LAUSR

Abstract Recognition of seabed sediment is one of the critical foundations of marine exploitation. This paper proposes a probabilistic neural network (PNN) based method to improve the identification accuracy of seabed sediment from side-scan sonar imagery. The feature set of side-scan images consists of two types of features, namely textural features and color features. In this study, partial eigenvalues of the gray co-occurrence matrix are selected as the textural feature, and the color features are represented by color moments. Combining textural features with color features, we get the input matrix, which is then fed into PNN for classification. PNN calculates the distance between the sample eigenvector to be predicted and the training sample eigenvector, then accumulates the probability belonging to a certain category. Finally, PNN outputs the forecasted class of samples eigenvector with the largest posterior probability. It is the first time that PNN has been used in seabed sediment classification from side-scan sonar imagery. Compared with the traditional clustering methods, PNN improved the accuracy of the classification and attained a highest accuracy of 92.2%.