On the use of machine learning based ensemble approaches to improve evapotranspiration estimates from croplands across a wide environmental gradient

Abstract

Accurately mapping of regional-scale evapotranspiration (ET) from the croplands using remote sensing is currently challenged by limited spatial information on crop and field management to properly characterize the biophysical constraints on ET. A multi-model ensemble can potentially address this challenge, however, conventional ensemble models using the simple average (MEAN) or Bayesian Model Average (BMA) assign a fixed weight to each model and may not fully utilize the strengths of individual models. To this end, we developed four ensemble ET Models (EEMs) that use different machine learning (ML) classifiers, namely K-nearest neighbors, random forest, support vector machine, and multi-layer perception neural network (MLP), to assign varying weights to assemble six physically-driven remote sensing-based ET models. These ML-based EEMs were compared against the six individual ET models and two conventional ensemble methods (MEAN and BMA) using latent heat fluxes (lambda E) observations from 47 cropland eddy covariance flux sites covering diverse environments across the globe. Results suggested that while MEAN and BMA can reduce some uncertainties in the individual models, ML-based EEMs can better integrate the capabilities of multiple biophysical constraints on ET used across the individual models. The four ML-based EEMs yielded daily lambda E for training, validation, and testing datasets with the coefficient of determination (R-2) and root mean squared error (RMSE) within 0.75 - 0.83 and 18 - 21 W m(-2), respectively, among which the MLP algorithm was found to be the most efficient with respect to accuracies and costs. These performance metrics were much better than those from the conventional ensemble models (R-2 = 0.69 - 0.71, RMSE = 23 - 25 W m(-2)) and six individual ET models (R-2 = 0.53 - 0.69, RMSE = 26 - 35 W m(-2)). Results suggested that ML-based EEMs perform much better than the conventional approaches and hence can be viable tools for mapping cropland ET across a wide environmental gradient.