Sarwar et al. (2026) Artificial intelligence for multiscale drought modeling and decision making

Identification

• Journal: Elsevier eBooks
• Year: 2026
• Date: 2026-01-01
• Authors: Abid Sarwar, Rui Gao, Mohammad Safeeq, J. T. Abatzoglou, Josué Medellín-Azuara, Joshua H. Viers
• DOI: 10.1016/b978-0-443-44625-2.00011-4

Research Groups

• Civil and Environmental Engineering Department, School of Engineering, University of California, Merced, CA, United States
• Management of Complex Systems, School of Engineering, University of California, Merced, CA, United States

Short Summary

This chapter assesses artificial intelligence and machine learning (ML) methodologies for enhancing drought evaluation and modeling across various scales, demonstrating significant improvements over traditional methods in forecasting accuracy, spatiotemporal pattern identification, and water usage efficiency.

Objective

• To assess artificial intelligence and machine learning (ML) methodologies that integrate sensory, satellite, and climate data to enhance drought evaluation and modeling, spanning from individual farms to national policy frameworks.

Study Configuration

• Spatial Scale: Individual farms, national policy frameworks, global onset events, field-scale.
• Temporal Scale: Seasonal droughts, 12 weeks (for DroughtCast predictions), 7 days (for flash drought detection lead time).

Methodology and Data

• Models used: Ensemble ML methods (Random Forest, Support Vector Machines), Deep Learning architectures (Recurrent Neural Networks, Convolutional Neural Networks, CNN-LSTM hybrid models, ConvLSTM networks), DroughtCast framework, Physics-informed neural networks.
• Data sources: Sensory data, satellite data, climate data.

Main Results

• Ensemble ML methods (Random Forest and Support Vector Machines) achieve an accuracy of 80% in forecasting seasonal droughts.
• Deep learning architectures, particularly CNN-LSTM hybrid models, demonstrate a 20% improvement in performance compared to individual algorithms.
• ConvLSTM networks are proficient in capturing spatiotemporal patterns but face computational scalability challenges in large-scale applications.
• The DroughtCast framework effectively predicts US Drought Monitor categories 12 weeks in advance using recurrent neural networks.
• ML-enhanced flash drought detection reaches an accuracy of 33% for global onset events with a 7-day lead time.
• IoT-enabled precision agriculture systems decrease water usage by 20%–40% while maintaining yield levels in drought conditions.
• Physics-informed neural networks enhance predictive accuracy by integrating physical process knowledge, especially with limited data.

Contributions

• Demonstrates notable improvements of artificial intelligence and machine learning methodologies over traditional methods for drought evaluation and modeling across multiple scales.
• Highlights the effectiveness of various advanced ML techniques, including ensemble methods, deep learning hybrids (CNN-LSTM), and physics-informed neural networks, for enhanced drought forecasting and management.
• Introduces specific frameworks and applications like DroughtCast for long-lead drought prediction and IoT for water conservation in agriculture.
• Identifies critical limitations of current ML approaches in drought modeling, such as inadequate model transferability, insufficient spatial resolution, and the need for data-informed design over architectural optimization.

Funding

Not specified in the provided text.

Citation

@article{Sarwar2026Artificial,
  author = {Sarwar, Abid and Gao, Rui and Safeeq, Mohammad and Abatzoglou, J. T. and Medellín-Azuara, Josué and Viers, Joshua H.},
  title = {Artificial intelligence for multiscale drought modeling and decision making},
  journal = {Elsevier eBooks},
  year = {2026},
  doi = {10.1016/b978-0-443-44625-2.00011-4},
  url = {https://doi.org/10.1016/b978-0-443-44625-2.00011-4}
}