Flood-driven waterborne disease risk mapping
Combining SAR flood extent, land-use classification and population density, satellite analytics identify where inundation is most likely to drive waterborne disease. The output is a spatial prior for field surveillance, not a pathogen forecast.
Sensors
- Sentinel-1 SAR (C-band, ESA): 10 m ground resolution in Interferometric Wide Swath mode, 6-day repeat at the equator (3-day with both satellites). Cloud-transparent, so it captures flood extent during and immediately after the storm events that most commonly trigger outbreaks. The primary source for inundation mapping.
- Sentinel-2 MSI (ESA): 10 m visible and near-infrared bands, 20 m shortwave infrared. Five-day revisit. Used for pre-flood land-use classification: distinguishing open defecation fields, waste-disposal sites, livestock paddocks and informal settlement density. Requires cloud-free acquisitions, which limits utility during active monsoon or cyclone periods.
- PlanetScope (Planet Labs): 3 m resolution, daily revisit in most regions. Useful for identifying latrine pits, waste piles and small livestock enclosures that are below Sentinel-2's classification threshold. Commercial tasking; archive access requires a licence.
- Copernicus DEM GLO-30: Global 30 m digital elevation model derived from TanDEM-X radar. Used to model flood flow direction and identify low-lying areas where contaminated water pools rather than drains. Vertical accuracy is approximately 4 m absolute at 90th percentile, which limits depth precision but is sufficient for flow-path analysis.
Why flooding and disease incidence are spatially predictable
Cholera and typhoid are not randomly distributed after a flood. Both are faecal-oral diseases whose transmission probability depends on what the water has contacted before it reaches people. Open defecation areas, pit latrines, animal holding areas and informal waste dumps are the proximate contamination sources. When floodwater inundates them and then recedes through densely populated low-lying areas, it carries Vibrio cholerae or Salmonella typhi to drinking-water sources, food and hands.
This spatial logic is well-documented. A 2017 analysis of the Yemen cholera outbreak and subsequent peer-reviewed work on Bangladesh, Haiti and Mozambique all show that case incidence clusters in areas combining inundation, poor sanitation infrastructure and high population density. The satellite contribution is not to detect the pathogen; no sensor can do that from orbit. The contribution is to map the spatial overlap of those three risk factors quickly enough to guide where field teams deploy oral rehydration points, water-quality test kits and case-finding capacity.
What the sensors actually see, and what they miss
Sentinel-1 SAR detects open water reliably. Flooded bare soil and flooded low vegetation are also detectable, though with higher false-positive rates in rough-surface conditions. Flooded urban areas and flooded dense forest are the hard cases: double-bounce returns from buildings can mimic dry-surface signals, and forest canopy masks the water beneath it entirely. These are not edge cases in disease-risk contexts; informal settlements and riparian vegetation are precisely where high-risk populations live.
Sentinel-2 land-use classification can resolve latrine compounds and waste sites at 10 m, but only from pre-flood imagery acquired under cloud-free conditions. During a monsoon-season flood, the most recent usable Sentinel-2 scene may be weeks old. PlanetScope's 3 m resolution helps identify small structures that Sentinel-2 misses, but it shares the cloud problem. The practical workflow therefore uses pre-flood optical classification as a static contamination-source layer, overlaid on near-real-time SAR inundation extent.
Population density is drawn from gridded products such as WorldPop or the Global Human Settlement Layer, both at 100 m or finer resolution. These are modelled estimates, not census counts, and they can misrepresent rapidly growing informal settlements. The risk index inherits that uncertainty.
Building the risk index: method and honest limits
The core analytic combines three raster layers: binary or probability-weighted flood extent from SAR, a contamination-source proximity score from optical land-use classification, and population density. Each cell in the output grid receives a composite score. The weighting between layers is a modelling choice, and different published frameworks (WHO WASH cluster guidance, the UNICEF flood-health rapid assessment method) make different choices. No single weighting scheme has been validated across all geographies.
The output is a spatial prior: a ranked map of where disease risk is elevated relative to the rest of the affected area. It is not a case-count forecast and should not be communicated as one. Epidemiologists use it to allocate scarce surveillance capacity, not to predict incidence rates. This distinction matters operationally. A high-scoring grid cell means 'send a team here first', not 'expect X cases here'.
Temporal latency is a genuine constraint. From satellite overpass to a processed risk layer, a well-resourced pipeline takes 12 to 48 hours depending on SAR processing queue, data access and analyst capacity. For a fast-moving flood, the inundation footprint may have changed materially in that window. Communicating the acquisition timestamp alongside the risk layer is not optional.
Elevation data as the connective tissue
The Copernicus DEM GLO-30 plays a quieter but important role. Flood extent from SAR is a snapshot of where water was at the moment of the overpass. DEM-based flow modelling extends that snapshot: it identifies connected low-lying areas that were likely inundated between passes, and it flags where water will pool as the flood recedes rather than drain freely. Pooled residual water is disproportionately dangerous because it concentrates contaminants and persists long enough for transmission chains to establish.
At 30 m horizontal resolution and roughly 4 m vertical accuracy, GLO-30 is adequate for catchment delineation and broad flow-path analysis in flat deltaic terrain. In steep or highly urbanised areas, the vertical error is large relative to the flood depths involved, and the flow model becomes unreliable. LiDAR-derived DEMs at sub-metre vertical accuracy exist for some cities but are rarely available in the lower-income settings where waterborne disease risk is highest.
From data layer to field decision
The risk map is most useful when it reaches public-health coordinators within the first 72 hours of inundation, before transmission chains are established. That requires pre-agreed data-sharing arrangements, a pre-built processing pipeline, and field teams who understand what the map represents. A risk layer delivered on day five to a team that has already deployed by intuition adds little value.
Satellize runs this analytic stack on open Sentinel and Copernicus data, with optional PlanetScope tasking where sub-10 m resolution is needed to resolve sanitation infrastructure. The methodology is the same used in published WASH-sector rapid assessments; what Satellize adds is processing speed and the ability to integrate client-supplied ground data (latrine GPS logs, waste-site surveys) into the contamination-source layer. The Tonga crop-estimation programme demonstrated the same principle in a different domain: satellite-derived spatial priors calibrated against ground-truth data produce more actionable outputs than either source alone.
Field teams should treat the risk index as a triage tool, not a substitute for water-quality testing. The index tells you where to test first. The test tells you whether to act.
Typical figures
| Primary flood-extent resolution | 10 m (Sentinel-1 IW mode) |
| SAR revisit (flood monitoring) | 3 to 6 days (Sentinel-1A/B combined); shorter with commercial SAR tasking |
| Land-use classification resolution | 10 m (Sentinel-2) or 3 m (PlanetScope, licensed) |
| Elevation model resolution | 30 m horizontal, ~4 m absolute vertical accuracy (GLO-30) |
| Minimum detectable flooded area | Approximately 0.1 ha in open terrain; larger in urban or forested areas |
| Processing latency (overpass to risk layer) | 12 to 48 hours depending on pipeline and data-access arrangements |
| Sentinel archive depth | Sentinel-1 from 2014; Sentinel-2 from 2015 |
| Output formats | GeoTIFF risk-index raster, GeoJSON contamination-source polygons, PDF situation report |
| Population density source | WorldPop or Global Human Settlement Layer, 100 m grid (modelled estimates) |
Analytics Satellize can run
| SAR flood-extent layer | Thresholded backscatter change detection on Sentinel-1 GRD, with reference to pre-event baseline | GeoTIFF binary or probability-weighted inundation mask, timestamped to overpass |
| Contamination-source classification | Supervised land-use classification on Sentinel-2 or PlanetScope pre-flood imagery, targeting latrine compounds, waste sites and livestock areas | GeoJSON polygon layer with source-type labels and confidence scores |
| Composite disease-risk index | Weighted raster overlay of flood extent, contamination-source proximity and WorldPop population density, following published WASH rapid-assessment frameworks | GeoTIFF risk-index raster (0 to 1 normalised), plus ranked priority-zone polygons for field deployment |
| Flood recession and pooling analysis | Multi-date SAR time series combined with GLO-30 flow-accumulation modelling to identify persistent water bodies post-peak | GeoTIFF showing residual water extent at 24 h intervals post-peak, with pooling-zone highlights |
| Risk-index change report | Differencing of consecutive risk-index rasters as flood extent evolves | PDF situation report with change maps, issued at each new SAR acquisition |
| Ground-truth integration layer | GPS-logged field data (latrine surveys, waste-site coordinates) merged into contamination-source raster to improve local accuracy | Updated GeoTIFF contamination-source layer incorporating client-supplied field observations |
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.