Detection of artifacts in bathymetric swath data suing artificial neural networks
Loading...
Files
Date
1990
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Due to the existence of new and more effective learning algorithms, artificial neural networks (ANNs) are now being used in many different fields such as in pattern recognition and classification. A study of the feasibility of using ANNs to detect certain types of artifacts (peaks) that occur in bathymetric swath surveying and to localize user-specified areas of ocean floor, was done. A peak detection network (PDNet) was designed and defined in terms of ANNs. This method of peak detection was labelled as the Neighborhood Averaging Method (NA). The PDNet is a parallel distributed processing and adaptive network that is able to continually monitor its inputs and detect data values outside some expected range (peaks). When some peak is detected the network is able to suggest what the "correct" value should be. Auxiliary average and variance modules (subnetworks) are defined in network terms. The Lateral Inhibition Method (LI), also biologically inspired, was tried for peak detection. This method, implemented as a "narrow" filter, was particularly effective in detecting peaks composed of single data points. Both the NA and the LI methods are data-profile oriented for execution speed. Both methods were tested on swath data considering different values of thresholds and neighborhood sizes. The results of these test were then visually verified using "Swath Viewer", an already existent data visualization tool. Interface software tools were created to facilitate result verification. One concludes that both the NA and LI methods are valuable methods for peak detection in swath data. ANNs were trained using the back-propagation (BP) learning algorithm to detect areas of ocean floor containing certain user-specified features. The BP equations are described. Preliminary tests were made on an IBM PC using profile data to define simple floor types such as ramps and hills. After these tests, a two-hidden layer BP ANN was implemented and tested on swath data. Auxiliary software tools were created to facilitate learning tasks and the handling of data files. One concludes that although ANNs are computational tools with some valuable features such as generalization and learning abilities, when used in swath data applications, such as in feature detection, their usefulness is of questionable value.