Time-variant error characterization of SMAP and ASCAT soil moisture using Triple Collocation Analysis

Abstract

Knowledge about spatiotemporal error characteristics of remotely sensed soil moisture (SM) products is essential for correctly interpreting observational information and optimally assimilating them into hydrological models. This work aims to (i) investigate the relative difference between time-invariant and time-variant daily SM errors of Advanced Scatterometer (ASCAT) and Soil Moisture Active Passive (SMAP) products and (ii) analyze correlation of the daily SM errors with surface biomass quantified by the Leaf Area Index (LAI) and rainfall. The time-invariant error denotes an aggregate error magnitude during the whole investigation period and the time-variant error at daily time scale also refers to an aggregate error magnitude but for the period of the 100-day time window centered at the day to be processed. The time-invariant and daily SM errors are estimated using the Triple Collocation Analysis (TCA) applied to ASCAT and SMAP SM retrievals along with the Global Land Data Assimilation System version 2.1 (GLDAS2) SM product for the period of April 2015 – January 2020. Results indicate the relative difference between time-invariant and time-variant daily errors is notable for both ASCAT and SMAP SM products. The relative difference in units percentage denotes the ratio of difference value between time-invariant and daily errors to time-invariant error itself. The daily TCA error fluctuated at 43% value from time-invariant error for ASCAT SM and at 47% value for SMAP SM on a global average. When averaged globally, temporal mean of time-variant daily errors was relatively smaller than time-invariant error by -27% for ASCAT SM and -18% for SMAP SM, respectively. In tropical areas, the relative difference between time-invariant and daily errors is large during the dry season and becomes small when rainy season comes. When the peak lagged correlation is used, daily errors exhibit a stronger correlation with rainfall than they do with LAI in 61% of landmass pixels for ASCAT SM and 66% of landmass pixels for SMAP SM. LAI cannot be used to predict temporal variability of time-variant SM errors in barren land. Lagged correlation analysis reveals rainfall peak coincides with SM error peak in areas featured with low vegetation cover, including barren land, grasslands, and open shrublands. By contrast, the LAI peak comes after the SM error peak in all cases. Savannas and woody savannas are a special case as SM error peak comes first, followed by rainfall then LAI peaks. In summary, ASCAT and SMAP time-variant error varies with a large deviation from time-invariant errors and its temporal variation shows a stronger association with rainfall than changing LAI.