Analysis of image-to-image registration performance of multitemporal KOMPSAT-3∙3A imagery with respect to convergence angle

Unlike numerous studies on the impact of a convergence angle between two stereo images, its effects on image-to-image registration have not been extensively investigated. In this study, we analysed the impact of the magnitude of the convergence angles of very-high-resolution (VHR) satellite image pairs on registration accuracy with respect to the two-dimensional (2D) geometric alignment. To perform image-to-image registration, we used an iterative process to extract a large number of well-distributed conjugate points (CPs) by combining area- and feature-based matching approaches. To evaluate the proposed method’s ability to extract CPs and demonstrate the relationship between the registration performance and the magnitude of the convergence angles, we conducted experiments using KOMPSAT-3 and KOMPSAT-3A satellite images acquired over Daejeon, South Korea. A total of 18 satellite images were used to generate 153 image pairs with different convergence angles. Two transformation models, affine and piecewise linear transformations, were used independently to warp a target image to a reference image, and their performances were compared. An analysis of the affine transformation with respect to the size of the convergence angles revealed that a larger convergence angle between images resulted in lower accuracy of the registration performance. This result differs from the convergence angle condition required in stereo images to extract accurate 3D information. However, the piecewise linear transformation achieved a reliable registration performance regardless of the magnitude of the convergence angles.