Yuanliang Ma
College of Marine Eng., Northwestern Polytechnical Univ., Xi'an 710072, People's Republic of China
Z. Y. Wang
Northwestern Polytechnical Univ.
G. Gimenez
D. Vray
CREATIS, INSA, Lyon, France
The paper presents a method of lake bottom sediment classification by artificial neural networks (ANN) using wideband echo signals. The samples to be classified were acquired by experiments at Lake Geneva. There are five types of sediments, namely, silt, rocks, pebbles, sand, and a mixture of sand and gravel. The pattern features are extracted from spectra of echo signals in subband energy expression. Different subband divisions for frequency-domain feature extraction are compared and it is shown that the contant Q method provides better results in comparison with the constant bandwidth method. Using the constant Q method in association with a BP-type locally connected neural network, 85.1% correct classification in average has been achieved for a testing data set. Wideband echo signals have outstanding superiority for classification in comparison with narrow-band signals. It contains more information representing the physical and architectural features of targets. The neural network ultilizes the information through careful optimization and provides a performance improvement up to 10%.