ARGOUBI25/drought-prediction-tunisia

# Spatiotemporal SHAP Analysis for Agricultural Drought Prediction ## A Multi-Source Machine Learning Framework in Semi-Arid Tunisia --- ## 📌 Overview This repository contains the full reproducible codebase for the research paper: > **Spatiotemporal SHAP Analysis for Agricultural Drought Prediction: A Multi-Source Machine Learning Framework in Semi-Arid Tunisia** > *Frontiers in Artificial Intelligence — Section: AI in Food, Agriculture and Water* > > **Authors:** Argoubi Majdi, Khaled Mili > **Paper DOI:** [Pending Publication] > **Dataset DOI:** 10.5281/zenodo.18733246 ### Study Area Five semi-arid governorates of central Tunisia: **Kairouan, Kasserine, Sidi Bou Zid, Gafsa, and Siliana**. ### Objectives 1. Predict agricultural drought (SSMI) using multi-source environmental and socioeconomic data. 2. Evaluate 6 ML/DL models (XGBoost, LightGBM, CatBoost, RF, BPNN, LSTM) with a focus on extreme event generalization (2021–2022). 3. Use **Tree SHAP** to explain global drivers and spatiotemporal drought dynamics. --- ## Key Results | Model | Test R² | Test SDI | Val R² (2021–22) | Val SDI | |-------|---------|----------|------------------|---------| | **BPNN** | **0.860** | **0.626** | 0.655 | 0.421 | | XGBoost | 0.817 | 0.572 | **0.696** | **0.448** | | LightGBM | 0.807 | 0.561 | — | — | | LSTM | 0.790 | 0.541 | 0.667 | — | | CatBoost | 0.756 | — | — | — | | RF | 0.681 | — | — | — | **Top SHAP predictors (XGBoost):** - MERRA-2 Surface SM: 26.0% - Temperature (T2M): 14.2% - Sand content: 10.0% - Precipitation: 8.6% --- ## Repository Structure ``` drought-prediction-tunisia/ ├── data/ # Raw and processed data (see data/README.md) ├── notebooks/ # Jupyter notebooks (exploration → figures) ├── src/ # Python source modules │ ├── data/ # Download and preprocessing scripts │ ├── models/ # Training, evaluation, metrics │ ├── shap/ # SHAP global, dependence, spatiotemporal │ └── figures …