Bayati et al. (2025) Evaluating the Functional Realism of Deep Learning Rainfall‐Runoff Models Using Catchment Hydrology Principles

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

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-12-30
• Authors: Ara Bayati, Ali Ameli, Saman Razavi
• DOI: 10.1029/2025wr040076

Research Groups

Not explicitly detailed in the abstract.

Short Summary

This study introduces a hydrology-specific Explainable AI (XAI) framework to evaluate the functional realism of Long-Short-Term-Memory (LSTM) networks in rainfall-runoff modeling. It reveals that despite high predictive accuracy, LSTMs often exhibit hydrologically implausible internal reasoning, particularly under varying climatic conditions.

Objective

• To investigate whether the internal functional reasoning of deep learning models like LSTMs, which achieve high predictive accuracy in rainfall-runoff modeling, is physically reliable.

Study Configuration

• Spatial Scale: 672 North American catchments.
• Temporal Scale: Event-scale to long-term climatic shifts, with analysis of lag-dependent and time-varying functional relationships.

Methodology and Data

• Models used: Long-Short-Term-Memory (LSTM) networks. A new hydrology-specific Explainable AI (XAI) framework is introduced, which extracts nonlinear, lag-dependent, and time-varying Impulse Response Functions (IRFs) to quantify functional relationships within LSTMs.
• Data sources: Not explicitly detailed in the abstract, but inputs include precipitation (P), temperature (T), and potential evapotranspiration (PET), and the output is simulated streamflow.

Main Results

• High predictive accuracy in LSTMs often masks hydrologically implausible internal reasoning.
• In over 70% of rain-dominated basins, short-term temperature rises unexpectedly raise simulated streamflow and enhance the celerity rate even without rainfall.
• In snow-dominated regions, potential evapotranspiration (PET) is misattributed as a driver of snowmelt-related flow and enhances the catchment's celerity rate.
• Correlation-driven learning in LSTMs can compromise the robustness of forecasts under weather extremes and short-term and long-term climatic shifts.

Contributions

• Introduces a novel hydrology-specific Explainable AI (XAI) framework for opening the black-box of LSTM models.
• Develops a method to extract nonlinear, lag-dependent, and time-varying Impulse Response Functions (IRFs) to quantify LSTM's internal functional relationships.
• Provides a scalable diagnostic tool for assessing the functional realism of deep learning models across diverse catchment types.
• Bridges deep learning with hydrologic understanding by offering a catchment hydrology perspective for evaluating model realism.
• Demonstrates that high accuracy in LSTMs does not guarantee physically reliable internal functioning, highlighting vulnerabilities for robustness under climate shifts.

Funding

Not explicitly detailed in the abstract.

Citation

@article{Bayati2025Evaluating,
  author = {Bayati, Ara and Ameli, Ali and Razavi, Saman},
  title = {Evaluating the Functional Realism of Deep Learning Rainfall‐Runoff Models Using Catchment Hydrology Principles},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2025wr040076},
  url = {https://doi.org/10.1029/2025wr040076}
}