Summary made by Quivr/gpt-4
This document is a research paper by F. Sattar et al., published in Applied Acoustics in 2016. The paper presents a scheme for identifying fish vocalizations, specifically grunts and growls, using auditory analysis and machine learning.
The researchers first partition hydrophone recordings of fish data into blocks of specific segments. Each 1D data block is then converted into a 2D feature map. A high-resolution feature set (descriptors) is then constructed from the feature maps and used as input to a Support Vector Machine (SVM) classifier.
The paper also discusses the use of a Sequential Floating Forward Selection (SFFS) algorithm for feature selection, which helps improve classification by removing redundant information in high-dimensionality spaces. The SFFS algorithm finds an optimum subset of features by appending and discarding features from subsets of selected features.
The researchers also discuss the use of a cost function that expresses a combination of two criteria: margin maximization and error minimization. This cost function minimization is subject to certain constraints.
The paper also mentions that the audio data for the research was collected off a private dock located on the east coast of Quadra Island in June 2012. The researchers used an HTI-96-MIN hydrophone for the recordings.
The paper concludes that the high-resolution features extract subtle and detailed information and contain more distinctive information than low-resolution features, making them effective for the identification of fish vocalizations.