An approach to seafloor classification using fuzzy neural networks combined with a genetic algorithm

Seafloor classification, either in terms of physical properties or geological provinces, is important in many fields including marine geology, hydrography, marine engineering, environmental sciences, and fisheries. The purpose of this research is to develop automated seafloor classification algorithms of the backscatter data from multibeam sonar. The algorithms include the processing of backscatter data, feature extraction and selection, and the classification approaches involved.; The raw backscatter data from multibeam sonar must be analyzed and processed, because it is difficult to use directly in the classification process. This research focuses on the correction algorithms of local bottom slope and near nadir reflection influences. Through these corrections and other compensations, we can obtain processed backscatter strength data which better reflects the features of the seafloor. These processes provide the data foundation for the later classification process.; Feature extraction and feature selection are essential steps in order to optimize a pattern classification system. In this work, a feature subset selection method using a Genetic Algorithm (GA) combined with a fuzzy ARTMAP neural network (FAMNN) classifier is proposed. The complete feature set is encoded in a chromosome and then optimized by GA algorithms with respect to both classification accuracy and number of selected features. The experimental results show that the classification performance is improved or at least kept similar using the feature set of 4 or 6 features selected from all 24 features.; Neural network classifiers are nonparametric and thus more robust than traditional statistical classifiers that typically require knowledge of the underlying probability distributions. In this thesis, at first a simplified fuzzy ARTMAP neural network (FAMNN) is investigated for the seafloor classification. The performance effects of variations in choice parameter, vigilance parameter, baseline vigilance parameter, voting strategy and the size of the training set are examined with a real data set. The performance of the FAMNN classifier has been compared with the traditional Bayesian classifier as a statistical benchmark on the same database. The FAMNN classifier outperforms the traditional Bayesian classifier in terms of every seabed type and total classification accuracies. However, the classification performance of the FAMNN classifier is highly dependent on an adequate number of samples to train the classifier. A novel fuzzy ARTMAP neural network variant (GA-FAMNN) is proposed which employs a GA strategy to search and generate new input pattern samples to fall near the boundaries between categories. The FAMNN classifier undergoes supervised training again with the original existing training set and the new augmenting samples. The two experimental results illustrate that the performance of the retrained FAMNN classifier has evidently improved using the proposed method. This is particularly so when there are a relatively small number of ground-truth samples.