Method of automatic coregistration of digital remote sensing images from different sources

In this paper, a method for the automatic alignment of diverse digital Earth remote sensing images using survey data is proposed. The method is designed to align color, grayscale, multispectral, and radar images, as well as their combinations, with potential differences in spatial resolution of up to four times (optionally – sixteen times). The main stages of the proposed method include: an optional upscaling (up to four times); an optional number of image channels reduction to three or one; keypoint detection, their description and alignment. To achieve a universal and robust solution in the latter stages, the best-known algorithms were compared: SIFT, SAR-SIFT, RIFT, and the trainable RoMa. Experimental studies using the indicated types of space images demonstrate a clear advantage of the trainable neural network model RoMa trained on a variety of heterogeneous images. For additional improvement of the alignment accuracy, we utilized a priori data about the images in the form of their georeferencing information.