TWO STAGE REGISTRATION OF REMOTE SENSING IMAGES WITH PREDICTIVE UNCERTAINTY

Remote sensing image registration (RSIR) is a process of spatial alignment of two or more images by a predicted spatial transformation or deformation, which can serve as an important preprocess stage for many tasks, such as image fusion, image stitching, and change detection. However, most existing RSIR methods focus just on improving the registration accuracy, overlooking the prediction uncertainty which may harm downstream tasks. In this paper, we propose an uncertainty term in a two-stage registration method’s loss to learn registration and uncertainty jointly. Additionally, we propagate the uncertainty to two downstream tasks, image stitching and change detection, to get more optimized results. Experiments on a self-assembled dataset demonstrate the effectiveness of the proposed method.