Drought Early Warning System
Probabilistic drought forecasting using multi-source data fusion and Kalman filtering
An automated early warning platform for anticipatory humanitarian action. The system ingests data from multiple sources — satellite vegetation indices, rainfall records, food prices, soil moisture, and seasonal forecasts — and fuses them using a Kalman filter-based state space model to produce probabilistic drought risk forecasts.
The Problem
Humanitarian organisations rely on manual, retrospective trigger scores to decide when to release pre-positioned aid. By the time a crisis is confirmed, the 6–12 week window needed for anticipatory action has already closed. The system replaces that manual process with an automated, forward-looking forecast.
Features
- Automated data ingestion from satellite and climate APIs
- Kalman filter fusion of heterogeneous, irregularly-observed indicators
- Probabilistic output — full distributions, not point estimates
- 1, 3, and 6-month forecast horizons
- Explicit uncertainty quantification that varies with data availability
- Structural break detection for climate regime shifts
Live App
earlywarning.gabriel-oduori.com
Technologies
- Python (state space models, Kalman filtering)
- Streamlit
- Multi-source remote sensing and climate data
Pilot deployment: Zimbabwe. Built on Harvey’s Structural Time Series framework.