The sensitivity kernel perspective on GRACE mass change estimates

Abstract

Mass change inferences from GRACE and GRACE-FO typically involve, first, the preparation of spherical harmonic (SH) datasets on global gravity field changes and, second, their subsequent analysis that leads to mass change estimates. This study addresses the second step, which builds on SH input datasets that comprise the monthly gravity field solutions as well as amendments to low-degree components and subtraction or re-addition of certain modeled geophysical signals. A variety of methods have been developed to estimate mass changes from SH input datasets. It remains a challenge to assess and compare different methods adopted by different studies and to understand the mechanisms by which their results differ. Methods are often distinguished as belonging to either the inverse or direct approach. In the inverse approach, mass changes are estimated using a set of predefined spatial patterns. In the direct approach, surface mass density variations are integrated by using a predefined weight function, or sensitivity kernel. In this paper, we recall that sensitivity kernels are inherent not only to the direct approach. They are also inherent and may be made explicit, for inverse approaches. We prove that certain implementations of the direct and inverse approach have identical sensitivity kernels, and are therefore equivalent, under the condition that they rigorously incorporate the same signal and error covariance information. We present sensitivity kernels for the example of four different methods to estimate Greenland Ice Sheet mass changes. We discuss the sensitivity kernels in relation to the underlying differences in the methods. We propose to use sensitivity kernels as a means of communicating, assessing and comparing methods of mass change estimates. Once the sensitivity kernels associated to a method are made explicit, any user can readily investigate the method in terms of leakage effects, error propagation from the input SH datasets, or effects of the choice of the SH input datasets.