LiganiumInc/Drought-Forecasting-Benin

Code for the paper : Prediction and uncertainty quantification of drought in North Benin # Prediction and Uncertainty Quantification of Drought in North Benin ## 📌 Objective This work aims to develop an **uncertainty-aware drought forecasting framework** for six key localities in the Alibori department of North Benin — **Banikoara, Gogounou, Kandi, Karimama, Malanville, and Segbana**. Our goal is twofold: 1. **Predict drought conditions** using state-of-the-art machine learning and deep learning models based on the Standardized Precipitation Index at a 6-month scale (SPI-6). 2. **Quantify prediction uncertainty** using the Ensemble Batch Prediction Interval (EnbPI) method, enabling more informed and trustworthy decision-making. --- ## 📥 Installation & Setup Follow these steps to set up the environment and run the experiments. ### 1️⃣ Create and activate a virtual environment Using **Python 3.10+**: ```bash # Create a virtual environment python -m venv env # Activate it # On Linux/MacOS: source env/bin/activate # On Windows (PowerShell): env\Scripts\activate ``` --- ### 2️⃣ Clone the repository ```bash git clone github.com cd Drought-Forecasting-Benin/models ``` --- ### 3️⃣ Install dependencies ```bash pip install -r requirements.txt ``` --- ## 🚀 Running Experiments ### ▶ Run for a specific city (example: Banikoara) ```bash python banikoara_generic_model_building.py ``` --- ### ▶ Run for all six cities at once ```bash chmod +x run_all.sh ./run_all.sh ``` --- ## 📊 Results Overview Our comparative study involved: * **6 Machine Learning models**: Linear Regression, Ridge Regression, Random Forest, XGBoost, LightGBM, SVR * **4 Deep Learning models**: Conv1D, LSTM, GRU, Conv1D-LSTM * **Evaluation metrics**: R², RMSE, MAE, Carbon Footprint * **Uncertainty metrics**: Empirical Coverage, Prediction Interval Width The **Conv1D-LSTM** model emerged as the top performer, offering an optimal balance between predictive accuracy and uncertainty coverage. --- ## 📚 Citation If you use this …