Detection of north atlantic right whale upcalls using local binary patterns in a two-stage strategy
📝 Abstract
In this paper, we investigate the effectiveness of two-stage classification strategies in detecting north Atlantic right whale upcalls. Time-frequency measurements of data from passive acoustic monitoring devices are evaluated as images. Vocalization spectrograms are preprocessed for noise reduction and tone removal. First stage of the algorithm eliminates non-upcalls by an energy detection algorithm. In the second stage, two sets of features are extracted from the remaining signals using contour-based and texture based methods. The former is based on extraction of time-frequency features from upcall contours, and the latter employs a Local Binary Pattern operator to extract distinguishing texture features of the upcalls. Subsequently evaluation phase is carried out by using several classifiers to assess the effectiveness of both the contour-based and texture-based features for upcall detection. Experimental results with the data set provided by the Cornell University Bioacoustics Research Program reveal that classifiers show accuracy improvements of 3% to 4% when using LBP features over time-frequency features. Classifiers such as the Linear Discriminant Analysis, Support Vector Machine, and TreeBagger achieve high upcall detection rates with LBP features.
💡 Analysis
In this paper, we investigate the effectiveness of two-stage classification strategies in detecting north Atlantic right whale upcalls. Time-frequency measurements of data from passive acoustic monitoring devices are evaluated as images. Vocalization spectrograms are preprocessed for noise reduction and tone removal. First stage of the algorithm eliminates non-upcalls by an energy detection algorithm. In the second stage, two sets of features are extracted from the remaining signals using contour-based and texture based methods. The former is based on extraction of time-frequency features from upcall contours, and the latter employs a Local Binary Pattern operator to extract distinguishing texture features of the upcalls. Subsequently evaluation phase is carried out by using several classifiers to assess the effectiveness of both the contour-based and texture-based features for upcall detection. Experimental results with the data set provided by the Cornell University Bioacoustics Research Program reveal that classifiers show accuracy improvements of 3% to 4% when using LBP features over time-frequency features. Classifiers such as the Linear Discriminant Analysis, Support Vector Machine, and TreeBagger achieve high upcall detection rates with LBP features.
📄 Content
1
DETECTION OF NORTH ATLANTIC RIGHT WHALE UPCALLS USING
LOCAL BINARY PATTERNS IN A TWO-STAGE STRATEGY
Mahdi Esfahanian, mesfahan@fau.edu* Hanqi Zhuang, zhuang@fau.edu Nurgun Erdol, erdol@fau.edu Edmund Gerstein, egerste1@fau.edu
Department of Computer and Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL 33431, USA
- Corresponding author. Tel: +1 561 929 6392 E-mail address: mesfahan@fau.edu 2
Abstract In this paper, we investigate the effectiveness of two-stage classification strategies in detecting north Atlantic right whale upcalls. Time-frequency measurements of data from passive acoustic monitoring devices are evaluated as images. Vocalization spectrograms are preprocessed for noise reduction and tone removal. First stage of the algorithm eliminates non-upcalls by an energy detection algorithm. In the second stage, two sets of features are extracted from the remaining signals using contour-based and texture based methods. The former is based on extraction of time- frequency features from upcall contours, and the latter employs a Local Binary Pattern operator to extract distinguishing texture features of the upcalls. Subsequently evaluation phase is carried out by using several classifiers to assess the effectiveness of both the contour-based and texture-based features for upcall detection. Experimental results with the data set provided by the Cornell University Bioacoustics Research Program reveal that classifiers show accuracy improvements of 3% to 4% when using LBP features over time-frequency features. Classifiers such as the Linear Discriminant Analysis, Support Vector Machine, and TreeBagger achieve high upcall detection rates with LBP features.
Index Terms— North Atlantic Right Whale, Acoustic Monitoring, Upcall Detection, Local Binary Patterns, Classification.
- Introduction The North Atlantic Right Whale (NARW), (Eubalaena glacialis), is one of the critically endangered species with an estimated count of fewer than 400 individuals in the North Atlantic Ocean [1, 2]. Right whales have been the “right” whale to hunt leading to their extreme low 3
numbers. North Atlantic right whales can be found in coasts of U.S. and Canada ranging from Bay of Fundy in Canada in the summer to Florida and Georgia coasts in the winter [3, 4]. Since their habitant regions are contaminated by human activities such as shipping traffic and fishing vessels, the anthropogenic mortality from collision with ships and entanglements in fishing gear are considered the main causes of their low population [1]. Therefore, many long-term studies have been conducted with the goal of monitoring NARWs in their habitats, especially where anthropogenic activities have increased. Results of such studies are also useful to marine biology research in conservation and behavioral changes. North Atlantic right whales produce a variety of vocalizations but our focus is mainly on one of the most commonly heard sounds known as “upcalls” or “contact call” which is characterized by an upsweep frequency from 50 to 350 Hz [5]. These are stereotyped frequency-modulated (FM) calls, about a second in duration. Their function is believed to establish and maintain contact between right whales. Variability of these contact calls is significant due to changes in initial frequency, FM rate, duration and bandwidth. Since upcalls are very common in their vocal repertoire and are highly species-specific, they are used as the primary basis for the acoustical detection of right whales [5, 6]. Other NARW vocalizations described in the literature as tonals, gunshots, hybrid, pulsive, and downcall [7] also exist but their proportion varies by season and habitat. Generally speaking, the dominant thought has been that the most accurate method for detecting NARW calls in large data sets is to employ human operators to evaluate data spectrograms visually and corroborate the results aurally. This method not only requires a great deal of time and can incur large labor costs, it is also limited by operator judgment which is often subjective and may include false detections. In recent years, Passive Acoustic Monitoring (PAM) 4
[8, 9] has received wide acclaim as one of the most effective techniques for detecting and localizing marine mammals. Acoustic data are also a fundamental component of studying the behavior of many cetaceans. One of the challenges that arise in using PAM systems is the level of ambient noise which can vary considerably over the course of data collection. This can make the data analysis difficult especially in low SNR environments. Another problem in the design of automatic detectors is the lack of a priori knowledge about noise and signals. The probability density function and the power spectral density (PSD) of signals and noise are both unknown. In addition, frequency components of upcalls may change depending on th
This content is AI-processed based on ArXiv data.