Bakar et al. (2026) Deep learning approaches for streamflow flash drought prediction across the contiguous United States

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-01
• Authors: Sophia Bakar, Hoang Hai Nguyen, Hyunglok Kim, Venkataraman Laskhmi
• DOI: 10.1016/j.jhydrol.2026.135956

Research Groups

• Center for Water, Agriculture, Natural Resources, and Energy Solutions, RTI International (Durham, NC, USA)
• Department of Environment and Energy Engineering, Gwangju Institute of Science and Technology (Gwangju, South Korea)
• Department of Civil and Environmental Engineering, University of Virginia (Charlottesville, VA, USA)

Short Summary

This study evaluates the ability of deep learning models to predict streamflow flash droughts (SFDs) across 671 catchments in the contiguous United States. The Temporal Fusion Transformer (TFT) model achieved the highest overall performance, outperforming the baseline Long Short-Term Memory (LSTM) and an LSTM with static catchment attributes.

Objective

• Evaluate the ability of deep learning models to predict streamflow flash droughts (SFDs) across a large-sample, multi-catchment dataset.
• Assess how differences in model architecture influence performance in capturing key characteristics of SFD events.

Study Configuration

• Spatial Scale: Continental United States (CONUS)
• Temporal Scale: Daily to pentad (5-day) scale

Methodology and Data

• Models used:
  • Baseline Long Short-Term Memory (LSTM) model
  • LSTM model with static catchment attributes
  • Temporal Fusion Transformer (TFT) model
• Data sources:
  • CAMELS dataset (catchment attributes and meteorology)
  • Daymet dataset (daily surface weather and climatological summaries)
  • USGS National Water Information System (NWIS) for streamflow observations

Main Results

• The TFT model achieved the highest overall performance, with a median Kling-Epstein-Gupta efficiency (KGE) of 0.87, root mean squared error (RMSE) of 0.13, and mean absolute error (MAE) of 0.10.
• The TFT model showed the highest median detection rate and lowest miss rate across all models for detecting SFDs.

Contributions

• This study contributes to the understanding of how deep learning models can improve prediction of rapid streamflow changes associated with flash drought development.
• The results demonstrate that Transformer-based models, such as the TFT, can capture complex relationships between atmospheric forcing and streamflow response.

Funding

• National Science Foundation (NSF)
• National Aeronautics and Space Administration (NASA)
• United States Geological Survey (USGS)

Citation

@article{Bakar2026Deep,
  author = {Bakar, Sophia and Nguyen, Hoang Hai and Kim, Hyunglok and Laskhmi, Venkataraman},
  title = {Deep learning approaches for streamflow flash drought prediction across the contiguous United States},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135956},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135956}
}