Examination of the spatial resolution and discrimination capability of various acoustic seafloor classification techniques based on MBES backscatter data
This thesis focuses on two major topics regarding acoustic seafloor classification techniques. The first topic is about acoustic class separation which affects the discriminative power of classification techniques and the quality of final results. The second topic is the spatial resolution of seafloor acoustic maps that is fundamentally coupled with acoustic class separation. The approach followed here, a) employs an advanced unsupervised classification technique and b) analyzes its implications on the angular response analysis (ARA) of acoustic backscatter. Moreover, a novel approach for improving the ARA technique is described. Applying an unsupervised Bayesian technique that performs an internal cluster validation test, we obtain objective classification of the entire backscatter dataset. This technique utilizes single-angle backscatter measurements from the middle range of the sonar swath offering better discrimination of acoustic classes. The main advantages of the Bayesian technique are that it does not require sonar calibration, it resolves along-swath seafloor variations and that it outputs ordinal categorical values for acoustic classes. Furthermore, the concept of the Hyper-Angular Cube (HAC) is applied and its results are compared with the Bayesian classification results. The HAC is built by several angular backscatter layers which can result either by interpolation of dense soundings or by normalization of backscatter mosaics at different incidence angles. The high dimensional data of the HAC is suitable for supervised classification using machine learning techniques and restricted amount of ground truth information. This approach takes angular dependence of backscatter into consideration and utilizes hydro-acoustic and ground truth data in a more efficient way than it was possible until now.
Use and reproduction:
No CC License (german copyright law applies)