|Title||Block Bundle Adjustment of Landsat 7 ETM+ Images over Mountainous Areas|
|Source||Photogrammetric Engineering and Remote Sensing vol. 69, no. 12, 2003 p. 1341-1349, https://doi.org/10.14358/PERS.69.12.1341 Open Access|
|Alt Series||Natural Resources Canada, Contribution Series 20181170|
|Publisher||American Society for Photogrammetry and Remote Sensing|
|Media||paper; on-line; digital|
|Subjects||geophysics; remote sensing|
|Program||Canada Centre for Remote Sensing Divsion|
|Abstract||This research study showed the potential of block bundle adjustment with nadir viewing sensor images, such as Landsat 7 ETM+. The method is based on the 3D analytical geometric model developed for
multisensor images at the Canada Centre for Remote Sensing, Natural Resources Canada. Different block sizes and configurations were tested and compared using an image bundle adjustment. The different tests using 15 Landsat 7 ETM+ images (five paths
and three rows) acquired over a study site in the Canadian Rocky Mountains, showed that the same results (around 25-m errors) could be obtained with image blocks as with a single image using a largely reduced number of ground control points (GCPs).
However, the combined image measurement and map errors of GCPs were included in the final error budget, and the internal accuracy of the blocks should be better (around one pixel or less). The number of GCPs to be used depended mainly on the
cartographic data accuracy: more GCPs than the minimum required reduced the error propagation in the least-squares block bundle adjustment. In addition, tie points with a known elevation value (elevation tie points, ETPs) instead of normal tie points
were used to link the adjacent images/strips (north-south and east-west) because the viewing-angle differences of overlapping images were smaller than 1° in north-south overlaps and around 10° in east-west overlaps. Better and more consistent results
were also obtained using strips of images in the blocks acquired from the same orbit and date instead of using independent images. Finally, using only GCPs in the outer strips/images and ETPs in each overlap achieved the same results (25-m errors).