Aflatoxin contamination risk mapping via drought-stress proxies
Aflatoxin contamination in maize and groundnut is triggered by drought stress during grain fill, weeks before any visible symptom appears. Satellite-derived stress indices can map that window spatially, giving food-safety programmes a probabilistic risk layer to direct ground sampling.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at the equator with both satellites. Provides NDVI, NDWI and red-edge chlorophyll indices to track canopy condition through grain-fill phenology. Cloud contamination in humid sub-Saharan zones can reduce effective revisit to once every 10-20 days during the rainy-season tail.
- Landsat 8/9 TIRS: Thermal infrared bands 10 and 11 at 100 m native resolution (resampled to 30 m in products), 16-day revisit per satellite, 8-day combined. Land surface temperature derived from TIRS identifies heat-stressed fields through crop water stress index approaches. Acquisition is daytime only, so diurnal temperature dynamics are partially captured.
- SMAP (Soil Moisture Active Passive): L-band passive radiometer at approximately 36 km spatial resolution, 2-3 day revisit globally. Root-zone soil moisture estimates derived from surface retrievals via exponential filter methods give a coarse but consistent signal of moisture deficit. Resolution is far too coarse to resolve individual smallholder fields; the signal is most useful as a district-level anomaly driver.
- MODIS (Terra/Aqua): Daily or near-daily land surface temperature and vegetation condition at 250 m to 1 km. Useful for constructing the historical baseline of Vegetation Condition Index (VCI) and Temperature Condition Index (TCI) that define what 'anomalous stress' means for a given pixel in a given dekad. Archive depth back to 2000 makes percentile baselines statistically meaningful.
Why the fungus wins during drought
Aspergillus flavus is a soil-borne opportunist. Under normal moisture conditions, competing microorganisms and the plant's own defences keep it in check. When a maize or groundnut crop experiences water deficit during grain fill, the plant redirects resources away from defence metabolites, kernel integrity degrades, and A. flavus colonises the developing grain. The aflatoxins it produces are heat-stable; they survive milling, cooking and most processing steps. The critical window is roughly two to four weeks before physiological maturity, which in practice means the stress event that matters most is invisible to any post-harvest inspection.
This is the logic that makes satellite observation useful. The fungus responds to a physical condition, drought stress, that leaves a measurable signature in canopy reflectance and land surface temperature well before harvest. A risk map built on those signals does not detect aflatoxin; it identifies where the environmental preconditions were met. That distinction matters enormously for how the output should be used.
What the indices actually measure, and what they do not
The primary stress signal comes from vegetation condition indices computed from Sentinel-2 time series. NDVI alone is a blunt instrument; red-edge bands (bands 5 and 6 at 20 m) are more sensitive to early chlorophyll loss before visible yellowing. A Vegetation Condition Index compares the current-season NDVI or red-edge index against the multi-year percentile range for the same calendar period, derived from the MODIS archive. A field sitting at the 10th percentile of historical greenness during grain fill is a candidate for elevated risk. It is not proof of contamination.
Land surface temperature from Landsat TIRS adds a second, partially independent line of evidence. Stressed crops transpire less, so their canopy temperature rises above that of well-watered neighbours. The Crop Water Stress Index formalises this comparison against a reference temperature envelope. The limitation is Landsat's 16-day revisit: a single cloud-free acquisition may miss the peak stress event entirely, or catch it after partial recovery following an isolated rain event.
SMAP root-zone soil moisture anomalies provide the district-scale moisture deficit context. Because SMAP's footprint is 36 km, it cannot distinguish one farmer's field from another. Its value is in confirming that a regional drought signal is real rather than a localised artefact in the optical data, and in extending the moisture-deficit record back through the season when optical data were cloud-obscured.
Building a risk layer that a food-safety officer can use
The analytic output is a gridded probability surface, typically at 10 m or 20 m resolution, expressing the relative likelihood that grain-fill stress conditions were met across the landscape. Inputs are combined through a weighted index or, in more sophisticated implementations, a logistic regression trained on historical ground-sampling data from the same agroecological zone. The weights assigned to VCI, TCI and soil moisture anomaly are not universal; they should be calibrated regionally because the relationship between stress timing and aflatoxin outcome varies with cultivar, soil type and the specific phenological calendar.
The risk layer needs ground truth to become actionable. Without paired aflatoxin measurements from the same season, the map is a prioritisation tool for sampling, not a food-safety decision in itself. A practical workflow sends mobile sampling teams to high-risk polygons first, concentrating laboratory capacity where satellite evidence suggests it is most likely to find exceedances above regulatory thresholds (typically 4 ppb total aflatoxin for human consumption under EU regulation, 20 ppb under US FDA guidelines). This is where the satellite product earns its cost: not by replacing laboratory analysis, but by making a fixed sampling budget go further.
Resolution floors, cloud and the smallholder problem
In the maize belts of eastern and southern Africa, and in the groundnut-growing regions of West Africa, fields are frequently smaller than one hectare. At Sentinel-2's 10 m resolution, a 0.5 ha field occupies roughly 50 pixels, which is workable. At MODIS 250 m, the same field is sub-pixel. The practical implication is that SMAP and MODIS can only provide regional context; field-level risk assignment depends entirely on Sentinel-2 and Landsat.
Cloud cover is a genuine problem in humid transition zones during the rainy-season tail, precisely when late-season drought stress is developing. In years with persistent cloud, the number of usable Sentinel-2 acquisitions during the critical grain-fill window can fall to two or three. Temporal compositing and gap-filling using MODIS as a coarse-resolution guide can partially recover the signal, but the uncertainty in the risk estimate widens considerably. Any delivery of this product should carry an explicit data-quality flag showing the number of cloud-free observations used in each pixel's stress assessment.
Calibration, validation and the honest limits of inference
The published literature on satellite-based aflatoxin risk mapping, including work by CIMMYT and IITA in sub-Saharan Africa, consistently shows that stress-proxy indices have moderate predictive skill when calibrated against local ground data, but perform poorly when models trained in one region are applied directly to another. Soil type, cultivar maturity class and the specific timing of the stress relative to phenological stage all shift the relationship. A model trained on Kenyan highland maize is not reliably transferable to Zambian valley-bottom fields without re-calibration.
Satellize builds these risk layers on open constellations, Sentinel-2, Landsat and MODIS, and incorporates SMAP anomalies as a regional moisture anchor. The approach follows the same index-based framework used in the Tonga crop-estimation programme, adapted for a different crop stress mechanism. Clients should plan for at least one season of paired satellite and laboratory data collection before treating the risk surface as operationally validated. The product is a screening tool. It is a good one, but it is not a substitute for the laboratory.
Typical figures
| Primary optical resolution | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge), 30 m (Landsat TIRS resampled) |
| Effective revisit (optical, cloud-free) | 5 days nominal Sentinel-2; degrades to 10-20 days in persistently cloudy humid zones |
| Thermal resolution | 100 m native TIRS, resampled to 30 m; 16-day per Landsat satellite, 8-day combined Landsat 8+9 |
| Soil moisture footprint | ~36 km (SMAP passive L-band); district-level context only, not field-level |
| Spectral bands used | Sentinel-2 B4 (red), B8 (NIR), B5/B6 (red-edge); Landsat TIRS Band 10 (10.6-11.2 µm); SMAP 1.4 GHz L-band |
| Minimum field size reliably assessed | ~0.25 ha at 10 m resolution; sub-hectare fields are at the margin of reliability |
| Historical baseline depth | MODIS archive from 2000; Sentinel-2 from 2015; Landsat from 1982 (TIRS from 2013) |
| Risk layer output format | GeoTIFF probability surface (0-1 relative risk); optional vector polygons by risk class |
| Latency from satellite acquisition | Sentinel-2 Level-2A available within 3-5 days of acquisition via Copernicus Data Space; Landsat within 24 hours via USGS |
| Output validated against | Requires client-supplied ground aflatoxin measurements for regional calibration; no universal threshold is transferable |
Analytics Satellize can run
| Grain-fill stress composite | Sentinel-2 red-edge NDVI and NDWI time-series compositing over the defined grain-fill window, compared against multi-year MODIS percentile baseline to produce Vegetation Condition Index per field polygon | GeoTIFF and field-polygon GIS layer showing VCI percentile rank for the current season's grain-fill period |
| Crop water stress index surface | Land surface temperature from Landsat TIRS compared against empirical upper and lower reference temperature bounds for well-watered and non-transpiring canopies, following published CWSI methodology | 30 m GeoTIFF of CWSI values for each cloud-free Landsat acquisition during grain fill, with temporal maximum stress flagged |
| Regional soil moisture anomaly layer | SMAP Level-3 root-zone soil moisture retrievals differenced against the 2015-present climatological mean for the same dekad; exponential filter applied to propagate surface signal to root-zone depth | District-level anomaly map (percentage of median) delivered as polygon GIS layer, updated every 2-3 days during the season |
| Combined aflatoxin risk probability surface | Weighted index combining normalised VCI, CWSI and soil moisture anomaly scores; weights derived from published CIMMYT/IITA index frameworks and adjusted by client-supplied calibration data where available | 10 m or 20 m GeoTIFF risk surface with per-pixel data-quality flag (number of cloud-free observations used); classified vector layer by risk tier for sampling prioritisation |
| Sampling prioritisation report | Spatial stratification of risk surface into high, medium and low tiers; area statistics by administrative unit; suggested sampling allocation based on proportional-to-risk design | PDF report and accompanying shapefile of recommended sampling locations, ranked by risk tier |
| Season-on-season risk trend | Multi-year stack of end-of-grain-fill risk composites to identify structurally vulnerable zones (fields that recur in the high-risk tier across multiple seasons regardless of rainfall year) | Multi-year GeoTIFF stack and summary statistics table by field or administrative polygon; suitable for input to national food-safety surveillance planning |
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.