Multi-view, broadband, acoustic classification of marine animals

Abstract

Acoustical methods provide rapid, non-invasive, and synoptic tools for studying marine ecosystems. Despite the dramatic advances in this technology during the past three decades, there is presently a large disparity between the demand for quantitative information about marine animals and the capability of acoustic systems to deliver this information. A primary reason for this disparity is the strong dependence of acoustic scatter from marine animals on their size, shape, in situ orientation, and taxa. In a typical setting, these parameters are unknown, and are difficult to determine using existing acoustic methods. To mitigate this problem, a multi-view, broadband approach to marine animal classification and size estimation is investigated in this thesis.

Initially, zooplankton classification was investigated for two ecologically important taxa: copepods and euphausiids. Numerical simulations compared physics-based feature transformations, Nearest Neighbor (NN), and Multi-Layer Perceptron (MLP) classifiers. Results indicate that combining frequency-correlation features with a MLP yields an accurate (> 90 % correct) classification algorithm. Based on these promising results, a laboratory system was developed to recorded multi-view, broadband scatter from live, individual copepods and mysids. Results using frequency correlation features indicate that these features yield very good separation between classes with non-overlapping standard deviations computed from eight individuals per class.

Next, sound scatter data from live, individual fish were used to develop several kernel-machine-based multi-view fusion algorithms. Performance was quantitatively compared as a function of the number of available views, feature spaces, and classification problem type. A collaborative fusion algorithm performs better than the others without requiring any assumption about view geometry, the number of views, or the type of features.

Finally, multi-view fish size and orientation estimation was investigated under three different approaches. Results indicate that classification-based size estimation can be effective with a limited aperture and limited number of views. Model-based and image-reconstruction-based estimation show very good performance with full aperture data.

This thesis demonstrates that the multi-view, broadband approach offers significant advantages for marine animal classification, sizing, and orientation estimation.