Performance evaluation of linear and multi-linear subspace learning techniques for object classification based on underwater acoustics

Underwater object identification based on acoustic sequence is a complex task, mainly, because of the non-stationary nature of the underwater environment. Moreover, the ambient conditions contribute heavily to varying temporal and spectral characteristics of the source. Further, the characteristic features of a source lie within its spectrum whereas pure spectral contents are more robust to variations along the time and frequency axis. In this work, performance of two different class of learning approaches i.e. linear and multi-linear subspace learning, have been evaluated. Moreover, spectral features are used as inputs to both the said approaches. Two linear subspace learning techniques, namely, principal component analysis (PCA) and linear discriminant analysis (LDA) along with one multi-linear subspace learning (MSL) technique, namely, multi-linear principal component analysis (MPCA) have been used. Performance of the system was evaluated using two sets of data i.e. raw acoustic dataset having samples belonging to 4 distinct classes of ships and a synthetic dataset downloaded from DOSITS, having acoustic samples belonging to 20 distinct classes of underwater objects i.e. sea species and man-made objects. For raw acoustic database, ships signatures were collected in the Indian ocean. Further, two-pass split window (TPSW) method was used to remove background noise from the processed raw acoustic samples. For classification, two neural classifiers were used, namely, robust variable learning rate feed-forward neural network (RVLR-NN) and convolution neural network (CNN). All simulations have been conducted in MATLAB. Further, the system was evaluated under the effect of noise i.e. additive white Gaussian noise (AWGN) at different levels of signal-to-noise ratio (SNR). In addition, dimensions of the feature set were also varied and effects of dimensionality reduction on classification accuracies were observed. Simulation results observed have shown that the combination of MPCA-CNN produced best classification results with an accuracy of up to 99.4%.