McEachran et al. (2025) Knowledge‐Guided Machine Learning for Operational Flood Forecasting

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-11-01
• Authors: Zachary P. McEachran, Rahul Ghosh, Arvind Renganathan, Somya Sharma, Lindsay Kelly, Michael Steinbach, John L. Nieber, Christopher Duffy, Vipin Kumar
• DOI: 10.1029/2024wr039064

Research Groups

Not specified in the abstract.

Short Summary

This study introduces a Factorized Hierarchical Neural Network (FHNN), a knowledge-guided machine learning framework for operational hydrologic forecasting at the catchment scale. The FHNN demonstrates superior streamflow prediction performance compared to expert human forecasters after the initial 12–18 hours, laying groundwork for AI-human collaboration in river forecasting.

Objective

• To develop and evaluate a knowledge-guided machine learning framework, the Factorized Hierarchical Neural Network (FHNN), for operational hydrologic forecasting at the catchment scale, specifically designed to model multi-scale processes and efficiently integrate real-time data.
• To compare the FHNN's flood forecast ability against a leading deep learning alternative (autoregressive LSTM) and expert human forecasters using physics-based models from the US National Weather Service (NWS).

Study Configuration

• Spatial Scale: Catchment scale, evaluated on the large-sample CAMELS-US data set.
• Temporal Scale: Operational hydrologic forecasting, with forecast horizons extending beyond 12–18 hours.

Methodology and Data

• Models used:
  • Factorized Hierarchical Neural Network (FHNN): Comprising inverse and forward models, designed for multi-scale process modeling and inference-based data integration.
  • Autoregressive Long Short-Term Memory (LSTM): A leading deep learning alternative used for comparison.
  • Physics-based model: Used by US NWS expert human forecasters.
• Data sources:
  • Observed precipitation data.
  • Observed temperature data.
  • Observed streamflow data.
  • CAMELS-US data set (large-sample).
  • Operational flood forecast data from the US National Weather Service (NWS).

Main Results

• The FHNN framework efficiently integrates real-time data through an inference-based inverse modeling approach, improving forecasts more efficiently than computationally intensive data assimilation methods.
• Expert human forecasters, utilizing physics-based models, produce more accurate forecasts within the first 12–18 hours of a forecast's issuance.
• The FHNN demonstrates significantly better streamflow predictions than NWS expert-derived forecasts after the initial 12–18 hours.

Contributions

• Introduction of the Factorized Hierarchical Neural Network (FHNN), a novel knowledge-guided machine learning framework for operational hydrologic forecasting that effectively models multi-scale processes and their interactions.
• Development of an efficient inference-based data integration approach within FHNN, offering an alternative to computationally intensive data assimilation methods.
• Quantitative demonstration of FHNN's superior long-term forecast performance (beyond 12–18 hours) compared to expert human forecasters from the US NWS, highlighting the potential for leveraging AI-based models with human expertise to enhance river forecasts.

Funding

Not specified in the abstract.

Citation

@article{McEachran2025KnowledgeGuided,
  author = {McEachran, Zachary P. and Ghosh, Rahul and Renganathan, Arvind and Sharma, Somya and Kelly, Lindsay and Steinbach, Michael and Nieber, John L. and Duffy, Christopher and Kumar, Vipin},
  title = {Knowledge‐Guided Machine Learning for Operational Flood Forecasting},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2024wr039064},
  url = {https://doi.org/10.1029/2024wr039064}
}