Deep Convolutional Neural Networks for Shark Behavior Analysis

Abstract

An important step in the study of free-ranging animals is to perform automatic identification and estimation of their natural different behaviors. This task is especially challenging for the species in the aquatic environment, for example, California horn sharks (Heterodontus francisci). Because they are relatively small, demersal, and active in the nighttime. It is quite impossible to conduct the observations in a continuous direct way. The shark lab at California State University Long Beach (CSULB) conducted laboratory trials to quantify acceleration signatures of horn sharks for different behaviors including resting, swimming, feeding, and nondeterministic movement (NDM). Currently, most of the existing methods have applied machine learning algorithms to estimate sharks’ different behaviors. However, there is still a lack of an efficient and effective way to conduct automatic prediction. In recent years, deep convolutional neural networks have shown the great promise in various computational biology, bioinformatics and neuroscience areas such as biological image analysis, gene expression pattern representation, 3D neuron reconstruction, automatic tumor detection, etc. In this work, we propose novel deep learning models to automatically classify four different shark behaviors using overall dynamic body acceleration (ODBA) through laboratory trial data sets. In specific, we design three deep convolutional neural networks (CNNs) to make fast and accurate predictions and classification. We perform thorough experiments, and the experimental results demonstrate that our proposed models overall produce the best performance than the prior traditional machine learning methods.