Rooftop rainwater harvesting potential mapping from imagery and rainfall data
Combining sub-metre building footprints, photogrammetric roof slopes, and long-term rainfall climatology lets planners rank every structure by harvestable volume before a single tank is installed.
Sensors
- Pleiades Neo: 30 cm panchromatic, 50 cm multispectral (6 bands), revisit roughly once per day at mid-latitudes with the four-satellite constellation. Primary source for building footprint delineation and photogrammetric DSM generation via stereo or tri-stereo acquisition.
- WorldView-3: 31 cm panchromatic, 1.24 m multispectral (8 bands), 13 cm SWIR. Stereo pairs yield dense point clouds from which roof pitch and ridge orientation can be extracted. Archive depth exceeds a decade for many cities.
- CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data): 0.05-degree (~5 km) gridded daily rainfall from 1981 to near-present, blending satellite cold-cloud duration with rain-gauge records. The primary source for long-term mean annual and seasonal rainfall totals used to compute expected catchment yield.
- ERA5 reanalysis: 31 km global hourly atmospheric reanalysis from ECMWF, 1940 to present. Useful where CHIRPS gauge coverage is sparse or for sub-daily intensity distributions that affect first-flush losses and tank sizing.
What a rooftop actually collects
The physics is simple. Annual harvestable volume equals effective catchment area multiplied by mean annual rainfall, reduced by a runoff coefficient that accounts for roofing material, slope, and first-flush losses. Corrugated iron at a 10-degree pitch typically returns a runoff coefficient of 0.80 to 0.90. Rough clay tile drops that to 0.60 to 0.75. Flat concrete with ponding can sit anywhere in between, depending on surface condition.
The complication is that none of those inputs, area, slope, material, or rainfall, are uniform across a city. A neighbourhood 15 kilometres inland from the coast can receive 30 percent less annual rainfall than the waterfront. Two adjacent buildings may have roof areas that differ by a factor of four. Mapping at the parcel level, rather than the ward or district level, is what makes the output actionable for a water authority deciding where to subsidise tank installation.
From pixels to catchment geometry
Building footprint extraction from Pleiades Neo or WorldView-3 imagery uses a combination of spectral contrast between rooftop materials and surrounding ground, and the sharp shadow edges that sub-metre resolution preserves. Deep-learning segmentation models trained on labelled urban datasets can achieve intersection-over-union scores above 0.80 on well-contrasted rooftops, though performance degrades on flat roofs that share spectral signatures with unpaved ground, and on rooftops obscured by overhanging trees. That degradation is honest and worth stating: tree canopy cover above roughly 40 percent in a neighbourhood will require manual correction or LiDAR supplementation.
Roof slope is extracted from a photogrammetric digital surface model generated by stereo or tri-stereo image matching. WorldView-3 tri-stereo acquisitions routinely produce DSMs with a vertical accuracy of 0.3 to 0.5 metres RMSE over urban areas, sufficient to distinguish a flat roof from a 15-degree pitch. Ridge orientation can be derived from the DSM gradient, which also informs which face of a pitched roof receives direct rainfall versus which drains to a neighbour's plot. The effective catchment area is the horizontal projection of the roof polygon, not the sloped surface area, because rainfall is measured vertically.
Rainfall climatology: CHIRPS versus ERA5
CHIRPS is the default input for this application. Its 40-plus-year daily record at 5 km resolution is long enough to compute reliable 10th, 50th, and 90th percentile annual totals, which matter for tank sizing under drought-year assumptions. The dataset is openly available and well-validated across sub-Saharan Africa, South and Southeast Asia, and Latin America, precisely the regions where piped-water access gaps are largest.
ERA5 adds value in two specific situations. First, where CHIRPS gauge density is very low (parts of Central Africa and island states), ERA5's physically consistent reanalysis can reduce systematic bias. Second, ERA5's hourly resolution allows estimation of storm-intensity distributions, which govern how much of a heavy rain event overflows a small tank before it can be absorbed. A city with the same annual total as another but with more of that total falling in a handful of intense convective events will need larger tank capacity per unit of roof area to capture the same fraction of annual yield. Ignoring intensity is a common error in desk-based harvesting assessments.
Prioritising where piped water does not reach
Harvesting potential only becomes a planning tool when overlaid with access deficits. Piped-network coverage can be approximated from utility records where these exist, or proxied from settlement morphology: informal settlements identified by irregular plot geometry, high building density, and narrow lane widths tend to have lower connection rates. This page does not cover informal settlement detection as a standalone topic, but the output of that analysis feeds directly into the prioritisation layer here.
The combined output is a ranked list of structures or sub-wards ordered by the product of harvestable volume and access deficit score. A large corrugated-iron roof in a high-rainfall, low-access area scores highest. A small flat-concrete roof in a well-served district with low annual rainfall scores lowest. Water authorities can use this ranking to target subsidy programmes, community outreach, or infrastructure investment without conducting expensive household surveys first.
One honest limit: the method estimates potential, not actual practice. A roof in good structural condition with clean guttering will approach the theoretical runoff coefficient. A degraded roof with holes, debris, or no guttering will not. Ground-truthing a sample of high-priority structures remains necessary before committing capital.
Resolution floors and what the method cannot see
At 30 cm resolution, Pleiades Neo can resolve individual rooftop features down to roughly one square metre in area, but the minimum useful building footprint for harvesting analysis is around 20 square metres, below which tank sizing becomes impractical. Very small structures, kiosks, informal lean-tos, are better excluded from the analysis than mapped with false precision.
Cloud cover is the persistent operational constraint. A single clear-sky acquisition is sufficient for the geometric analysis, but in persistently cloudy tropical climates, obtaining a usable stereo pair over a target city can require a tasking window of several weeks. Commercial archive depth helps: WorldView-3 has imaged most large cities multiple times since 2014, and a cloud-free stereo pair may already exist. Where it does not, tasking lead times of two to six weeks are realistic. Satellize's analytics workflow for the Kingdom of Tonga crop-estimation programme has encountered similar cloud-persistence challenges in Pacific island environments, and the same scheduling logic applies here.
Rainfall grids at 5 km cannot capture fine-scale orographic effects within a city. A city on a hillside may have rainfall variation of 20 percent or more across a few kilometres that CHIRPS smooths over. Where topographic relief is significant, downscaling using a digital elevation model and local lapse rates is advisable before computing per-structure yields.
From analysis to a water authority's desk
The deliverable is a GIS layer, typically a polygon feature class with per-building attributes: footprint area in square metres, mean roof slope in degrees, dominant roofing material class, mean annual rainfall in millimetres from CHIRPS, estimated annual harvestable volume in litres under median and 10th-percentile rainfall assumptions, and a composite priority score. This can be served as a web map, exported to a water utility's existing GIS platform, or summarised as a ward-level tabular report for non-GIS users.
Update frequency depends on how rapidly the building stock changes. In fast-growing cities, an annual refresh of the footprint layer against new imagery is reasonable. The rainfall climatology updates incrementally and rarely changes materially year to year. The most expensive input to refresh is the stereo DSM; in practice, a new DSM acquisition every three to five years is sufficient unless a major urban renewal programme has altered the roofscape.
Typical figures
| Primary imagery resolution | 30 cm (Pleiades Neo pan), 50 cm (Pleiades Neo MS), 31 cm (WorldView-3 pan) |
| DSM vertical accuracy | 0.3 to 0.5 m RMSE from tri-stereo matching over urban areas (WorldView-3 published performance) |
| Minimum useful building footprint | ~20 m² for practical harvesting analysis; smaller structures excluded |
| Rainfall grid resolution | ~5 km (CHIRPS 0.05°); ~31 km (ERA5 native) |
| Rainfall record length | 1981 to near-present (CHIRPS); 1940 to near-present (ERA5) |
| Revisit for tasking | ~1 day (Pleiades Neo 4-satellite constellation); ~1 to 4.5 days (WorldView-3 at mid-latitudes) |
| Cloud-free acquisition latency | Days to weeks depending on climate; persistent cloud regions may require multi-week tasking windows |
| Archive depth (commercial imagery) | WorldView-3 from 2014; Pleiades from 2012; Pleiades Neo from 2021 |
| Delivery format | GeoPackage or Shapefile with per-building attributes, GeoTIFF DSM, CSV ward-level summary, optional web-map service |
Analytics Satellize can run
| Building footprint layer with roofing material classification | Deep-learning semantic segmentation (U-Net class architectures) applied to VHR multispectral imagery; spectral indices distinguish metal, tile, concrete, and thatch | Polygon GIS layer with material class and confidence score per structure |
| Photogrammetric DSM and roof-slope raster | Dense image matching from stereo or tri-stereo VHR pairs; slope derived from DSM gradient; ridge orientation from aspect raster | GeoTIFF DSM at 0.5 to 1 m posting; derived slope and aspect rasters |
| Per-structure annual harvestable volume estimate | Catchment yield formula (area × rainfall × runoff coefficient) applied per polygon; runoff coefficients assigned by material class and slope band; CHIRPS percentile statistics as rainfall input | Attribute table with median and 10th-percentile annual yield in litres per structure |
| Rainfall climatology surface (downscaled) | CHIRPS 40-year percentile statistics; optional orographic downscaling using SRTM DEM where relief exceeds 100 m within the study area | GeoTIFF rainfall grids (mean, P10, P90) at native or downscaled resolution |
| Priority ranking layer for subsidy targeting | Composite score combining harvestable volume rank and piped-access deficit proxy (derived from settlement morphology or utility records); scored 0 to 100 per structure | Ranked polygon layer and ward-level summary table exportable to Excel or GIS |
| Change detection for building-stock refresh | Bi-temporal change detection between baseline and update imagery; new structures flagged for re-analysis; demolished structures removed from active inventory | Change polygon layer with status flags; updated master footprint dataset |
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.