With the advances of meteorological observation instruments, a variety of meteorological observations can be used in numerical prediction models. However, due to the observational error, especially systematic deviations in unconventional observations, the effect of observational data assimilation has not been fully examined. Thus, a variational assimilation method for temporal and spatial gradient information is proposed to eliminate such errors. The principle is that no a priori estimates of the systematic bias are needed, but a gradient information operator is used to transform the original variables so as to implicitly avoid this systematic error. A series of results of four-dimensional variational assimilation ideal experiments based on a shallow water model shows that this assimilation could completely eliminate the impact of smoothness systematic deviation on the assimilation results. The model could provide a good assimilation effect for the variables having a small value, and could estimate the scope of application for the ones having a large value. Due to the uncertainty of the optimal solution, the assimilation effect absorbs more of the overall temporal and spatial gradient trends of the observation field rather than the observation value itself, which is applicable to observational data with low credibility.