T31SHA/Drought-Forcasting_Prediction

A production-grade machine learning pipeline delivering 1-, 3-, and 6-month SPEI drought forecasts for Algeria's Oran region — with 95.1% explained variance and 98.3% directional accuracy. # Drought-Forcasting_Prediction # 🌵 SPEI Drought Forecasting & Intelligence Platform > Multi-horizon drought prediction for Northwestern Algeria using 76 years of climate data. --- ## Overview This project builds a complete end-to-end machine learning pipeline to **forecast drought conditions** using the **Standardised Precipitation-Evapotranspiration Index (SPEI)** at a grid point in Northwestern Algeria (35.75°N, 0.75°E). The dataset spans **January 1950 to February 2026** — 914 monthly observations across 48 SPEI accumulation windows. The primary target is **SPEI-12**, the 12-month accumulation window and the international standard for hydrological drought monitoring. --- ## Live Demo **→ drought-intelligence.streamlit.app** Upload the CSV file and the full pipeline — data exploration, model performance, uncertainty quantification, and feature intelligence — runs automatically in the browser. --- ## What is SPEI? SPEI is a dimensionless z-score measuring the balance between precipitation and atmospheric water demand. Values follow N(0,1) by construction. | SPEI-12 Value | Classification | |:---:|:---| | +1.0 | 🔵 Wet conditions | --- ## Repository Structure ``` Drought-Forcasting_Prediction/ │ ├── drought_forecasting_.ipynb # Full ML pipeline (10 stages) ├── app.py # Streamlit interactive dashboard ├── SPEI_0.75_35.75.csv # Raw SPEI data (Jan 1950 → Feb 2026) ├── model_results.pkl # Trained model objects + predictions ├── metrics_comparison.csv # RMSE, MAE, R², MAPE, Bias, Hit Rate └── classification_report.csv # WMO drought class precision/recall/F1 ``` --- ## Pipeline — 10 Stages | Stage | Description | |:---:|:---| | 1 | Data ingestion & datetime parsing | | 2 | Exploratory data analysis — 6 publication-quality plots | | 3 | Feature engineering — lags, rolling statistics, interactions | | 4 | Chronological train / validation / test split | | 5 | Model training — Ridge, XGBoost …