Physically-constrained evapotranspiration models with machine learning parameterization outperform pure machine...

Area of Science:

• Environmental Science
• Machine Learning
• Hydrology

Background:

• Terrestrial evapotranspiration (ET) is vital for water, energy, and carbon cycles.
• Pure machine learning (ML) models struggle with ET estimation, particularly under extreme conditions.
• Physics-informed ML, specifically hybrid models, show promise for improved ET accuracy.

Purpose of the Study:

• To investigate the mechanisms behind the improved performance of hybrid ET models.
• To compare six novel hybrid ET models with a pure ML model.
• To identify optimal parameterizations for hybrid ET modeling.

Main Methods:

• Developed six hybrid models integrating different physical ET formulations with the random forest algorithm.
• Trained and compared models using daily ET observations, meteorological data, and satellite remote sensing.
• Analyzed the correlation between model error (RMSE) and sensitivity to machine-learned parameters.

Main Results:

• A strong correlation (r=0.93) was found between ET estimate sensitivity and model error (RMSE).
• Reduced sensitivity to machine-learned parameters minimizes error propagation and enhances model performance.
• The most accurate hybrid model used a novel, stable empirical parameter, outperforming pure ML and other hybrid models.

Conclusions:

• Conventional parameterizations may need reevaluation for optimal integration of physical models with ML.
• Domain knowledge is crucial for developing effective hybrid models.
• This study provides insights for advancing hybrid modeling beyond ET estimation.