Evapotranspiration estimation using four different machine learning approaches in different terrestrial ecosystems

Abstract

Elucidating the biophysical mechanisms governing the exchange of water vapor between land and the atmosphere is particularly crucial for addressing water scarcity under climate change. Owing to the rapid development of machine learning techniques, a series of powerful tools have been proposed over the past two decades, allowing the scientific community to obtain new insights into the patterns of evapotranspiration (ET) on different spatial scales ranging from ecosystem to global. The primary focus of this study was to investigate the feasibility and effectiveness of both extreme learning machine (ELM) and adaptive neuro-fuzzy inference system (ANFIS) to estimate the daily ET with flux tower observations in four main types of ecosystems. A comparative research was undertaken to evaluate the potential of the models compared with the conventional artificial neural network and support vector machine models. All the developed models were evaluated according to the following performance indices: coefficient of determination (R2), Nash-Sutcliffe efficiency (NSE), root mean square error and mean absolute error. The results showed that all the applied models had high performance for modeling daily ET (e.g., R2 = 0.9398–0.9593 and NSE = 0.8877–0.9147 in forest ecosystem). Among the applied ELM models, the three hybrid ELM methods outperformed the original ELM method in most cases at the four sites and the computational time required for learning these ELM models has been considerably reduced. The subtractive clustering and fuzzy c-means clustering algorithms for ANFIS generally performed better than the grid partitioning algorithm. It was concluded that the advanced ELM and ANFIS models can be recommended as important complements to traditional methods due to their robustness and flexibility. Moreover, significant difference regarding the modeling performance existed among the four major ecosystems types. The models generally achieved the best performance in forest ecosystem, while provided the worst in cropland ecosystem.