In this paper, we study the use of image structures as observations in the context of geophysical data assimilation. Instead of assimilating the apparent radiance values of an image sequence to monitor the physical state, we here focus on the shape comparison between a variable of the model and its observation given by images. More precisely, three observation operators based on wavelet transform, image gradient structure and image gradient orientation are proposed and compared with the basic pixel-to-pixel comparison. We also study formally the influence of observation error and its description into the assimilation process in order to define a robust cost function. The effectiveness of the proposed approach is demonstrated on twin experiments performed on a 2D shallow-water model.