whole_BuelensBartAndreHendrikLutgart2008_thesis.pdf (22.9 MB)
Scientific data mining for spatio-temporal hydroacoustic data sets
thesisposted on 2023-05-26, 22:50 authored by Buelens, Bart Andre Hendrik Lutgart
Managing natural marine resources for sustainable exploitation of the oceans and the flora and fauna they contain is a challenging task. Decisions by policy makers are based on advice from the scientific community. Through surveying and monitoring programs, scientists study the marine environment to gain insight into its structure and function. Employing acoustic techniques, sonar systems are often the best tools available to effectively observe aquatic environments. Important applications include fisheries and seafloor mapping. Fish stock assessments are typically conducted using single beam echosounders, while bathymetric surveys are conducted with multibeam sonar. Multibeam sonar instruments that are capable of collecting samples for the complete water column are an emerging technology. Since they collect acoustic data over much greater sampling volumes than single beam instruments, significant improvements in fisheries studies are expected. The combined collection of seafloor and water-column data will lead to survey cost savings and to an integrated, ecosystem-based approach to monitoring the ocean environment. While standard data analysis procedures are established for single beam fisheries and standard multibeam bathymetric applications, this is not the case for full water-column multibeam sonar data. In this thesis, a data mining approach for handling such data is proposed. The developed method consists of a preprocessing algorithm based on an inversion technique, followed by a pattern analysis algorithm using kernel clustering methods. The preprocessing algorithm applies a deconvolution as a model inversion method to reduce the data set in size and to convert the acoustic measurements into a generic vector representation. Each vector has a spatial and a temporal component as well as a number of additional features typically relating to the acoustic backscatter energy. These spatio-temporal vectors are then subjected to pattern analysis algorithms. Two clustering algorithms are selected: a density based spatial clustering algorithm, and a clustering algorithm based on kernel methods. A new method is developed to allow the kernel clustering algorithm to make use of the spatial and non-spatial components of the data in a combined fashion. This results in a powerful, flexible and versatile clustering procedure. The outcome is a segmentation of the data into coherent structures, for example fish schools and the seabed. Classification is achieved through the differentiation between data clusters indicative of different fish species or seabed habitats. The effectiveness of the data mining methods is demonstrated in a number of case studies. It is hoped that the developed approach will facilitate routine use of water-column multibeam sonar data for fisheries applications in particular, and for ecosystem studies and marine resource management in general.
Rights statementCopyright 2008 the author No access or viewing until 16 November 2010. Thesis (PhD)--University of Tasmania, 2008. Includes bibliographical references.