The Input–Output System Theory (IOST) method is primarily based on the spectral combination of heterogeneous data taking into account their statistical properties and approximating the Power Spectral Density (PSD) functions of the signals and their errors. In this study a Multiple Input–Multiple Output System (MIMOS) is proposed, where the input measurements as well as the input and output signals are different gravity field observables. The work presents a sensitivity analysis of the IOST method towards the input data noise and the effect of integral kernel modifications to the error prediction estimates. Simulated input noise fields along with the contribution of the input data resolution to the output prediction errors are investigated towards the analysis of the 2D error covariance functions. Special attention is paid to the contribution of different kernels to the error prediction estimates and 2D planar and spherical discrete kernels are tested considering the overall prediction improvement of the spectral procedure. The MIMOS system is finally assessed by a number of numerical tests and conclusions are drawn in terms of the optimal modelling of the input data noise and the significance of the spherical kernels in the improvement of the prediction results.