RayDalio10/aurora-nigeria-dust

Harmattan dust forecasting and validation over Nigeria using the Aurora foundation model. # Aurora Nigeria Dust — Harmattan Dust Forecasting and Validation Evaluating whether the Aurora Earth-system foundation model can forecast Harmattan dust over Nigeria with enough skill to support early warning. ## Summary Using the pre-trained Aurora 0.4-degree air-pollution model (inference only), this project forecasts dust (PM10) over Nigeria from CAMS input data and validates the forecasts against independent satellite observations (MODIS aerosol optical depth; Sentinel-5P Absorbing Aerosol Index). Multi-season benchmark result (two Harmattan seasons, 2023/24 and 2024/25; 26 forecast dates): Aurora shows statistically significant dust-forecast skill at 1-2 day lead - Day 1 Pearson r = 0.79 (95% CI 0.42-0.95), Day 2 r = 0.55 (95% CI 0.15-0.84) - with skill not distinguishable from zero beyond about two days. At a 2-day lead the system detects 75% of dust events with a 25% false-alarm ratio. This defines a useful early-warning horizon of about 1-2 days. Results are from a pilot benchmark (n = 26) with wide confidence intervals that would tighten with a larger multi-season sample. A single-season case study (January 2024, n = 10) gave a higher but sample-fragile Day-1 correlation (~0.93); the multi-season benchmark above is the robust, defensible result. ## Data sources - CAMS global atmospheric-composition forecasts (Copernicus ADS) - model input and dust-AOD reference. - Aurora checkpoint aurora-0.4-air-pollution.ckpt (Hugging Face microsoft/aurora). Weights are CC-BY-NC-SA (non-commercial). - MODIS MOD04_L2 AOD (NASA Earthdata) - Dark Target / Deep Blue Combined product. - Sentinel-5P Absorbing Aerosol Index (Google Earth Engine). ## Setup This project requires Python, a GPU-enabled runtime for Aurora inference, and access to three open data services: Copernicus ADS, NASA Earthdata, and Google Earth Engine. ### 1. Clone the repository bash git clone github.com cd aurora-nigeria-dust ### 2. Install dependenc …