hazard zone, broken down by settlement type. That requires three inputs: the classified settlement layer, a population distribution model, and one or more hazard footprints.
Population distribution is typically estimated by disaggregating census data using the classified building footprint as a dasymetric weight. WorldPop and LandScan publish gridded population surfaces at 100 m and 90 m resolution respectively, derived from exactly this approach, and they can be used as a prior. The hazard layers, flood return-period extents, landslide susceptibility indices, or probabilistic seismic intensity maps, come from national disaster-management agencies, UNDRR, or global datasets such as the AQUEDUCT flood layers from the World Resources Institute. Overlaying the three inputs produces a table: settlement type, area, estimated population, and return-period exceedance probability for each hazard. That table is what a civil-protection planner can act on before an event occurs.
Keeping the baseline current
A pre-disaster exposure assessment is not a one-time product. Informal settlements in rapidly urbanising regions change faster than any annual mapping cycle can track. The practical answer is a tiered update strategy. Sentinel-2 time series, free and continuous, flags areas of significant land-cover change at 10 m. When the change signal exceeds a threshold, typically a normalised difference built-up index shift of 0.05 or more sustained across two or three acquisitions, it triggers a commercial tasking order for sub-metre imagery over that specific area only. This concentrates expensive tasking budget on locations that have actually changed rather than re-mapping the entire city on a fixed schedule.
Change thresholds need local calibration. Construction activity, seasonal vegetation, and flooding all produce spectral changes that can mimic settlement expansion. A classifier trained on one city's spectral conditions will degrade when applied elsewhere without retraining. This is not a solvable problem in the abstract; it is a reason to maintain city-specific training libraries and to validate against ground truth at least annually.
What the analytics pipeline produces
The outputs of a well-structured informal-settlement exposure programme are concrete and machine-readable. A classified building-type raster at the resolution of the input imagery, typically 30 cm to 1 m per pixel, forms the base layer. Object-based segmentation then produces a vector polygon file of individual structure footprints with attached attributes: estimated material class, confidence score, and area. Aggregating footprints to administrative units or community boundaries yields a settlement-density surface. Intersecting that surface with hazard return-period layers produces the exposure table described above.
Satellize runs this pipeline on commercial imagery acquired under client licence, combining open Sentinel-2 baselines with tasked WorldView-3 or Pleiades Neo for the sub-metre classification layer. The approach is similar in structure to the methods applied in the Tonga crop-estimation programme, where Sentinel-2 time series flags change and targeted commercial imagery resolves ambiguous areas. Deliverables are GeoTIFF rasters, GeoPackage or Shapefile vector outputs, and a PDF summary report with exposure statistics by administrative unit and hazard type, formatted for upload to national disaster-management platforms.
Honest limits, stated plainly
Sub-metre optical mapping of informal settlements is the best available method for pre-disaster exposure assessment in the absence of comprehensive cadastral records. It is not infallible. Classification accuracy in complex urban environments rarely exceeds 90 per cent even under ideal conditions. Cloud cover can delay or prevent acquisition during the seasons when flood risk is highest. Population estimates derived from dasymetric disaggregation carry uncertainty that compounds the classification error; treat them as order-of-magnitude figures rather than census-quality counts.
The method also says nothing about structural vulnerability. Knowing that a roof is corrugated metal tells you the material; it does not tell you whether the walls are load-bearing masonry or timber frame, or how steep the slope beneath the structure is. Seismic and landslide risk assessments that depend on structural vulnerability need field survey or engineering judgement to complement the satellite-derived exposure layer. The satellite product defines where to look. It does not replace looking.
Typical figures
| Best available spatial resolution | 30 cm (WorldView-3 panchromatic, Pleiades Neo) |
| Multispectral resolution | 1.24 m (WorldView-3 8-band VNIR+SWIR); 50 cm (SkySat); 10 m (Sentinel-2) |
| Revisit for commercial tasking | 1 day or better (WorldView-3, Pleiades Neo constellation); up to 12 passes/day (SkySat) |
| Sentinel-2 free baseline revisit | 5 days global; 2–3 days at latitudes above 40° |
| Minimum resolvable structure | Individual roof panels from ~2 m² at 30 cm resolution; lane widths >1.5 m detectable |
| Key spectral bands for material discrimination | SWIR 1 (1195–1225 nm) and SWIR 2 (1550–1590 nm) on WorldView-3; NIR and SWIR on Sentinel-2 |
| Cloud penetration | None. Optical sensors require clear or thin-cloud conditions; SAR is needed for cloud-persistent coverage |
| Archive depth | Sentinel-2: from 2015. WorldView family (including WorldView-1/2): from 2007 in some areas. SkySat: from 2014 |
| Typical classification accuracy (favourable conditions) | 85–92% overall accuracy reported in published studies; drops to 70–80% in spectrally ambiguous contexts |
| Deliverable formats | GeoTIFF (classified raster), GeoPackage/Shapefile (structure polygons), PDF exposure report |
Analytics Satellize can run
| Informal settlement extent layer | Object-based image analysis (OBIA) with SWIR spectral indices and morphological texture features; supervised classification trained on labelled roof samples | GeoTIFF raster and vector polygon GeoPackage with material-class attributes and confidence scores |
| Building footprint density surface | Kernel density estimation over classified structure centroids, aggregated to 50 m or 100 m grid; normalised by administrative unit area | GeoTIFF density raster at configurable grid resolution |
| Population exposure by hazard zone | Dasymetric disaggregation of census population using building-footprint weights; spatial intersection with flood return-period, landslide susceptibility, or seismic intensity layers | CSV and PDF exposure table by administrative unit and hazard return period (10-, 50-, 100-year) |
| Settlement change alert | Sentinel-2 normalised difference built-up index (NDBI) time series; threshold-based change flagging triggers commercial tasking order for affected tiles | Automated alert GeoJSON with bounding box of changed area and magnitude estimate; delivered within 48 hours of Sentinel-2 acquisition |
| Stereo-derived surface model for vertical disambiguation | Pleiades Neo stereo or tri-stereo photogrammetry; normalised digital surface model (nDSM) differenced against terrain model to estimate building height | GeoTIFF nDSM at 0.5–1 m resolution; building height attribute appended to footprint polygons |
| Priority re-mapping schedule | Change-rate modelling from historical Sentinel-2 NDBI time series; ranked list of sub-areas by estimated annual growth rate | Tasking priority report with recommended acquisition windows and estimated cloud-free probability by month |
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.