Efforts to extract information regarding the surficial composition of the ocean bottom have increased in the last decade as increases in the availability of computing power have corresponded with advances in signal processing techniques. The ability to extract information from acoustic echosounders is especially desirable due to the relatively low cost and ease of deployment of such systems. Products already exist for the acquisition and logging of echosounder returns.
An acoustic return is comprised of the incoherent backscatter from individual scatterers within the annulus of insonification that occurs when a spherically-spreading transmit pulse intersects with the ocean floor. The return is a convolution of the source ping, and the impulse response modeled by the backscatter profile. Most echosounders, generate an envelope of the received signal. The bottom impulse response undergoes a dilation linear with depth due to simple geometry which can be corrected with time-scale normalization. Under certain circumstances it may be necessary to deconvolve the source ping from the envelope of the return prior to time-scale normalization. It is shown that this can be done by modelling the envelope generation function with a finite sum discrete convolution and the Hilbert transform of the source signal. A second-order Volterra kernel can be derived using a standard predictor network with constrained optimization.
Other factors which contribute to the quality of the return include off-vertical transducer angles which in fact improve the classification by eliminating the nulls that occur in the bottom impulse response due to transducer beam pattern. Spatial averaging can have the effect of beam widening if the transducer angle varies.
Simple feature extraction algorithms are shown to be moderately effective in providing separability. The computational cost of combining the resulting feature sets can be reduced if the individual feature sets are scaled appropriately, reduced and then combined, prior to a reduction to the final dimensionality. The resulting feature space axes contain contributions from both the principal axes of the individual feature sets, as well as cross-algorithmic terms.
Blind clustering of the data is provided through a two-step modification of the K-means algorithm. The first step generalizes it to use arbitrary classification metrics, and the second embeds this generalized kernel within a second kernel which modifies the covariance. The resulting K-stats kernel is very robust when successively applied to a growing number of clusters. / Graduate
Identifer | oai:union.ndltd.org:uvic.ca/oai:dspace.library.uvic.ca:1828/9808 |
Date | 01 August 2018 |
Creators | Caughey, David Arthur |
Contributors | Kirlin, R. Lynn |
Source Sets | University of Victoria |
Language | English, English |
Detected Language | English |
Type | Thesis |
Format | application/pdf |
Rights | Available to the World Wide Web |
Page generated in 0.0015 seconds