Evaluation of water flux predictive models developed using eddy covariance observations and machine learning: a meta-analysis

Abstract. With the rapid accumulation of water flux observations from global eddy-covariance flux sites, many studies have used data-driven approaches to model site-scale water fluxes with various predictors and machine learning algorithms used. However, systematic evaluation of such models is still limited. We therefore performed a meta-analysis of 32 such studies, derived 139 model records, and evaluated the impact of various features on model accuracy throughout the modeling flow. SVM (average R-squared = 0.82) and RF (average R-squared = 0.81) outperformed over evaluated algorithms in both cross-study and intra-study (with the same training dataset) comparisons. The average accuracy of the model applied to arid regions is higher than other climate classes. The average accuracy of the model was slightly lower for forest sites (average R-squared = 0.76) than for cropland and grassland sites (average R-squared = 0.8 and 0.79), but higher than for shrub sites (average R-squared = 0.67). Among various predictor variables, the use of net/sun radiation, precipitation, air temperature, and the fraction of absorbed photosynthetically active radiation improved the model accuracy. Among the different validation methods, random cross-validation shows higher model accuracy than spatial cross-validation and temporal cross-validation, but spatial cross-validation is more important for the application for water flux predictive models when used for spatial extrapolation. The findings of this study are promising to guide future research on such machine learning-based modeling.