Logo image
Developing an rCNN detector for Rice's Whale (Balaenoptera ricei) long-moan variant calls in the Western Gulf of Mexico
Report   Open access

Developing an rCNN detector for Rice's Whale (Balaenoptera ricei) long-moan variant calls in the Western Gulf of Mexico

Lia Alexa Caldwell
2025-12

Abstract

Rice's whale long-moan passive acoustic monitoring deep learning automatic dector faster-rCNN SouthEast Fishereies Science Center SEFSC High Frequency Acoustic Recording Packages HARPs Machine Learning Gulf of Mexico
Rice’s whales (Balaenoptera ricei) are a critically endangered species of baleen whale endemic to the Gulf of Mexico (GOMx). Understanding the population distribution of this species is critical for informing conservation efforts that support its recovery. To support monitoring of vulnerable marine mammal species throughout the Gulf, the Southeast Fisheries Science Center (SEFSC) has deployed moored SoundTraps (SoundTrap ST500 & ST600, Ocean Instruments, New Zealand), and moored High Frequency Acoustic Recording Packages (HARPs) in partnership with Scripps Institution of Oceanography. Because the acoustic data from these deployments consist of continuous long-duration recordings, automated call detectors and classifiers are used to find Rice’s whale calls in these large datasets. An automatic call detector using a region-based convolutional neural network (rCNN) was recently developed to aid in efficiently detecting and classifying Rice’s whale long-moan and downsweep sequence calls off the West Florida shelf. While the long-moan detector has proven effective for detecting Rice’s whale calls in the eastern Gulf within the species’ known core distribution area, it is less effective at classifying Rice’s whale long-moan call variants in the western Gulf. The purpose of this project was to develop a similar detector to detect and classify western variants of Rice’s whale long-moan calls. We tested several detectors to optimize hyperparameters and to examine how different color limits and single- versus multi-class detection affect model training and performance. We found that the model performed best when trained to detect a single long-moan class and with spectrograms produced with consistent color limits displaying 70 dB range and a minimum at the 5th percentile noise level. The resulting detector achieved an average precision of 0.95 during initial testing and an average precision of 0.86 when applied to an unseen dataset. This model improves current detector capabilities by reducing missed detections and increasing precision, thereby enhancing the detection and classification of Rice’s whale long-moans in the western GOMx, contributing to more efficient workflows and improved understanding of the species’ distribution.
pdf
Caldwell, Lia - Final Report - 01.14.20261.48 MBDownloadView
Open Access

Metrics

1 Record Views

Details

Logo image