Image co-registration at very high spatial resolutions
To follow crop development over time means we need to follow pixels over time. To that end, pixels must align perfectly, and this is hard to achieve. STARS is currently working on co-registration of Very High Resolution (VHR) satellite images in farming areas. The problem is hard due to the absence of stable and reliable ground control points in such images. State-of-the-art methods like SIFT, ASIFT and SURF fail due to absence of strongly oriented geometric structures. We chose trees as the best suited natural ground control points. Tree detection is followed by image matching. Our results obtained so far are satisfactory, but can be further refined with correction of distortion induced by elevation of the tree crowns with respect to terrain. This work will be discussed in detail in the upcoming STARS Technical Workshop in Dakar, Senegal.
- Automatic image registration by tree detection. Master image and detected candidate objects for matching from master (blue) and from slave (green) images, respectively. Circular candidate objects are detected by automatic procedure and are characterized by position and radius. Most, but not all, objects are corresponding to trees. Left: objects from the slave image do not match objects from the master image. Right: objects from the slave image are matched to the objects from the master image and geometric transformation is applied on the slave objects. The transformed slave objects match the master objects well.