The Performance of Multiple Model-Simulated Soil Moisture Datasets Relative to ECV Satellite Data in China

Reliability and accuracy of soil moisture datasets are essential for understanding changes in regional climate such as precipitation and temperature. Soil moisture datasets from the Essential Climate Variable (ECV), the Coupled Model Intercomparison Project Phase 5 (CMIP5), the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP), the Global Land Data Assimilation System (GLDAS), and reanalysis products are widely used. These datasets generated by different techniques are compared in a common framework over China in this study. The comparison focuses on four aspects: spatial pattern, temporal correlation, long-term trend, and the relationships with precipitation and the Normalized Difference Vegetation Index (NDVI). The results indicate that all soil moisture datasets reach a good agreement on the spatial patterns of wet and dry soil. These patterns are also consistent with that of precipitation. However, there are considerable discrepancies in the absolute values of soil moisture among these datasets. In terms of unbiased Root-Mean-Square Difference (unRMSE, i.e., removing the differences in absolute values), all modeled datasets obtain performances comparable with ECV observations. Our results also suggest that a multi-model ensemble of soil moisture datasets can improve the representation of soil moisture conditions. The optimal dataset from which the wetting/drying trends in soil moisture have the highest consistency in terms of changes in precipitation and NDVI varies by season. Specifically, in spring, CMIP5 in northwest China shows that the trends in soil moisture are consistent with the changes in precipitation and NDVI. In summer, ECV presents the most identical performance compared to the changes in precipitation and NDVI. In autumn, GLDAS and Reanalysis have better performance in south China and parts of north China. In winter, GLDAS performs the best in the east of south China, followed by the Reanalysis dataset. These discrepancies among the datasets present various changes in different regions, which should be well noted and discussed before use