Study on Echo Features and Classification Methods of Fish Species

Acoustic identification of fish species is important to effective fisheries resources management. However, classification of fish species is still a challenge in acoustic research on marine animals. This work focuses on classification performance comparisons of three different features with four different methods, in order to find the optimal feature-method combination of twelve kinds of combinations. First, a modified simulation model is introduced for generating fish school echoes of different fish species. Then three kinds of common feature sets including echo waveforms, echo spectra and echo spectrograms are employed for classification performance comparisons. Each feature set is utilized by four kinds of typical classification methods including decision tree, adaptive boosting (AdaBoost), artificial neural networks (ANN) and convolutional neural networks (CNN). The simulation results show that the highest classification rate is reached by CNN with echo spectrograms as the input, which is verified by the subsequent tank experiment.