Slum upgrading programme progress tracking over time
Very-high-resolution multi-epoch imagery tracks physical changes in roof material, structure regularity and road access across informal settlements, providing an independent audit of whether slum-upgrading programmes are delivering measurable outcomes on the ground.
Sensors
- Airbus Pléiades Neo: 30 cm native panchromatic, 50 cm multispectral (4-band plus red-edge and deep-blue on Neo). Revisit of 1–2 days at mid-latitudes with constellation of two satellites. The resolution is sufficient to distinguish corrugated metal from tile or concrete slab at roof level, which sits at the practical threshold for material classification.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral (8 bands including SWIR). The SWIR bands (1195–2365 nm) add material discrimination that visible-only sensors cannot match, separating asbestos-cement sheet from galvanised iron on the basis of reflectance rather than colour alone. Revisit approximately 1 day at off-nadir.
- Planet SkySat: 50 cm panchromatic, 1 m multispectral (4-band). A constellation of 21 satellites enables frequent tasking, useful for catching construction activity between the less frequent Pléiades Neo or WorldView-3 acquisitions. Less spectral depth than WorldView-3 but adequate for structure-regularity and road-access metrics.
- Sentinel-2 MSI: 10 m multispectral (13 bands), 5-day revisit at the equator with both satellites. Too coarse for individual roof classification but invaluable for establishing the temporal context of a neighbourhood: tracking vegetation clearance, large-scale drainage works or bulk infrastructure rollout over years before the VHR tasking begins.
What a roof actually tells you
Roof material is the most consistently observable proxy for housing permanence from space. Corrugated galvanised iron, the dominant roofing in most informal settlements, has a high near-infrared reflectance and a spectrally flat visible signature. Fired clay tile and concrete slab both show lower reflectance and distinct spectral slopes. At 30–50 cm resolution these differences are detectable; below that threshold, mixed pixels start to dominate and classification accuracy degrades sharply. The 50 cm boundary is not arbitrary: it corresponds roughly to the minimum roof span that can be resolved as a single-material surface rather than a blended edge pixel.
WorldView-3's SWIR bands push this further. Asbestos-cement sheet, still common in older informal settlements and a health hazard that upgrading programmes specifically target, has a SWIR reflectance signature that separates it from galvanised iron even when both appear similarly bright in the visible. That distinction matters for programme monitoring: replacing asbestos with iron is an improvement, but it is not the same as replacing it with concrete slab, and a visible-only sensor cannot tell them apart reliably.
Structure regularity and road access as secondary indicators
Roof material is the sharpest signal, but it is not the only one. Structure regularity, meaning the degree to which building footprints follow consistent orientation and spacing, is a measurable geometric property. Upgraded structures tend to be more rectilinear and more uniformly spaced than organically grown informal construction. Comparing the distribution of building orientations and inter-structure gaps between a baseline epoch and a follow-up epoch gives a quantitative regularity score that correlates with planned reconstruction rather than incremental self-improvement.
Road access is the bluntest but most policy-relevant metric. A paved or compacted-surface access road wide enough for an emergency vehicle (typically 3.5 m or more) is visible at 30–50 cm resolution as a distinct linear feature with consistent spectral response. Its presence or absence before and after a programme intervention is a binary check that requires no spectral sophistication, only a clean acquisition and a competent analyst. Drainage channels, retaining walls and communal water points are also detectable at this resolution, though they require manual verification rather than automated classification.
Building a multi-epoch comparison that holds up to scrutiny
The analytical workflow starts with image normalisation. VHR sensors from different dates, different viewing angles and different atmospheric conditions produce images that are not directly comparable in raw digital number. Radiometric normalisation, typically using pseudo-invariant features such as concrete roads or large rooftops that should not have changed, brings epochs into a consistent reflectance space before any change detection is run.
Change detection itself is run at the object level, not the pixel level. Individual building footprints are segmented from the baseline image, then the spectral and geometric properties of each footprint are compared in the follow-up image. This produces a per-structure change record: roof material before and after, regularity score before and after, proximity to the nearest paved road before and after. Aggregated to the programme area, these records become the independent audit layer. They can be cross-tabulated against the programme's own reported outputs, street by street or block by block, to identify where reported improvements are visible in imagery and where they are not.
The attribution problem: programme or organic change?
This is the hardest methodological question in this use case, and any honest analysis must confront it directly. Informal settlements improve organically. Residents replace iron sheet with tile when they can afford to, without any government programme. A satellite cannot distinguish a government-funded roof replacement from a self-funded one.
The standard approach is a difference-in-differences comparison: measure the rate of improvement inside the programme boundary against the rate in a comparable control area outside it, over the same time period. If the programme area improves at a significantly higher rate than the control, the excess improvement is plausibly attributable to the intervention. This is not proof of causation. It is the same logic used in public-health programme evaluation, and it carries the same caveats: the control area must be genuinely comparable in baseline conditions, economic trajectory and proximity to infrastructure investment. Selecting that control area requires local knowledge that imagery alone cannot supply. Satellite analysis narrows the uncertainty; it does not eliminate it.
Archive depth and the baseline problem
A programme that began in 2018 needs a 2017 or earlier baseline to measure against. Pléiades 1A and 1B (predecessors to Neo, operational since 2011 and 2012) and WorldView-3 (operational since 2014) have commercially available archives that cover many cities. SkySat archive goes back to approximately 2014. For settlements in cities that were tasked for other purposes, a usable baseline image may already exist. For others, the archive is sparse and the analyst must work with whatever is available, which may mean accepting a baseline from six months into the programme rather than before it started. That is a genuine limitation and should be stated in any programme report.
Sentinel-2, free and open, provides a consistent 10 m record from 2015 onwards for every point on Earth. It cannot classify individual roofs, but it can establish that a neighbourhood's overall spectral character shifted in a direction consistent with increased built permanence, giving temporal context even when VHR archive is absent.
What an independent audit actually delivers
The output of this analysis is not a judgement on programme management. It is a spatial, quantifiable record of physical change that programme managers, finance ministries, donor agencies and affected communities can all read from the same map. Structures where roof material improved: mapped. Blocks where road access was extended: mapped. Areas where no detectable change occurred despite reported works: mapped, with the caveat that some improvements, such as internal plumbing or electrical connections, are invisible to any optical sensor.
Satellize runs this class of multi-epoch VHR analysis on client-licensed commercial imagery, applying the same object-based change-detection methods used in its Tonga crop-estimation programme to built-environment contexts. The deliverable is a GIS layer set with per-structure change attributes, a summary statistics table by programme zone, and a written audit note that explicitly states what the imagery can and cannot confirm. Commissioning agencies receive a product they can table in a programme review without needing to explain its limits themselves.
Typical figures
| Spatial resolution (VHR) | 30 cm panchromatic (Pléiades Neo, WorldView-3); 50 cm panchromatic (SkySat) |
| Spatial resolution (temporal context) | 10 m multispectral (Sentinel-2 MSI) |
| Revisit (VHR tasking) | 1–2 days (Pléiades Neo); ~1 day (WorldView-3 off-nadir); ~1 day (SkySat constellation) |
| Revisit (Sentinel-2) | 5 days at equator (2-satellite constellation) |
| Spectral bands | Pléiades Neo: PAN + 6-band MS (blue, green, red, red-edge, NIR, deep-blue). WorldView-3: PAN + 8-band VNIR + 8-band SWIR. SkySat: PAN + 4-band MS. Sentinel-2: 13 bands (443–2190 nm) |
| Minimum detectable roof span | ~1.5 m at 30 cm resolution (practical limit for single-material classification); classification reliability degrades below 50 cm pixel size |
| Archive depth (VHR) | Pléiades 1A/1B from 2011/2012; WorldView-3 from 2014; SkySat from ~2014; coverage varies by city tasking history |
| Archive depth (Sentinel-2) | Global continuous record from June 2015 |
| Cloud limitation | Optical sensors require cloud-free acquisitions; tropical cities may require multiple tasking attempts across weeks to obtain usable imagery in wet season |
| Delivery format | GeoTIFF orthorectified imagery; GeoPackage or Shapefile per-structure change layer; CSV summary statistics; PDF audit note |
Analytics Satellize can run
| Baseline roof-material classification map | Object-based image analysis (OBIA) with supervised spectral classification; WorldView-3 SWIR bands used for asbestos-cement vs metal discrimination | GIS polygon layer with per-structure roof-material class and confidence score at programme start date |
| Multi-epoch change detection: roof material | Radiometrically normalised bi-temporal OBIA comparison; per-structure material transition matrix (e.g. iron to tile, iron to concrete slab) | GIS change layer with transition type, epoch pair and area per structure; tabular summary by programme zone |
| Structure regularity index | Geometric analysis of segmented building footprints: orientation variance, inter-structure gap distribution, footprint rectangularity score | Per-block regularity score time series; map showing blocks with statistically significant regularity increase |
| Road-access extension mapping | Supervised linear-feature extraction on VHR imagery; surface-type classification (paved, compacted, unpaved) by spectral response and texture | Road-network change layer showing new or upgraded segments with estimated surface type and width |
| Difference-in-differences attribution estimate | Comparison of improvement rates inside programme boundary vs analyst-selected control area over matched time intervals; statistical significance test on rate difference | Written attribution note with confidence range, control-area justification and explicit statement of assumptions and limits |
| Sentinel-2 temporal context profile | NDBI (Normalised Difference Built-up Index) and bare-soil index time series at 10 m over full programme period from 2015 onwards | Annual index raster stack and trend chart showing neighbourhood-level built-environment trajectory as context for VHR findings |
| Programme audit report | Spatial cross-tabulation of satellite-detected improvements against programme-reported outputs by zone; gap analysis identifying areas with no detectable physical change | PDF audit report with mapped findings, per-zone improvement rates, and explicit list of improvement types not detectable by optical imagery |
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.