A novel approach for line-based correspondence in multi-sensor high-resolution satellite image georegistration

High-resolution satellite imagery offers a cost-effective and dependable means for acquiring geospatial data and conducting analysis. Georegistering such images is imperative for change detection and analysis of time-series image sequences. While most georegistering algorithms rely on points, linear features have the advantage of improving the geometric strength of the georegistering process and are found in most satellite images. For precise georegistration, we first need to identify accurate conjugate lines between satellite images. This paper presents an automated technique to extract straight lines from satellite images and directly find matching pairs of straight lines among satellite images. Our technique draws its foundation from the Scott and Longuet-Higgins method. However, to accommodate large changes in geometry and terrain variations that are common in satellite images, we modified the algorithm by introducing a new proximity matrix. Our proximity matrix is based on different forms of line-based 2D transformation models. In addition, to account for occlusion and illumination variations along straight lines, we proposed a new measure to estimate radiometric likeness between two straight lines. We tested our algorithm on satellite images with different resolution covering two sites with different topography and elevation changes. Our findings indicated that by utilizing the 2D line-based projective transformation model, a 95% success rate was attainable when matching various pairs of straight lines. Furthermore, the outcomes demonstrated that incorporating radiometric data into the matching process did not lead to substantial changes in the results.