Evaluation of ten machine learning methods for estimating terrestrial evapotranspiration from remote sensing

Remote sensing retrieval of evapotranspiration (ET), or surface latent heat exchange (LE), is of great utility for many applications. Machine learning (ML) methods have been extensively used in many disciplines, but so far little work has been performed systematically comparing ML methods for ET retrieval. This paper provides an evaluation of ten ML methods for estimating daily ET based on daily Global LAnd Surface Satellite (GLASS) radiation data and high-level Moderate-Resolution Imaging Spectroradiometer (MOD'S) data products and ground measured ET data from 184 flux tower sites. Measurements of accuracy (RMSE, R-2, and bias) and run time were made for each of ten ML methods with a smaller training data set (n = 7910 data points) and a larger training data set (n = 69,752 data points). Inclusion of more input variables improved algorithm performance but had little effect on run time. The best results were obtained with the larger training data set using the bootstrap aggregation (bagging) regression tree (validation RMSE = 19.91 W/m(2)) and three hidden layer neural network (validation RMSE = 20.94 W/m(2)), although the less computationally demanding random kernel (RKS) algorithm also produced good results (validation RMSE = 22.22 W/m(2)). Comparison of results from sites with different ecosystem types showed the best results for evergreen, shrub, and grassland sites, and the weakest results for wetland sites. Generally, performance was not improved by training with data from only the same ecosystem type.