Geogenic arsenic groundwater risk mapping via surface geology proxies
Naturally occurring arsenic in alluvial aquifers cannot be seen from orbit, but the sedimentary and redox conditions that concentrate it leave spectral fingerprints. Satellite-derived geology proxies produce probabilistic risk surfaces that guide where to test, not what the water contains.
Sensors
- Landsat 8/9 OLI/TIRS: 30 m multispectral resolution, 16-day revisit per satellite (8-day combined). OLI bands 2-7 support clay mineral and iron-oxide indices; TIRS thermal bands at 100 m (resampled to 30 m) provide land surface temperature as a soil moisture and organic matter proxy.
- Sentinel-2 MSI: 10-20 m multispectral resolution across 13 bands, 5-day revisit at the equator. Red-edge bands (B5, B6, B7) and SWIR bands (B11, B12) sharpen clay mineralogy mapping and vegetation indices used to infer organic carbon accumulation in reducing environments.
- SRTM DEM: ~30 m (1 arc-second) global elevation model from the 2000 Shuttle Radar Topography Mission. Provides topographic wetness index, slope and drainage accumulation layers that identify low-lying depositional zones where fine-grained, arsenic-bearing sediments preferentially settle.
- MODIS Terra/Aqua: 250-500 m resolution, daily global coverage. MODIS surface reflectance and land surface temperature time series serve as coarse soil moisture and vegetation productivity proxies, useful for characterising seasonal inundation patterns across large river basins where fine-resolution data are computationally expensive.
Why geology at the surface predicts chemistry at depth
Geogenic arsenic in groundwater is not uniformly distributed. It concentrates in young alluvial aquifers, particularly in the Bengal Basin, the Red River Delta and the Mekong lowlands, where two conditions coincide: fine-grained sediments rich in iron oxyhydroxides that adsorb arsenic during weathering, and strongly reducing groundwater chemistry that causes those oxyhydroxides to dissolve and release arsenic into solution. Both conditions are linked to depositional environment, sediment age and organic carbon content, all of which have surface expressions.
The surface expressions are subtle but real. Low-lying floodplain positions with high topographic wetness indices accumulate finer sediments and sustain longer waterlogging, promoting reducing conditions below. High soil organic carbon, inferred from spectral indices and vegetation productivity, drives microbial oxygen consumption that deepens reducing zones. Iron-oxide abundance in surface soils, detectable through band ratios in Landsat OLI and Sentinel-2 SWIR bands, correlates with the iron-rich parent material that carries arsenic into the aquifer. None of these signals measures groundwater chemistry. Together, they constrain where high arsenic is probable.
What the sensors actually measure, and what they do not
Landsat OLI's clay mineral index (typically the ratio of SWIR band 6 to SWIR band 7, or equivalent Sentinel-2 B11/B12 combinations) responds to hydroxyl-bearing minerals including kaolinite and smectite. The iron-oxide index (OLI band 4 divided by band 2) highlights ferruginous surface soils. These are established geological remote-sensing techniques, documented in USGS and ESA literature, and they work best on bare or sparsely vegetated soil. Dense rice paddy cultivation, which covers precisely the high-risk zones in South and Southeast Asia, suppresses the soil signal for much of the year. Seasonal compositing, using Landsat or Sentinel-2 scenes acquired during the dry season when fields are fallow, partially recovers it.
SRTM topography is more consistently interpretable. Topographic wetness index values above roughly 8 to 10 (dimensionless, calculated as ln(catchment area / tan slope)) reliably identify depositional flats and backswamps in published studies of the Bengal Basin. MODIS-derived soil moisture anomalies add a seasonal dimension. The honest limit of the whole approach is that it produces a spatial prior: a probability surface saying arsenic exceedance above the WHO guideline of 10 micrograms per litre is more or less likely here than there. It cannot replace a single well-water test.
Published methods and what they have demonstrated
Several peer-reviewed studies, published in journals including Remote Sensing (MDPI) and Environmental Science and Technology, have combined these inputs into logistic regression and random-forest classifiers trained on well-water arsenic measurements from national databases in Bangladesh, Cambodia and Vietnam. Reported area-under-curve values for cross-validated models have generally fallen in the 0.75 to 0.85 range, meaning the satellite-derived risk surface correctly ranks high-risk versus low-risk locations substantially better than chance, but far from perfectly. The models perform best where training well data are dense and worst in geologically heterogeneous areas where shallow and deep aquifers are spatially interleaved at scales below the sensor resolution.
Lithological maps, where available at 1:250,000 or finer, add explanatory power that pure spectral indices cannot match. The combination of published geological survey data with satellite-derived indices consistently outperforms either input alone. This is not a surprise: the satellite sees the surface, the geological map records what surveyors found below it. The two are complementary, not redundant.
Honest limits: resolution floors, cloud and the validation gap
Cloud cover is a genuine operational problem. The Bengal Basin and Mekong Delta experience monsoon cloud cover for five to seven months per year. Optical sensors are blind through cloud. Dry-season compositing is the standard mitigation, but it means the soil signal captured is from a different season than peak inundation, which is when reducing conditions are most active. Synthetic aperture radar from Sentinel-1 can characterise inundation extent through cloud, but SAR backscatter does not directly measure the mineralogical properties that matter for arsenic prediction.
Spatial resolution is a second constraint. At 30 m, Landsat cannot resolve individual hand-dug wells, which may tap aquifers with sharply different chemistry over distances of tens of metres. Published studies consistently find that the satellite-derived risk surface is most reliable at the village or sub-district scale and becomes unreliable as a guide to individual well safety. Any programme that uses these maps to prioritise testing must communicate this clearly to field teams and to the communities being served. The risk surface tells you which villages to visit first; it does not tell you which well to drink from.
Building a risk surface: from raw imagery to a probabilistic map
A practical workflow starts with cloud-free dry-season composites from Landsat 8/9 or Sentinel-2, typically median composites over two to three dry seasons to reduce noise. From these, analysts derive the clay mineral index, iron-oxide index, normalised difference vegetation index (NDVI) as an organic carbon proxy, and land surface temperature from Landsat TIRS as a secondary soil moisture indicator. SRTM-derived topographic wetness index and slope are added as static layers. Published lithological data, where available, are encoded as categorical inputs.
These layers feed a classifier, most commonly a random forest or gradient-boosted model, trained on arsenic measurements from existing well surveys. Output is a continuous probability surface of exceedance above the WHO 10 microgram per litre threshold, delivered at 30 m or 10 m grid resolution depending on the primary optical input. The surface is then aggregated to administrative units for public-health prioritisation, with uncertainty bounds derived from cross-validation held-out predictions. Satellize applies this workflow on open Landsat and Sentinel archives; the Tonga crop-estimation programme uses similar multi-index compositing logic, adapted here to a very different geochemical question.
Validation against an independent well-testing dataset is not optional. It is the only honest way to know whether the model is useful in a new basin. A risk map without a validation plan is a hypothesis, not a product.
From map to action: what a government or NGO does next
A probabilistic risk surface has immediate operational value even before a single additional well is tested. It allows a health ministry or NGO to rank sub-districts by expected burden, concentrating scarce field-testing kits on the areas where exceedance is most probable. In a country like Bangladesh, where an estimated 20 million people may be exposed to arsenic above the WHO guideline (a figure drawn from published epidemiological literature, not from satellite data alone), the difference between random sampling and risk-stratified sampling can mean thousands of additional cases identified per testing campaign.
Once field testing returns results from the prioritised areas, those measurements feed back into the model as additional training data, improving the next iteration of the risk surface. This is a standard active-learning loop in spatial epidemiology. The satellite layer does not replace the well-testing programme. It makes the testing programme cheaper and faster by reducing the search space.
Typical figures
| Primary optical resolution | 10-30 m (Sentinel-2 at 10-20 m; Landsat 8/9 at 30 m) |
| Topographic input resolution | ~30 m (SRTM 1 arc-second global DEM, 2000 acquisition) |
| Revisit for optical compositing | 5-day (Sentinel-2 combined); 8-day (Landsat 8+9 combined) |
| Usable dry-season window (South/SE Asia) | Typically November to April; cloud fraction below 20% in most years |
| Spectral bands used for mineralogy | Landsat OLI bands 4, 6, 7; Sentinel-2 B5, B6, B7, B11, B12 |
| Thermal band for soil moisture proxy | Landsat TIRS bands 10-11 at 100 m native (resampled to 30 m) |
| Archive depth | Landsat from 1972 (TM from 1982); Sentinel-2 from 2015; SRTM static 2000 |
| Output risk surface resolution | 30 m grid (Landsat-based) or 10 m grid (Sentinel-2-based); aggregated to admin units on request |
| Typical model performance (published range) | AUC 0.75-0.85 in cross-validated studies; degrades in geologically heterogeneous areas |
| Delivery formats | GeoTIFF probability raster, GeoPackage admin-unit summary, PDF technical report with uncertainty bounds |
Analytics Satellize can run
| Dry-season spectral composite | Median compositing of cloud-masked Landsat 8/9 OLI or Sentinel-2 MSI scenes across 2-3 dry seasons; standard USGS/ESA surface reflectance products as input | GeoTIFF multi-band composite at 10-30 m, clipped to area of interest, ready for index derivation |
| Clay mineral and iron-oxide index layers | Published band-ratio indices (SWIR ratio for clay; visible-to-NIR ratio for iron oxide) applied to spectral composite; validated against geological survey literature | GeoTIFF index rasters with documented band equations and value ranges |
| Topographic wetness index surface | SRTM DEM hydrological conditioning, flow accumulation and slope calculation using standard GIS methods; ln(catchment area / tan slope) formulation | GeoTIFF TWI raster at 30 m, with drainage network and depositional flat polygons extracted at user-specified threshold |
| Probabilistic arsenic-exceedance risk surface | Random forest or logistic regression classifier trained on existing well-arsenic measurements; features include spectral indices, TWI, MODIS soil moisture proxy and lithological class; cross-validated with held-out well data | GeoTIFF probability-of-exceedance raster (WHO 10 µg/L threshold) with uncertainty band rasters; admin-unit summary table ranked by estimated burden |
| Field-testing prioritisation map | Risk surface aggregated to village or sub-district polygons; population-weighted ranking using gridded population data (e.g. WorldPop); stratified sampling design output | GeoPackage with ranked administrative units, recommended sample sizes per stratum, and printable field maps at 1:50,000 |
| Model update after field validation | New well-test results ingested as additional training points; model retrained and performance metrics recalculated; change surface showing where risk estimates shifted | Updated probability raster and technical note documenting AUC improvement and spatial shifts in high-risk zones |
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.