Fine-tuned Faster-R-CNN for detection and classification of blue and fin whale calls in low-frequency audio data

Read full text Summarise with VIRAT
Authors: Michaela N. Alksne, Marie A. Roch, Kaitlin E. Frasier, John A. Hildebrand, Shane Andres, Dolapo Adesanya, Lauren M. Baggett, Joshua M. Jones, Xiaobai Liu, Ana Širović, Xixin Zhang, Joshua Zingale
+1 more Simone Baumann‐Pickering
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This repository contains the weights for a fine-tuned Faster R-CNN object detection model trained to detect blue and fin whale calls in 60-second spectrograms from passive acoustic recordings. The model detects five call types: blue whale A, B, and D calls, and fin whale 20 Hz and 40 Hz calls. The model uses a ResNet-50 backbone with a feature pyramid network and was fine-tuned from a pretrained checkpoint over three rounds of human-in-the-loop refinement on high-frequency acoustic recording package (HARP) and California Cooperative Oceanic Fisheries Investigation (CalCOFI) sonobuoy recordings from the Southern California Current Ecosystem. The final training dataset comprised 11,676 annotated calls across both recording platforms. The model operates on 60-second spectrograms computed with a 1-s Hamming window and 90% overlap, band-limited to 10–150 Hz. Audio is partitioned into non-overlapping 60-second clips, spectrograms are normalized individually between 0–1, and mapped to an 8-bit integer tensor. Code for running inference or retraining the model is available at https://github.com/m1alksne/WhaleMoanDetector.

Fields of Study & Subjects

Marine animal studies overview Music and Audio Processing Underwater Acoustics Research

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN
Document type
Dataset