Information based uncertainty decomposition in SMAP modified Dual Channel Algorithm retrieve soil moisture

Abstract:

Soil moisture is a key component in earth system dynamics; yet accurate observation of soil moisture remains a challenge. The recent launched NASA-SMAP mission focuses on characterizing surface soil moisture globally at a relatively fine temporal scale. Although the SMAP soil moisture products have been validated against in situ soil moisture measurements via a sequence of core and sparse validations sites, there is still a lack of systematic characterization of when, where, and why the retrieval algorithm deviates from the in situ observations. Here we decomposed the information loss from horizontally (T_Bh) and vertically (T_Bv) polarized brightness temperature input to SMAP modified Dual Channel Algorithm (MDCA) model output via a mutual information approach. The results shown that on average 12% of the uncertainty is caused by the loss of information in the MDCA model itself while the remainder is induced by inadequacy of T_Bh and T_Bv datasets. We find the fraction of MDCA induced uncertainty is positively correlated (pearson r of 0.28) with errors between in-situ observations and SMAP estimates. A decomposition of mutual information between T_Bh, T_Bv and MDCA soil moisture shows that on average 55% of the information is redundantly shared by T_Bh and T_Bv, while 14% of the information content of output is provided when they are synergized. The information provided uniquely from both T_Bh and T_Bv is 15%. The information redundantly provided by T_Bh and T_Bv was found to be tightly correlated (pearson r = -0.7) to how well the in situ and MDCA output were correlated. Thus, as T_Bh and T_Bv provide more redundant information for the MDCA its overall quality improves. This suggests the informational redundancy between these remotely sensed observations can be used as independent metric to assess the overall quality of algorithms using these data streams. This study provides a baseline approach that can also be applied to evaluate other remote sensing models and understand informational loss as satellite retrievals are translated to end user products.