The ocean floor, which covers approximately 70% of the Earth's surface, still contains vast areas which are largely unexplored. In truth less of the ocean floor has been mapped than either of the surfaces of the Moon or Venus, which has been mapped in recent years by NASA's Magellan spacecraft. Developments over the last 20 years have produced tools which allow the research community and industry to map the ocean floor in greater detail and over more extensive area than previously possible. These tools are exclusively remote sensing ones and are extremely varied. Their application covers a wide range of the electromagnetic spectrum and include ship-borne gravity and magnetics; satellite measurement of sea surface height and temperature; and seismic data. reflection and refraction, as well as sidescan sonar and swath bathymetry techniques which produce images of the seafloor. All of these methods, their associated processing techniques and the interpretations they produce are invaluable for research into the processes which shape the ocean floor.
ParticipantsProject aims
We employ state of the art textural image analysis and computer programs in the form of neutral networks to develop an automated technique for seafloor classification. Research into textural image analysis is a fairly recent field as is the use of neural networks but their application to sea floor remote sensing is a new frontier. Textural image analysis techniques have been developed for medical, recognition of man-made structures and patterns and whilst there is a reasonable bank of knowledge for standard statistical methods, involving gray level, probability distribution functions of different forms there has been little application of the techniques to seafloor characterization. In more recent times novel approaches to textural analysis involving unitary sinusoidal transforms have been developed for texture analysis and classification. Textural analysis can be used in two ways, (1) to extract features or (2) classify an image. Once an image has been analyzed and classified, by looking at as small an area as is possible it is then necessary to use these classes and classify the entire image rather than isolated areas. Neutral networks are being used for this task due to their capability to be trained.
