Dargham et al. (2026) Development of intensity-duration-frequency curves using machine learning and satellite-derived precipitation data

Identification

• Journal: Frontiers in Water
• Year: 2026
• Date: 2026-01-29
• Authors: Elias Dargham, Cynthia Andraos
• DOI: 10.3389/frwa.2026.1727182

Research Groups

• Regional Center for Water and Environment, Faculty of Engineering, Saint Joseph University of Beirut, Beirut, Lebanon

Short Summary

This study develops more accurate and robust Intensity-Duration-Frequency (IDF) curves by leveraging satellite-based precipitation data and advanced machine learning techniques. It finds that deep learning models, particularly Temporal Convolutional Attention Networks (TCAN), significantly outperform traditional statistical methods in capturing complex rainfall patterns and reducing uncertainty.

Objective

• To develop more accurate and robust Intensity-Duration-Frequency (IDF) curves using satellite-based precipitation datasets and advanced machine learning techniques, thereby lowering uncertainty and improving the reliability of construction under non-stationary rainfall trends.

Study Configuration

• Spatial Scale: Beirut, Lebanon (approximate center coordinate 33.89°N, 35.50°E; metropolitan area ~67 km²; Beirut International Airport 33.83°N, 35.50°E).
• Temporal Scale: Daily (24-hour) and sub-daily (30-minute) precipitation data from January 1998 to February 2025.

Methodology and Data

• Models used: Gumbel distribution (statistical benchmark), Support Vector Regression (SVR), Artificial Neural Networks (ANN), Temporal Convolutional Networks (TCN), Temporal Convolutional Attention Networks (TCAN).
• Data sources: Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG) Version 07 Final Run (daily and 30-minute precipitation data). Ground-based station measurements from Beirut International Airport and government sources for validation.

Main Results

• Deep learning models (TCAN, TCN, ANN) significantly outperformed the Gumbel statistical method in constructing IDF curves.
• TCAN achieved the highest performance with R² = 0.961, Nash-Sutcliffe Efficiency (NSE) = 0.959, Root Mean Squared Error (RMSE) = 6.609 mm/h, and Mean Absolute Error (MAE) = 3.841 mm/h.
• TCN followed closely with R² = 0.958, NSE = 0.956, RMSE = 6.870 mm/h, and MAE = 4.069 mm/h.
• The Gumbel statistical method showed lower performance with R² = 0.623 and NSE = 0.596.
• ML-based curves improved prediction errors for extreme events by at least 61.24% (up to 80.0%) compared to the statistical method.
• Advanced ML models significantly reduced uncertainty: TCAN by 62.2%, TCN by 60.8%, ANN by 50.3%, and SVM by 41.0% relative to the Gumbel distribution.
• TCAN achieved excellent calibration, exceeding the 95% coverage target for prediction intervals at 95.6%.

Contributions

• First study to integrate satellite-based precipitation data with advanced deep learning models, specifically Temporal Convolutional Networks (TCNs) and attention mechanisms (TCANs), to directly construct Intensity-Duration-Frequency (IDF) curves.
• Demonstrates the superior performance of advanced deep learning models over traditional statistical methods (Gumbel distribution) in terms of accuracy, explained variance, and uncertainty reduction for IDF curve development.
• Provides a robust, adaptive, and data-driven framework for IDF curve construction, particularly beneficial for data-scarce regions and in the context of non-stationary climate conditions.
• Quantifies the reduction in epistemic and total uncertainty, and the improvement in prediction interval precision and calibration offered by advanced machine learning models for hydrological design.

Funding

The authors declared that financial support was not received for this work and/or its publication.

Citation

@article{Dargham2026Development,
  author = {Dargham, Elias and Andraos, Cynthia},
  title = {Development of intensity-duration-frequency curves using machine learning and satellite-derived precipitation data},
  journal = {Frontiers in Water},
  year = {2026},
  doi = {10.3389/frwa.2026.1727182},
  url = {https://doi.org/10.3389/frwa.2026.1727182}
}