Informal settlement detection from satellite imagery
Sub-metre optical imagery and spectral analysis can distinguish informal settlements from formal low-rise housing by roof material, street geometry, and building density. The method is powerful but not infallible: dense formal housing creates genuine ambiguity that honest analysis must flag.
Sensors
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral across 8 bands including SWIR. The SWIR bands are particularly useful for discriminating metal roofing from asbestos and concrete. Revisit roughly 1–4.5 days depending on latitude and tasking priority.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral in 6 bands. Stereo and tri-stereo capable, which allows building-height estimation to complement footprint mapping. Revisit 1–2 days at mid-latitudes.
- Planet PlanetScope: 3 m multispectral, daily global revisit. Insufficient resolution to resolve individual roof panels, but well suited to tracking settlement boundary expansion month-on-month and flagging areas for higher-resolution tasking.
- Sentinel-2 MSI: 10 m visible and NIR, 20 m SWIR, 5-day revisit at the equator. Usable for large-area triage and temporal change, but individual structures are below the detection floor. Free and openly archived from 2015.
What a corrugated roof gives away
Informal settlements are not random. They have consistent physical signatures that remote sensing can exploit. The most diagnostic is roof material. Corrugated galvanised iron, the dominant roofing in unplanned settlements across sub-Saharan Africa, South and Southeast Asia, and Latin America, has a high and spectrally flat reflectance in the visible and near-infrared bands, with a characteristic dip in the SWIR that differs measurably from concrete tiles, clay, or asphalt. WorldView-3's eight-band SWIR capability was partly designed with exactly this kind of material discrimination in mind.
Texture is the second signature. When imaged at 30 cm, informal settlements show high spatial frequency variation: rooftop edges at irregular angles, gaps between structures that are too narrow to be streets, and the absence of the regular parcel geometry that formal planning imposes. Grey-level co-occurrence matrix (GLCM) texture features extracted from panchromatic imagery have been used in peer-reviewed studies to separate informal from formal fabric with overall accuracies in the 80–90 percent range, depending on city and training data quality. The third signature is street-network topology. Informal areas tend to produce organic, low-connectivity graphs when street centrelines are extracted, whereas formal grids show higher betweenness centrality and more regular intersection spacing.
Classification methods: from texture statistics to neural networks
Two broad method families dominate the published literature. The first uses hand-crafted features: GLCM statistics, morphological building indices, and normalised difference built-up indices derived from multispectral bands, fed into support vector machines or random forest classifiers. These are interpretable, require relatively modest training data, and perform well when the analyst understands the local context. Their weakness is transferability. A classifier trained on Nairobi's Mathare settlement does not automatically generalise to Manila or Karachi without retraining.
The second family uses convolutional neural networks (CNNs) trained on labelled slum polygons, often derived from OpenStreetMap or humanitarian mapping exercises. Fully convolutional networks and U-Net architectures applied to sub-metre imagery have demonstrated precision and recall figures in the 85–93 percent range on held-out test cities in published work, though these figures depend heavily on how well the training set represents the target city's building stock. Transfer learning reduces but does not eliminate the need for local labels. Both approaches benefit from pansharpening: fusing the high-resolution panchromatic band with multispectral data to get colour information at 30–50 cm effective resolution.
The ambiguity problem: when formal looks informal
Dense formal low-rise housing is the persistent source of false positives. In many cities, low-income but legally titled neighbourhoods use similar roof materials and achieve similar building densities to genuinely informal areas. Post-war reconstruction housing, social housing blocks from the 1970s, and traditional courtyard housing in North African medinas can all produce spectral and texture signatures that overlap substantially with informal settlements.
No honest methodology papers over this. The standard mitigation is multi-source fusion: combining spectral classification with cadastral records where they exist, OpenStreetMap building tags, nighttime light intensity as a proxy for service connectivity, and field validation sampling. Where cadastral data is absent or unreliable, which is precisely the situation in most cities where informal settlement mapping matters most, the ambiguity cannot be resolved from imagery alone. Analysts should report confidence intervals and flag ambiguous zones explicitly rather than delivering a binary map that implies false certainty.
Temporal change: where PlanetScope earns its place
A single classification epoch answers where informal settlements exist. Governments and planners usually need to know how fast they are growing and in which direction. PlanetScope's daily revisit at 3 m resolution is well suited to this. Change detection between monthly composites can resolve settlement boundary expansion at the scale of individual city blocks, which is sufficient to trigger a higher-resolution tasking order on WorldView-3 or Pléiades Neo for detailed characterisation of the newly built area.
The practical workflow is a two-tier system. PlanetScope provides the surveillance layer: cheap, frequent, and broad. Commercial sub-metre imagery provides the characterisation layer: expensive, targeted, and detailed. Sentinel-2 sits between them, useful for cities where budget constraints rule out commercial tasking and where 10 m resolution is sufficient to track gross boundary change over multi-year periods. Sentinel-2's archive runs from 2015, which gives a decade of baseline data for cities that have not been previously mapped.
Honest limits: what imagery cannot tell you
Satellite imagery classifies built form. It does not classify tenure, services, or living conditions directly. A neighbourhood can have informal-looking rooftops and be fully serviced with clean water, electricity, and legal title. Conversely, a neighbourhood can have concrete roofs and be entirely without services or legal recognition. The physical signature and the legal or social status are correlated but not identical, and conflating them in a policy document causes real harm.
Cloud cover is a practical constraint in tropical cities, which is where most of the world's fastest-growing informal settlements are located. Optical imagery is blind through cloud. A WorldView-3 tasking over Lagos or Dhaka during monsoon season may return unusable imagery for weeks. Synthetic aperture radar can penetrate cloud and has been used to detect building density changes, but at resolutions of 1–3 m for commercial SAR systems it lacks the fine-grained roof discrimination that makes optical imagery useful for this application. Analysts working in persistently cloudy regions need to plan around seasonal windows or accept longer change-detection intervals.
Satellize's analytics team has worked through similar classification-under-uncertainty problems in the Tonga crop-estimation programme, where spectral ambiguity between crop types required multi-date compositing and confidence-weighted outputs rather than hard classifications. The same discipline applies here.
From map to decision: what the output should look like
A useful informal settlement map for a municipal government is not a single raster layer. It is a set of polygon boundaries with associated attributes: estimated roof-material class, building density per hectare, confidence score, date of last imagery, and a flag indicating whether the area falls in an ambiguous zone requiring ground verification. Delivered as a GeoPackage or shapefile, it can be loaded directly into a city's GIS and joined to administrative boundaries for ward-level reporting.
Change reports comparing two epochs should express growth in hectares and in estimated structure count, not just in percentage terms, because percentage figures obscure absolute scale. A settlement that grows 40 percent from 2 hectares is a different planning problem from one that grows 40 percent from 200 hectares. Prioritising ground-truthing effort toward the high-confidence, high-growth polygons is where the map actually saves field teams time and money.
Typical figures
| Best spatial resolution (optical) | 30 cm panchromatic (WorldView-3, Pléiades Neo); 3 m (PlanetScope); 10 m (Sentinel-2) |
| Revisit frequency | Daily (PlanetScope); 1–4.5 days (WorldView-3, tasking dependent); 5 days (Sentinel-2 at equator) |
| Key spectral bands | Visible, NIR, and SWIR (WorldView-3 8-band SWIR most diagnostic for roof-material discrimination); visible + NIR (Pléiades Neo, PlanetScope, Sentinel-2) |
| Minimum detectable settlement area | Approximately 0.1–0.5 ha at sub-metre resolution; 1–5 ha practical floor at 3 m (PlanetScope) |
| Classification accuracy (published range) | 80–93% overall accuracy reported in peer-reviewed studies; varies with city, training data, and method |
| Archive depth | Sentinel-2 from 2015 (open); WorldView constellation from ~2007 (commercial, licensed access); PlanetScope from ~2016 |
| Cloud cover constraint | Full optical blackout under cloud; tropical cities may face weeks of unusable imagery during wet season |
| Typical delivery format | GeoPackage or shapefile polygon layer with confidence attributes; GeoTIFF classification raster; PDF change report |
| Latency from tasking to delivery | 3–10 days for commercial sub-metre (imagery acquisition plus processing); near-real-time for Sentinel-2 open data |
Analytics Satellize can run
| Informal settlement extent map | GLCM texture classification or U-Net CNN on pansharpened sub-metre imagery, trained on labelled polygons | Polygon GIS layer with roof-material class, building density, and per-polygon confidence score |
| Settlement boundary change report | Bi-temporal or multi-temporal change detection on PlanetScope monthly composites or Sentinel-2 time series | Tabular and cartographic report showing growth in hectares and estimated structure count between two epochs |
| Ambiguity-flagged classification | Posterior probability thresholding from random forest or CNN softmax outputs; zones below confidence threshold flagged for ground verification | GIS layer with explicit uncertain-zone polygons and recommended field-sampling locations |
| Roof-material spectral index map | Band-ratio analysis using WorldView-3 SWIR bands to discriminate corrugated iron, asbestos, concrete, and plastic sheeting | GeoTIFF raster with classified roof-material categories and legend |
| Multi-year growth trajectory | Annual settlement boundary extraction from Sentinel-2 archive (2015 to present) using supervised classification and post-classification comparison | Time-series chart and animated GIS layer showing decadal expansion direction and rate |
| High-priority growth zone alert | Automated PlanetScope change detection with threshold trigger for new construction above a set area, triggering commercial sub-metre tasking order | Alert notification with coordinates and recommended tasking parameters for follow-on characterisation |
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.