Understanding the spatial patterns of evapotranspiration estimates from land surface models over China

Land surface models are important tools to represent and predict the spatiotemporal variability of evapotranspiration, which is a key variable in terrestrial water, energy and carbon cycles. However, evapotranspiration estimates from land surface models may suffer from various uncertainties in land surface modeling. Therefore, assessing the performance of evapotranspiration simulation plays a vital role in understanding the deficiencies of land surface modeling. Most of the evaluation studies of modeled evapotranspiration relied on comparisons with flux site observations (point scale) and water budget-derived evapotranspiration (basin scale), which have certain drawbacks and limitations. Thus, the evaluation results may be misleading for understanding the performance of land surface models on representing the spatial variability of evapotranspiration. In this study, a thorough spatial evaluation of the new and reprocessed Global Land Data Assimilation System evapotranspiration products is performed across China based on three bias-insensitive spatial evaluation methods, including the empirical orthogonal function analysis, the connectivity analysis and the fractions skill score. These evapotranspiration products were estimated from three land surface models, namely Noah, VIC and CLSM. The conventional evapotranspiration evaluation against eddy covariance measurements is also performed. The results show that all three products have consistent trends with the observed evapotranspiration series at both daily and monthly time scales. Noah and VIC have comparable performances in terms of different statistic metrics and outperform CLSM at both time scales. The spatial evaluation methods can provide additional valuable information to diagnose the model errors. VIC has the worst spatial performance during the warm months. Despite its inferior performance in late winter and early spring, Noah, overall, has the best spatial performance among the three. The gained insights of this study can help to improve the spatial performance of these models and further promote the system development.