Frequency Domain Transformations and CNNs to Predict Unlabeled Shark Behavior With GPS Data

Abstract

This paper provides a comprehensive analysis of frequency domain transformations applied to convolutional neural networks (CNN) to model and predict unlabeled shark behavior in the open ocean with GPS position data. The frequency domain-based CNN networks are compared against the time domain CNN to contrast the two CNN architectures. The shark behavior data were obtained through two datasets where tri-axis accelerometer data were collected from live sharks. The first dataset was from the CSULB Shark Lab and consisted of labeled shark behavior into four shark behavioral categories. The second dataset used in this study was unlabeled and recorded from sharks in the open ocean and had GPS positioning data and depth data points. Findings show that the CNN architecture based on the frequency domain slightly outperforms time-based CNNs for classifying California horn shark behavior. Through spectral density analysis, prominent features are extracted and allow for distinguishing the shark behaviors.