What Milestone 2 means​

The prototype phase (March–August 2026) is where the EOHA platform transitions from a design artefact into a running system. The core deliverables are:

• Automated GEE pipeline — a Python-triggered Cloud Function that pulls Sentinel-2 and MODIS imagery on an 8-day cycle, runs zonal statistics against ward boundaries, and writes directly to the Ward Risk store without manual intervention.
• Server-side risk scoring — calculateDynamicRisk() moves from the browser to a Cloud Run service, enabling real-time scoring for all 9 Limpopo municipalities and laying the groundwork for multi-country expansion.
• XGBoost ML forecasting — the first live run of the Malaria Risk Forecast pipeline against the baseline dataset, producing outbreak_probability_score values per zone per prediction period.
• NCD live data — replacing the static NCD baseline with live GEE-computed scores using Copernicus/Sentinel atmospheric data (NO2, SO2, aerosol index, UV radiation).

Design artefacts carried into M2​

The Design Specification Document (v2) produced in M1 defines the full entity model the prototype must implement:

What's shipping first​

Our first sprint targets the Malaria pipeline end-to-end: GEE ingest → zonal statistics → XGBoost score → alert threshold check → USER_RISK_ALERT insert. Once that loop is closed on Limpopo data, we extend to NCD and begin multi-country boundary ingestion.

Follow progress on the Milestones page and in the live platform.

Malaria gets the satellite imagery attention. But cardiovascular disease, diabetes, and chronic respiratory conditions kill far more people annually across Southern Africa — and their environmental drivers (air quality, heat stress, urban sprawl, lack of green space) are all measurable from orbit.

What we're tracking​

The NCD risk module currently maps four countries: South Africa, Botswana, Zimbabwe, and Lesotho. The variables we pull from GEE are:

Air quality — We use Sentinel-5P TROPOMI's tropospheric NO₂ column as a proxy for ambient air pollution. In the Gauteng region and Bulawayo corridor, concentrations are consistently elevated compared to rural areas.

Urban heat stress — Land surface temperature deviation from the seasonal baseline. Cities absorb and retain heat; high chronic heat exposure correlates with hypertensive and cardiovascular outcomes.

Green space access — Sentinel-2 NDVI as a proxy for access to parks and vegetated areas. Low NDVI in high-density urban areas signals limited green space — a risk factor for sedentary lifestyle and mental health outcomes.

Built-up density — ESA WorldCover Class 50 (built-up) as a proportion of the ward area. This directly maps urban sprawl and the shift toward car-dependent, sedentary lifestyles.

The challenge: scale​

Malaria risk is meaningful at ward level. For NCDs, the relevant unit of analysis shifts depending on what you're measuring. Air quality needs a city-region level because pollution disperses. Healthcare access is highly local — a single clinic can serve a community of 10,000 within 5 km but nothing beyond. Obesity prevalence comes from surveys, not satellites.

Our current prototype solves this by working at the country → province → municipality hierarchy, with the understanding that finer-grained NCD data will be layered in as Milestone 2 progresses.

What the prototype shows​

The four-country choropleth already reveals some expected patterns. Zimbabwe's composite NCD risk score (63%) is driven largely by air quality issues in Harare and Bulawayo combined with limited formal healthcare infrastructure. South Africa's 58% reflects high urban density and lifestyle risk in metros offset by relatively stronger healthcare access. Lesotho and Botswana both sit in the moderate range.

These are rough starting values for the prototype. They'll be replaced by computed, ward-level scores in Milestone 2.

The interactive map​

The NCD dashboard lets you click into South Africa, select a province, and zoom to municipality level. The current data shown in the environmental panel (population density, AQI, healthcare access, obesity prevalence) is illustrative — static values while we wait for the full GEE extraction pipeline to be wired up.

But the infrastructure is working: the GADM ADM1 and ADM3 GeoJSON files are loading, the Leaflet layers are toggling correctly, and the dropdown sync is functioning.

What's coming in Milestone 2​

• Live GEE-computed NCD scores at ward level for all four countries
• Sentinel-5P TROPOMI NO₂ layer visualised directly on the map
• Time slider for NCD trend analysis (2023–2026)
• Validation layer showing SADHS survey-derived prevalence data alongside the EO-derived risk scores

The goal is a tool that a provincial health department can actually use to prioritise where to deploy NCD screening programmes — not just another nice map.

What we built​

The core of Milestone 1 is proving that a satellite-to-dashboard pipeline actually works — not in theory, but with real data, real boundaries, and real health-relevant indices. Here's what we have:

• Monthly snapshots for every ward in Limpopo's 9 municipalities
• 13 variables per ward per month: surface temperature, soil moisture, NDWI, NDVI, agricultural coverage, population density, and habitat class
• A live Leaflet dashboard that lets you drill down from province → municipality → ward and see a composite malaria risk score

The risk engine is straightforward but grounded in the epidemiology: Anopheles mosquitoes need moisture, warm temperatures (25–30°C), and standing water. If a ward has all three, it gets flagged high. If the soil is dry, it stays green.

The GEE extraction process​

The honest version: the first extraction attempt produced a dataset with mismatched column names, coordinate precision issues, and a few municipalities where the ward centroids fell slightly outside the boundary polygons. We fixed all of that before the data hit the dashboard.

The key decisions we made:

1. Monthly median composites — single dates are too noisy in Limpopo's rainy season. Medians of all cloud-free scenes per month are much more stable.
2. ADM3-level aggregation — we're using GADM ward polygons as the unit of analysis. This keeps the data at a resolution that's meaningful for district health offices.
3. Safe CSV export — we named the file _Safe for a reason. We removed any columns that could indirectly identify individuals when combined with external data.

What the data shows​

Even at this early stage, some patterns are clear. January and February 2025 show markedly higher soil moisture across Greater Giyani and Thulamela wards — which matches the historical malaria burden in those areas. Surface temperatures peak in October–November 2025, right on schedule.

We're not making causal claims yet. The validation work happens in Milestone 3. But the EO signals are picking up the right seasonal patterns, which is encouraging.

What's next​

Before Milestone 1 closes in March 2026, we need to:

• Finalise the architecture documents (Annexure A and B) for submission to TuksNovation
• Complete a wireframe review session with the NICD team
• Begin scoping the live GEE → API pipeline for Milestone 2

The NCD module is also taking shape — the Southern Africa choropleth is working and the province drill-down for South Africa is functional. Expanding to real ward-level NCD data is a Milestone 2 task.

If you're working with EO data for health applications in Southern Africa and want to compare notes, open an issue on GitHub or reach out directly.

Google Earth Engine gives you petabyte-scale access to satellite imagery. The challenge isn't getting the data — it's getting the right data at the right spatial and temporal resolution, then wrangling it into something a health surveillance dashboard can actually use.

Here's exactly what we did.

Setting up the boundary​

Everything aggregates up to GADM ADM3 ward polygons. We load Limpopo's wards directly from the FAO GAUL dataset in GEE:

import ee
ee.Initialize(project='your-project-id')

limpopo = (ee.FeatureCollection('FAO/GAUL/2015/level2')
    .filter(ee.Filter.eq('ADM1_NAME', 'Limpopo')))

This gives you ~567 ward polygons. We add a WardLabel property by concatenating the ADM2 code with a sequential ward number — that becomes the primary key throughout the dataset.

NDWI from Sentinel-2​

The Normalised Difference Water Index picks up surface water and waterlogged soils:

def add_ndwi(image):
    ndwi = image.normalizedDifference(['B3', 'B8']).rename('NDWI')
    return image.addBands(ndwi)

s2 = (ee.ImageCollection('COPERNICUS/S2_SR_HARMONIZED')
    .filterDate('2025-01-01', '2025-01-31')
    .filterBounds(limpopo.geometry())
    .filter(ee.Filter.lt('CLOUDY_PIXEL_PERCENTAGE', 20))
    .map(add_ndwi))

ndwi_monthly = s2.select('NDWI').median()

Values above −0.1 suggest meaningful surface water presence — a key condition for Anopheles breeding habitat.

Land surface temperature from MODIS​

MODIS MOD11A2 gives us 8-day composites at 1 km. We take the monthly mean and convert from the raw DN to Celsius:

lst = (ee.ImageCollection('MODIS/061/MOD11A2')
    .filterDate('2025-01-01', '2025-01-31')
    .select('LST_Day_1km')
    .mean()
    .multiply(0.02)
    .subtract(273.15)
    .rename('LST_Surface_C'))

The 25–30°C window is where Anopheles gambiae and A. arabiensis thrive. LST values outside that range contribute 0 to the risk score.

Soil moisture from SMAP​

The SMAP 10 km product gives us a direct volumetric soil moisture estimate:

smap = (ee.ImageCollection('NASA/USDA/HSL/SMAP10KM_soil_moisture')
    .filterDate('2025-01-01', '2025-01-31')
    .select('ssm')
    .mean()
    .rename('Soil_Moisture'))

Soil moisture above 0.35 m³/m³ is our high-risk threshold (+40 points in the risk model). This is the single highest-weighted variable because saturated soils are where larvae actually develop.

Aggregating to ward level​

Once we have all the band composites, we stack them and run reduceRegions:

composite = ndwi_monthly.addBands(lst).addBands(smap)

ward_stats = composite.reduceRegions(
    collection=limpopo,
    reducer=ee.Reducer.mean(),
    scale=1000  # matches MODIS resolution
)

Then export to Drive:

task = ee.batch.Export.table.toDrive(
    collection=ward_stats,
    description='limpopo_jan2025',
    fileFormat='CSV',
    selectors=['WardLabel', 'ADM2_NAME', 'NDWI', 'LST_Surface_C', 'Soil_Moisture']
)
task.start()

One thing that tripped us up​

The scale parameter in reduceRegions matters a lot. If you're mixing Sentinel-2 (10 m) and MODIS (1 km) bands in the same composite, GEE will resample everything to your chosen scale. We use 1000 m because SMAP is coarse — running at 10 m scale would be false precision and would time out on the export anyway.

For the ESA WorldCover land cover layer (10 m native), we run a separate reduction at scale=10 and join it in afterward.

What we do with the output​

The exported CSV gets light post-processing in Python (fixing column names, computing Agric_Percentage as the fraction of Class 40 pixels, clipping any implausible LST values) and then lands in the platform's /data/ directory as Limpopo_Risk_Jan25_Jan26_Safe.csv.

PapaParse loads it at runtime in the browser. No backend required — at least for the prototype.

In Milestone 2, this export → post-process → upload cycle will be replaced by a Cloud Function that runs automatically when a new month's worth of GEE data is ready.