Modeling the Ecosystem Water Use Efficiency of Croplands Using Machine Learning and MODIS Data

Abstract

The ecosystem water use efficiency (WUE), which is the ratio of carbon gain (gross primary productivity, GPP) to water consumption (evapotranspiration, ET), is widely regarded as a crucial link in coupling the carbon and water cycles of terrestrial ecosystems on a global scale. The estimation of cropland WUE is considerably important for improving agricultural management, simulating the carbon–water cycle processes of croplands, and enhancing crop yields. Currently, the estimate of WUE primary relies on the modeled GPP and ET data. However, differences in model structures and parameter uncertainties cause remarkable differences in the accuracy of the estimated WUE among different models. The eddy covariance technology enables the continuous observation and research of carbon and water fluxes at a site scale, ensuring simulation accuracy. However, the technology is constrained by its limited spatial coverage. In this study, we integrated data from 20 global cropland sites and built 7 machine learning models based on remote sensing to directly estimate cropland WUE, thereby expanding the spatial scope of WUE simulations, omitting the simulation process for obtaining GPP and ET, and ensuring a certain degree of precision. Results show that, compared with other machine learning algorithms, the ensemble learning model has the strongest ability to reproduce cropland WUE, with a determination coefficient of 0.90 and a root mean square error of 1.54–1.75 . These results indicate that a combination of machine learning methods and remote sensing can reasonably simulate the WUE of agricultural ecosystems.