A framework for estimating actual evapotranspiration through spatial heterogeneity-based machine learning approaches

Actual evapotranspiration (ET) is a key variable controlling the exchange of energy and water in terrestrial ecosystems. The spatial heterogeneity of ET patterns poses a great challenge for regional ET estimation over heterogeneous landscapes. In this study, we proposed a framework for estimating ET through interactive detector for spatial associations (IDSA)-based machine learning (ML) approaches by combining data from remote sensing and four flux towers distributed in the Hai River Basin. The IDSA model was applied to explore individual and interactive determinants of ET over the Hai River Basin. In addition, the geographical regions of ET were determined according to the spatial heterogeneity. Then we simulated ET in each geographical region separately using three ML models, including random forest, gradient boosting decision tree, and Cubist. General ML models without considering spatial heterogeneity were used for comparison, and evaluations were conducted with the eddy covariance flux tower observations at the site scale, also with water balance ET at the basin scale. The results demonstrated that the spatial patterns of ET were difficult to be explained by individual environmental variables (64.9 %). The maximum interpretability was improved by about 11.2 % through the interaction of air temperature (T) and normalized different vegetation index (NDVI). The framework developed in this study had excellent performance with coefficient of determination (R2) ranging from 0.807 to 0.811, root mean square error (RMSE) ranging from 0.654 mm/day to 0.661 mm/day, and mean absolute error (MAE) ranging from 0.470 mm/day to 0.485 mm/day. It was superior to general ML models at the site and basin scales. Additionally, the ET simulated by IDSA-based ML models were in good agreement with the reference product, further illustrating its reliability. The proposed framework provides a theoretical basis for in-depth understanding of spatial heterogeneity of ET and a new perspective for ET prediction over large scales.