Sentinel-2 built-up area expansion and urban growth mapping

Sentinel-2's 10-metre multispectral bands resolve the spectral contrast between concrete, bare soil and vegetation well enough to classify built-up extent and quantify urban expansion rates from bi-temporal or time-series imagery. Cloud cover and construction-site ambiguity are real limits; SAR coherence fills both gaps.

Sensors

Sentinel-2 MSI (ESA): 10-metre resolution in bands 2, 3, 4 and 8; 20-metre in bands 11 and 12 (SWIR). Five-day global revisit with two satellites. The SWIR bands are critical for separating impervious surfaces from bare soil. Free and open archive from 2015.
Sentinel-1 C-band SAR (ESA): 10-metre ground range resolution in Interferometric Wide swath mode. Coherence between repeat passes (6- or 12-day) drops sharply over vegetation and water but stays high over stable built structures, making coherence a reliable built-up indicator regardless of cloud cover.
Landsat 8/9 OLI (USGS/NASA): 30-metre resolution across comparable spectral bands including SWIR. Sixteen-day revisit per satellite, 8-day combined. Archive extends to 1972 (Landsat 1 MSS), making it indispensable for multi-decade urban growth studies where Sentinel-2 coverage does not reach.
Planet Dove: 3-metre resolution, daily revisit globally. Four-band (RGBNIR) constellation useful for resolving individual buildings and narrow roads that fall below Sentinel-2's detection floor. Commercial licence required; no open archive.

What the spectrum reveals about concrete

Impervious surfaces, rooftops, roads and compacted ground have a characteristic spectral signature: moderate reflectance in the red (band 4, 665 nm), high reflectance in the near-infrared (band 8, 842 nm) and notably high reflectance in the shortwave infrared (band 11, 1610 nm). Vegetation suppresses SWIR reflectance and amplifies NIR through the red-edge; water absorbs almost everything beyond 750 nm. That contrast is the physical basis of every built-up index derived from Sentinel-2.

The most widely used index is the Normalised Difference Built-up Index (NDBI), computed from SWIR and NIR: (B11 - B8) / (B11 + B8). Positive NDBI values indicate built-up or bare surfaces; negative values indicate vegetation or water. Combining NDBI with NDVI and the Modified Normalised Difference Water Index lets analysts assign pixels to four broad classes: built-up, vegetation, bare soil and water. At 10 metres per pixel, a single Sentinel-2 tile covers 100 km² per pixel column and resolves structures larger than roughly 20 to 30 metres across. A suburban house is detectable; a footpath is not.

The construction-site problem

Bare agricultural soil and active construction sites are spectrally nearly identical in Sentinel-2 bands 4, 8 and 11. Both show high SWIR reflectance, low NIR and low NDVI. A freshly ploughed field outside Lagos and a new housing development outside Nairobi can produce the same pixel value. This is the single largest source of commission error in urban mapping from optical data alone.

Several strategies reduce the confusion. Temporal consistency helps: agricultural bare soil is seasonal and reverts to vegetation within weeks; construction sites stay bare for months or years. A time series of six or more observations across a growing season separates the two behaviours reliably. The Global Human Settlement Layer (GHSL), produced by the European Commission's Joint Research Centre, uses exactly this logic combined with morphological filtering to distinguish built-up patches from agricultural patterns. The GHSL release at 10-metre resolution (R2023A) draws on Sentinel-2 imagery and provides epoch layers at 2018 and 2020 that are publicly downloadable and widely cited in urban planning literature.

SAR coherence is the more direct solution. A Sentinel-1 coherence pair computed over 12 days shows high coherence (values approaching 1.0) over stable rooftops and low coherence over disturbed soil, crops or water. Fusing an optical built-up index with a SAR coherence layer reduces the false-positive rate from agricultural bare soil substantially, though the gain depends on local crop calendars and soil moisture conditions.

Measuring expansion: bi-temporal versus time series

Bi-temporal change detection compares two classified maps, one from a baseline date and one from a target date, and flags pixels that switched from non-built to built. It is fast and interpretable. The weakness is that any classification error in either epoch propagates directly into the change layer. A 5% error rate in each map compounds to a change layer that may be 10% wrong by area.

Time-series methods are slower to compute but more defensible. Algorithms such as CCDC (Continuous Change Detection and Classification), developed at Boston University and published in the peer-reviewed literature, fit spectral trajectories to every pixel across hundreds of observations and detect statistically significant breaks. Urban conversion produces a permanent step-change in SWIR and NIR that CCDC identifies with a date estimate accurate to within a few weeks. The Landsat archive, which extends back to 1984 in consistent format, is the standard input for multi-decade CCDC runs; Sentinel-2 extends the approach forward at higher spatial resolution from 2015 onwards.

The 10-metre floor and what sits below it

Ten metres is a hard resolution limit, not a soft one. A road 6 metres wide occupies less than one Sentinel-2 pixel. A single-storey rural structure of 8 by 10 metres covers one pixel at most. Both will be classified as whatever spectral class dominates the surrounding area. In practice, Sentinel-2 urban maps undercount low-density peri-urban development: scattered small buildings surrounded by vegetation are systematically missed because the pixel-level signal is dominated by the green canopy.

This matters operationally for infrastructure planning and tax-base assessment, where individual structures count. Planet Dove at 3 metres resolves individual rooftops and narrow roads. The trade-off is cost, proprietary archive access and the absence of SWIR bands in the standard four-band product, which reintroduces the construction-site ambiguity. There is no single sensor that solves all three problems simultaneously. Analysts working on peri-urban fringe mapping typically use Sentinel-2 for area-level statistics and task commercial imagery for ground-truth sampling.

SAR coherence as a cloud-proof complement

Persistent cloud cover over tropical cities can reduce usable Sentinel-2 observations to fewer than four per year in some regions. Sentinel-1, operating in C-band at 5.4 GHz, penetrates cloud completely. Its coherence product does not classify built-up area directly in the way an optical index does, but it identifies stable scatterers: surfaces that return a consistent phase between repeat passes. Rooftops, walls and paved surfaces are stable; vegetation, water and disturbed soil are not.

The standard workflow fuses a mean annual coherence layer with optical classification results. Pixels that are optically ambiguous (moderate NDBI, low NDVI) but show high SAR coherence are reclassified as built-up. The method was validated in published studies across sub-Saharan African cities where cloud cover made optical-only classification unreliable. Coherence is not a perfect signal: metal roofs produce very high coherence but so do some rock outcrops; temporal decorrelation from wind-blown vegetation can mimic low coherence over sparse urban areas. Cross-checking against the optical record remains necessary.

What an analyst can actually deliver

A credible urban expansion product from Sentinel-2 and Sentinel-1 data includes: a classified land-cover map at 10-metre resolution for each epoch; a change layer showing net built-up gain by administrative unit; an annualised expansion rate in hectares per year; and a confidence layer derived from observation density and SAR coherence agreement. Accuracy assessments against high-resolution reference imagery typically report overall accuracies of 85 to 92% for built-up class at city scale, with lower figures in peri-urban fringe areas.

Satellize runs this workflow on open Sentinel and Landsat archives for government planning clients, with the same analytical pipeline used in the Tonga crop-estimation programme adapted for urban rather than agricultural classification. The honest caveat for any prospective client is that accuracy varies by geography, cloud climatology and the density of the urban fabric being mapped. A city in the Sahel with minimal cloud cover and high spectral contrast between concrete and sand will produce better results than a humid tropical city with dense tree cover over informal settlements. Scoping a pilot area first is not caution; it is the correct methodology.

Typical figures

Spatial resolution (Sentinel-2 optical)	10 m (bands 2, 3, 4, 8); 20 m (bands 11, 12 SWIR)
Spatial resolution (Sentinel-1 coherence)	10 m ground range (IW mode); coherence computed at 20–40 m effective resolution after multilooking
Revisit (Sentinel-2, two satellites)	5 days at equator; 2–3 days at mid-latitudes
Revisit (Sentinel-1, two satellites)	6 days (Europe, some regions); 12 days globally
Minimum detectable built-up feature	Structures roughly 20–30 m across for reliable classification; sub-pixel features missed
Archive depth	Sentinel-2 from 2015; Landsat from 1984 (OLI from 2013); GHSL epochs from 1975
Spectral bands used	Red (665 nm), NIR (842 nm), SWIR (1610 nm, 2190 nm) for NDBI and index fusion
Cloud sensitivity	Optical bands fully blocked by cloud; SAR coherence unaffected
Typical classification accuracy (built-up class)	85–92% overall at city scale; lower in peri-urban fringe and under persistent cloud
Delivery formats	GeoTIFF (classified map, change layer, confidence layer), GeoPackage, COG, area statistics as CSV

Analytics Satellize can run

Bi-temporal built-up extent change map	NDBI/NDVI/MNDWI index fusion with threshold classification, per-class accuracy assessment against reference sample	GeoTIFF change layer with net built-up gain/loss by administrative unit; PDF accuracy report
Annualised urban expansion rate	Time-series classification (CCDC or equivalent) on Sentinel-2 and Landsat stack; pixel-level change date extraction	Tabular expansion rate in hectares per year per district; time-series chart
SAR-coherence-fused built-up map	Sentinel-1 IW coherence computed over 12-day pairs, fused with optical NDBI classification to resolve cloud-obscured and spectrally ambiguous pixels	Merged GeoTIFF with confidence layer; confusion matrix comparing optical-only versus fused result
Construction activity monitoring	Monthly NDBI and coherence time series flagging pixels with persistent bare-soil signature and stable SAR return, distinguishing active construction from seasonal agriculture	Monthly GIS layer of active construction zones; alert feed for new sites above a configurable area threshold
GHSL-aligned settlement density layer	Classification aligned to Global Human Settlement Layer R2023A methodology; output compatible with GHSL epoch comparisons for longitudinal planning studies	10-metre settlement density GeoTIFF; comparison table against GHSL baseline epochs
Peri-urban fringe delineation	Morphological filtering and spatial adjacency analysis on classified map to separate contiguous urban core from scattered peri-urban development	Vector polygon layer of urban core, fringe and rural zones; area statistics by zone

Sources

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.