Change detection involves comparing two images to find the differences between them. With sonar data, this can be used to identify changes to both the seafloor and objects located on it. As well as mine-hunting applications, it has been used to spot changes made to the seafloor by bottom-feeding fish, detect the effects of sediment transport on the surface topography of the seafloor, and identify changes to the water properties.

The ability to perform change detection is dependent on the images from the two passes of the sonar being precisely registered (aligned). The navigation data alone is not sufficiently accurate to perform this registration. Therefore, the registration must be corrected via a data-driven approach, estimating the errors from the data itself. Registering images of a bland seafloor has been described as the worst-case scenario due to the lack of distinguishing features to match between the images.

This thesis presents a number of modifications to an existing rough facet synthetic aperture sonar (SAS) simulator to allow the generation of suitable repeat-pass imagery. The implementation of the scattering model is extended to allow the temporal decorrelation of a facet between passes to be simulated. These repeat-pass images are validated against published statistical models for the coherence of two speckle images.

As with any coherent imaging system, SAS imagery is corrupted by multiplicative speckle noise. Despite its random appearance, speckle is a deterministic process. If the images are aligned, then the speckle patterns in corresponding areas are coherent (assuming the temporal decorrelation is sufficiently low). This implies that the speckle patterns in two images can be used to register them. A model of the position errors caused by incorrect track information is presented, along with a block-based correlation method to estimate these errors from two images. Using these detected errors, parameter estimation can be used to find the translation and heading errors in the assumed sonar track.

Bland repeat-pass images generated by the simulator are used to test the performance of this registration algorithm. It is shown to be robust against both additive noise and temporal decorrelation, and capable of detecting the track errors. Guidelines are provided for both the design of the system and the implementation of the presented algorithm in order to maximise the registration accuracy.