Feature Extraction, Selection, and K-Nearest Neighbors Algorithm for Shark Behavior Classification Based on Imbalanced Dataset

Abstract

This paper presents the feature extraction, selection and K-Nearest Neighbors (K-NN) algorithm to classify behaviors of sharks based on the data collected by tri-axial acceleration data loggers (ADLs). Because these behaviors are hard to observe in the wild and do not occur frequently, being able to adequately classify them is extremely challenging. In the proposed scheme, we first employ several transformations to enrich the static and dynamic acceleration data. Then, the enhanced data is converted from time to the frequency domain through the fast Fourier transform (FFT) for noise removal. A modified K-NN approach integrated with feature selection is developed and conducted on the frequency domain data to improve the F1-score for minority classes. The acceleration data of California horn sharks ( Heterodontus francisci ) gathered through ADLs mounted on the first dorsal fin is used to demonstrate our algorithm. A comparison study shows that the features extracted and selected by our proposed scheme can significantly improve the performance of classification on the imbalanced dataset.