recherchemse/SmartSDG-Tunisia

# Satellite-derived Composite Drought Index Modelling Framework and Forecasting using Deep Learning ## Project Overview This project develops a novel CDI using multi-source Earth Observation data (CHIRPS, ERA5-Land, MODIS) at 1 km resolution. It integrates SPI, SPEI, soil moisture anomaly (SMA), NDVI anomaly, and LST anomaly via a logic-based cause-effect framework for drought monitoring and classifies into Normal, Watch, Warning, Alert-1/2, and Urgency stages. Deep learning forecasts CDI using 12 models (e.g., TimeFormer, SSSLN, LSTM), with TimeFormer performing best (Accuracy: 0.9057). ## Key Features - High-resolution (1 km) national drought dataset for Tunisia 2000-2025. - Captures major events (e.g., 2000-2002, 2016-2018, 2021-2025 droughts). - Validated against ground stations (R²=0.75 for SPI). - Designed for reproducibility, extensibility, and transferability to other regions. ## Data Sources | Product | Variables | Resolution | Period | Source | |---------|-----------|------------|--------|--------| | CHIRPS | Rainfall (SPI) | 5 km | 1981-present | UCSB Climate Hazards Center | | ERA5-Land | Temp, Evap, Soil Moisture (SPEI, SMA) | 9 km | 1981-present | Copernicus CDS | | MODIS | LST (MOD11A1), NDVI (MOD13A3) | 1 km | 2000-present | NASA LP DAAC | Data harmonized to 1 km monthly scale via clipping, re-projection, and gap-filling. ## Methods 1. Compute indicators: SPI-1 (gamma dist.), SPEI-1 (log-logistic), z-scores for SMA/NDVI-A/LST-A. 2. CDI logic: Thresholds (e.g., SPI/SPEI/SMA/NDVI ≤ -1; LST-A ≥ 0.5). 3. Forecasting: Train/test 12 DL models on CDI time-series (train: 2000-2016; test: 2024-2025). 4. Validation: Wavelet analysis, ground SPI correlation. ## Installation - Python environment with PyTorch/TensorFlow, Google Earth Engine. - Public datasets downloadable via APIs (CHIRPS, ERA5, MODIS). ## Usage Run scripts for data processing, CDI computation, and forecasting. ## Funding This work was supported by the Ministry of Higher Education a …