Estimating evapotranspiration by coupling Bayesian model averaging methods with machine learning algorithms.

Area of Science:

• Hydrology
• Remote Sensing
• Climate Science

Background:

• Evapotranspiration (ET) is crucial for the water cycle, energy balance, and climate change.
• Existing ET models have significant uncertainties, necessitating improved estimation methods.
• Accurate ET estimation is vital for water resource management and climate modeling.

Purpose of the Study:

• To reduce model errors and uncertainties in daily evapotranspiration (ET) estimation.
• To improve the accuracy of daily ET estimates by assembling multiple ET models.
• To evaluate the effectiveness of the Bayesian model averaging (BMA) method for ET estimation.

Main Methods:

• Employed the Bayesian model averaging (BMA) method to combine eight different ET models.
• Utilized Landsat 8 satellite data for ET estimation.
• Included four surface energy balance models (SEBS, SEBAL, SEBI, SSEB) and four machine learning algorithms (polymars, random forest, ridge regression, support vector machine).
• Validated model performances using in situ measurements in a semi-arid region.

Main Results:

• The BMA method significantly outperformed all eight individual ET models.
• The most influential models identified by BMA were random forest, support vector machine (SVM), SEBS, and SEBAL.
• The BMA approach, integrating machine learning, substantially enhanced the accuracy of daily ET estimates.
• Reduced model uncertainties by leveraging the strengths of both empirical and physically-based models.

Conclusions:

• The Bayesian model averaging (BMA) method offers a robust approach to improve daily ET estimation accuracy.
• Coupling BMA with machine learning techniques effectively reduces uncertainties in multi-model ET predictions.
• This integrated approach provides a more reliable ET estimate by capitalizing on diverse model strengths.