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

Evapotranspiration (ET) is one of the most important components of global hydrologic cycle and has significant impacts on energy exchange and climate change. Numerous models have been developed to estimate ET so far; however, great uncertainties in models still require considerations. The aim of this study is to reduce model errors and uncertainties among multi-models to improve daily ET estimate. The Bayesian model averaging (BMA) method is used to assemble eight ET models to produce ET with Landsat 8 satellite data, including four surface energy balance models (i.e., SEBS, SEBAL, SEBI, and SSEB) and four machine learning algorithms (i.e., polymars, random forest, ridge regression, and support vector machine). Performances of each model and BMA method were validated through in situ measurements of semi-arid region. Results indicated that the BMA method outperformed all eight single models. The four most important models obtained by the BMA method were ranked by random forest, SVM, SEBS, and SEBAL. The BMA method coupled with machine learning can significantly improve the accuracy of daily ET estimate, reducing uncertainties among models, and taking different intrinsic benefits of empirically and physically based models to obtain a more reliable ET estimate.

Automated summary

This study utilizes the Bayesian model averaging method to combine multiple ET models, incorporating machine learning algorithms and remote sensing data to improve the accuracy of daily ET estimates. Results show that the BMA method outperforms individual models, further reducing model uncertainties.