Fusion Local Descriptors To Improve Multispectral Satellite Image Matching

Image matching is determining the correspondence between two images of the same scene. It is considered one of the most critical processes in computer vision and remote sensing. This paper aims to match multispectral satellite images using the fusion of feature-based descriptors. The images used in this article were obtained from the Sentinel-2 satellite. Reference images are RGB, and query images are SWIR, which were subjected to geometric transformation. To match these images, first, the corner and edge features are extracted using the phase congruency algorithm, and the PC image is made by combining these two features. Then, using the KAZE, key points are extracted from the PC image. In the next step, the extracted features are described by RIFT2 and KAZE and then merged. By combining these two descriptors, a more robust descriptor was created for radiometric and geometric transformations. The proposed method, with an accuracy of 96.71 and repeatability of 0.0397, performed better than feature-based matching algorithms SIFT, KAZE, RIFT2, and deep learning algorithms LOFTR and Superglue.