Skip to content
GeoVitaIntelligence

Live demo · Africa

Malaria endemicity on a living globe.

Annual Plasmodium falciparum parasite rate in children aged 2–10 (PfPR₂–₁₀) across 1.46 million grid cells — observed by the Malaria Atlas Project and predicted by GeoVita's ConvLSTM2D model from nine years of environmental history.

PfPR₂–₁₀ measures endemicity (prevalence), not clinical incidence. The MAP surfaces are model-based references; forecasts complement, but never replace, epidemiological surveillance.

Model benchmark

Six trained models, two honest baselines.

Every approach was tested twice: on geographically separated 250 km blocks (spatial transfer) and on an untouched 2022–2024 period (future-year forecasting).

01

Best spatial transfer

ConvLSTM2D ranked first on unseen regions: RMSE 0.022, r 0.993, NSE 0.986.

02

A demanding benchmark

On 2022–2024, simple persistence ranked first overall; ConvLSTM2D was the best trained model, with no clear RMSE difference between them.

03

Humidity, wind, canopy water

Specific humidity, wind speed, canopy interception and air temperature carried the most predictive signal; the previous year mattered most.

RMSE by model — Spatial CV

Lower is better

  • Model
  • Baseline
ConvLSTM2D0.0218Persistence0.0331CNN-LSTM0.0672Random Forest0.0675GRU0.0759XGBoost0.0792LSTM0.0894Historical Mean0.1057

Forecast error by horizon (RMSE)

ConvLSTM2D leads one year ahead; the persistence baseline catches up as the horizon grows.

  • ConvLSTM2D
  • Persistence
  • Other models
0.0000.0250.0500.0750.100H1 · 2022H2 · 2023H3 · 2024Years after last observed surfaceConvLSTM2DPersistence

What drives the ConvLSTM2D forecast

Increase in RMSE when each predictor is shuffled (permutation importance). Predictive dependence, not causation. ΔRMSE ×10⁻³.

Specific humidity2.95Wind speed2.48Canopy interception2.43Air temperature2.06NDVI1.38Precipitation1.22Evapotranspiration1.18Soil moisture0.44Shortwave radiation0.12Surface temperature0.09

Which years matter most

Increase in RMSE when all predictors in one lag year are shuffled. ΔRMSE ×10⁻³.

t−15.38t−22.18t−32.63t−40.75t−50.38t−60.48t−70.44t−80.51t−9-0.06

Case study · high resolution

Finding hidden larval habitats at 10 cm.

In Qaleh-Ganj, south-east Iran, a digital surface model from 189 UltraCam aerial images was fused with land-use and soil-type maps to locate water-holding depressions where Anopheles larvae can develop within mosquito flight range of the town.

  1. 01Aerial stereo images → 8.1 billion-point dense cloud → DSM & orthomosaic
  2. 02Depression extraction with depth, area and volume
  3. 03SVM land-use & soil classification (silt and clay beds hold water)
  4. 04Filter to a 1.5 km mosquito flight buffer → ranked high-risk map
image resolution
10 cm
high-risk depressions identified
223
long-lasting field habitats matched
12 / 14
classification kappa
0.93

Coarse SRTM, NDVI and NDWI layers missed most temporary habitats — evidence for why GeoVita pairs satellite screening with high-resolution aerial and UAV assessment.

Youssefi et al. (2022), International Journal of Applied Earth Observation and Geoinformation.

Read paper

Want this for your region?

We co-design pilots with local partners, data and validation.

Propose a pilot