Deep learning for cross-region streamflow and flood forecasting at a global scale.

Area of Science:

• Hydrology
• Artificial Intelligence
• Environmental Modeling

Background:

• Streamflow and flood forecasting present persistent challenges in hydrology.
• Traditional physically based models struggle with parameter scarcity and complex calibration, especially in ungauged catchments.

Purpose of the Study:

• To introduce a novel hybrid deep learning model, ED-DLSTM, for global streamflow forecasting.
• To address limitations of traditional models in gauged and ungauged catchments.

Main Methods:

• Development of an encoder-decoder double-layer long short-term memory (ED-DLSTM) model.
• Application of the model to over 2,000 catchments across North America, Europe, and the UK.
• Inclusion of a spatial attribute encoding module for interpretability and regionalization.

Main Results:

• Achieved a mean Nash-Sutcliffe efficiency (NSE) of 0.75 across diverse catchments.
• Demonstrated successful application in 160 ungauged catchments in Chile, with 76.9% achieving NSE > 0.
• Validated the model's cross-region predictive capabilities and hydrological regionalization effects.

Conclusions:

• Deep learning methods, like ED-DLSTM, offer a powerful solution to overcome hydrologic data scarcity.
• The proposed model shows significant potential to improve streamflow forecasting accuracy and reliability.
• ED-DLSTM advances the application of AI in hydrology, addressing physical model deficiencies.