|dc.description.abstract||Image matching, or registration, is the process where two or more images are compared to find corresponding areas or objects. Several different methods are used in the approach of quantizing displacements (e.g. cross-correlation methods, Fourier methods, least-square based methods, wavelet based methods (Brown, 1992; Zitova & Flusser, 2003)), and they are being used in a variety of different fields. In geosciences, digital image matching have been used to measure displacements in a range of studies (including mass movements and slope deformations, ice sheet motion, arctic and mountain glacier and rockglacier displacements and terrain model generation).
In this thesis, one spatial domain (normalized cross-correlation (NCC)) and two Fourier based image matching methods (phase and gradient correlation) are compared and evaluated based on different parameterizations and several test images covering glaciers and rock glaciers. Geometric and radiometric corrections are considered, as well as pre and post-processing techniques. Additionally, an experimental software including image matching algorithms has been developed. The code development process and implementations are discussed.
Three cases have been tested, with several tests in each case. Results showed that, compared to NCC methods, Fourier based methods generally were (1) more robust against snow cover and shadow differences, (2) proved to have better filtering capabilities and (3) processed approximately 3 times as fast. NCC based methods however, allowed for more rotation and deformation of image features in the matching process, but generally achieved a lower signal to noise ratio (SNR) in the results. The implemented quad-tree operator, designed and developed to improve the NCC technique by automatically adjusting the reference window sizes, did not achieve significantly more robust results compared to ordinary NCC methods.
Among the algorithms tested in this work, the gradient correlation algorithm is considered the most suitable approach for quantifying glacier displacements from repeat imagery. It is not sensitive to surface cover differences, generally allows for acceptable amounts of image feature deformations, and is one of the fastest algorithms tested.
Results from the rock glacier in Muragl valley (images from 1981 and 1994), Tokositna glacier (2000 and 2001) and Columbia glacier (2002), showed a maximum average displacement of respectively 0.46 m/y, 3.1 m/d and 7.5 m/d.||eng