Evaluating Different Machine Learning Methods for Upscaling Evapotranspiration from Flux Towers to the Regional Scale

TLDR

Evapotranspiration links surface energy, water, and carbon cycles, but in situ measurements are spatially limited and regional‑scale modeling remains challenging. The study aims to upscale ET from flux‑tower sites to the regional scale using machine‑learning algorithms. Five machine‑learning algorithms—artificial neural network, Cubist, deep belief network, random forest, and support vector machine—were trained on 65 site‑years from 36 flux towers and applied to daily 1 km × 1 km grid cells across the Heihe River Basin (2012–2016). The ANN, Cubist, RF, and SVM produced comparable ET estimates with slightly lower error than the deep belief network, RF had the lowest regional uncertainty, and all methods captured the spatial and temporal patterns of ET, performing best over densely vegetated areas.

Abstract

Abstract Evapotranspiration (ET) is a vital variable for land‐atmosphere interactions that links surface energy balance, water, and carbon cycles. The in situ techniques can measure ET accurately but the observations have limited spatial and temporal coverage. Modeling approaches have been used to estimate ET at broad spatial and temporal scales, while accurately simulating ET at regional scales remains a major challenge. In this study, we upscale ET from eddy covariance flux tower sites to the regional scale with machine learning algorithms. Five machine learning algorithms are employed for ET upscaling including artificial neural network, Cubist, deep belief network, random forest, and support vector machine. The machine learning methods are trained and tested at 36 flux towers sites (65 site years) across the Heihe River Basin and are then applied to estimate ET for each grid cell (1 km × 1 km) within the watershed and for each day over the period 2012–2016. The artificial neural network, Cubist, random forest, and support vector machine algorithms have almost identical performance in estimating ET and have slightly lower root‐mean‐square error than deep belief network at the site scale. The random forest algorithm has slightly lower relative uncertainty at the regional scale than other methods based on three‐cornered hat method. Additionally, the machine learning methods perform better over densely vegetated conditions than barren land or sparsely vegetated conditions. The regional ET generated from the machine learning approaches captured the spatial and temporal patterns of ET at the regional scale.