abrhamadddis/weatherbench-ethiopia

# WeatherBench · Ethiopia ### Benchmarking ML Weather Forecasting Models over the Horn of Africa --- ## Overview This project uses the WeatherBench2 framework to rigorously evaluate three state-of-the-art ML weather forecasting models over **Ethiopia** — a region of high meteorological importance in East Africa due to its complex topography, monsoon systems, and sensitivity to global climate variability. Models are evaluated against **ERA5 reanalysis** (the standard ground truth in weather forecasting) across four deterministic metrics — RMSE, MAE, Bias, and ACC — over five atmospheric variables at two pressure levels, for the full year 2020 (lead times up to 10 days). --- ## Model Selection The three models — **GraphCast**, **Pangu-Weather**, and **FuXi** — were selected from the **top-performing models on the WeatherBench2 global leaderboard**, representing the current state of the art in ML-based Numerical Weather Prediction (ML-NWP): | Model | Institution | Architecture | WB2 Leaderboard | |---|---|---|---| | **GraphCast** | Google DeepMind | Graph Neural Network | Top-ranked overall | | **Pangu-Weather** | Huawei | 3D Earth Vision Transformer | Top-ranked (deterministic) | | **FuXi** | Fudan University | Cascaded U-Net | Top-ranked (extended range) | > **Note on Resolution:** The **64×32 (~5.6°/grid point)** low-resolution dataset variant was chosen > to make evaluation feasible on limited computational resources. Higher-resolution variants exist but > require significantly more memory and I/O bandwidth. > **Recommended Environment:** Run the notebook on **Google Colab**, > which is co-located with Google Cloud Storage for fast GCS data streaming and provides free compute > resources. A local environment works for `--analyze` (offline), but `--evaluate` requires GCS access. --- ## Key Results at 72h Lead Time > Evaluated over Ethiopia (3–15°N, 33–48°E), averaged across 500 hPa and 850 hPa pressure levels. | Variable | Best RMSE | Best MAE …