Weather Monitoring
Background
Africa faces a critical challenge: a lack of reliable, localized weather forecasts. This data gap directly impacts sectors like agriculture, which forms the backbone of many African economies. While the TAHMO initiative has deployed nearly 800 weather stations across the continent, spatial coverage remains uneven, and significant gaps persist. High-quality, localised data is vital—not only for daily farming decisions but also for climate resilience planning.
To tackle this, we are developing a machine learning-based forecasting pipeline that integrates reanalysis datasets like ERA5 with observational ground truth from TAHMO stations. The approach involves:
- Ingesting and preprocessing ERA5 data via the Copernicus Climate Data Store.
- Running inferences using ML models such as GraphCast.
- Performing retrospective forecasts (hindcasts) for past periods (e.g., 2024) and comparing predictions to actual TAHMO observations to assess accuracy and perform quality control (e.g., detecting outliers via statistical methods).
The broader goal is to power a mobile or web-based app that serves localized forecasts directly to African farmers—helping bridge the information gap in weather awareness.
Accomplishments
- Built a robust data pipeline to download and preprocess ERA5 data from online archives.
- Integrated GraphCast into the pipeline to generate forecasts over the 6-month period period between January 2024 and June 2024 using overlapping sliding windows as a way of evaluating predictions.
- Designed experiments with two different input sets: ERA5-only and ERA5 augmented with TAHMO observations.
- Developed an evaluation framework that compares predictions to ground truth to both validate forecasts and identify anomalous station behavior.
- Implemented a median-based quality control method that uses nearest-neighbor comparisons to flag outlier TAHMO stations—improving the reliability of the ground truth data used for validation.
Next Steps
- Model Spatial Influence Mapping: Investigate how each TAHMO station within an ERA5 grid influences forecasted values at the surrounding grid points. This involves learning a representation based on spatial distance, climatic zones, and other features.
- Pipeline Completion: Finalize the end-to-end pipeline—integrating data extraction, forecasting, evaluation, and delivery—to power an application that can serve predictions to end users, particularly smallholder farmers.