Aderdour et al. (2026) Advanced Drought Prediction Using Hybrid Deep Learning Models: A Case Study of the High Atlas and Anti-Atlas Mountains

Identification

• Journal: Open Access CRIS of the University of Bern
• Year: 2026
• Authors: Nacer Aderdour, Ikram Essajai, Mohamed El Ghazouani, Abdelmajid Bessate, Henri Rueff, Mehdi Maanan, Hassan Rhinane
• DOI: 10.48620/94044

Research Groups

• Centre for Development and Environment (CDE), University of Bern, Switzerland.
• Hassan II University of Casablanca, Morocco.
• Copernicus Publications (Publisher).

Short Summary

This study develops a hybrid deep learning framework using Gated Recurrent Units (GRU) to predict the Standardized Precipitation Index (SPI) at a 5 km resolution in the High and Anti-Atlas mountains. The model achieves over 91% accuracy by integrating multi-source remote sensing and climate data, providing a robust early-warning tool for regions with sparse meteorological stations.

Objective

• To develop an operational high-resolution drought prediction framework for semi-arid, topographically complex regions by benchmarking various recurrent neural network (RNN) architectures against traditional models.

Study Configuration

• Spatial Scale: High Atlas and Anti-Atlas Mountains, Morocco (5 km spatial resolution).
• Temporal Scale: 1990–2024 (Historical data and training); 2021–2024 (Validation/held-out period).

Methodology and Data

• Models used: Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Random Forest (RF), and AutoRegressive Integrated Moving Average (ARIMA).
• Data sources: Multi-source synthesis of remote sensing and climate variables, including the Standardized Precipitation Index (SPI), Normalized Difference Vegetation Index (NDVI), soil moisture, precipitation, and temperature.
• Feature Engineering: 128 engineered features, including rolling statistics, seasonality indicators, lag dependencies, and cross-variable interactions.

Main Results

• The GRU model was identified as the most effective architecture, achieving 91.89% predictive accuracy within a ±0.2 SPI threshold.
• Deep sequence learning models (GRU, LSTM, Bi-LSTM) significantly outperformed traditional baseline models such as Random Forest and ARIMA.
• The integration of 128 engineered features allowed the model to effectively capture complex drought dynamics in rugged terrain despite sparse ground-based observations.

Contributions

• Establishes a high-resolution (5 km) drought forecasting framework specifically optimized for semi-arid mountainous environments where conventional monitoring is limited.
• Demonstrates the superior performance of GRU-based hybrid models in processing multi-variable climate sequences for month-ahead early warnings.
• Provides a scalable methodology for drought risk management in North Africa using advanced feature engineering and remote sensing synthesis.

Funding

• LH MENA.

Citation

@article{Aderdour2026Advanced,
  author = {Aderdour, Nacer and Essajai, Ikram and Ghazouani, Mohamed El and Bessate, Abdelmajid and Rueff, Henri and Maanan, Mehdi and Rhinane, Hassan},
  title = {Advanced Drought Prediction Using Hybrid Deep Learning Models: A Case Study of the High Atlas and Anti-Atlas Mountains},
  journal = {Open Access CRIS of the University of Bern},
  year = {2026},
  doi = {10.48620/94044},
  url = {https://doi.org/10.48620/94044}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)