Frequently collected multitemporal multispectral images mostly present strong temporal redundancies that can be exploited for data compression in temporal domain considering the fact that the user already has a previous reference image. While the spatial and spectral prediction model is applied, the compression considering temporal correlation needs to be explored. In this paper a gradient-based temporal prediction approach has been proposed where the image of a scene is predicted from the previously taken image of the same scene. The geometrically co-registered reference image and the recent image are used for sequential prediction in order to minimize the model residual. The model parameters are optimized automatically to achieve minimum residual entropy for lossless compression. Experimental results demonstrate the effectiveness of proposed method, especially when the new data are not highly correlated to the previous data due to the real changes experienced between the two data collection dates
