Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WhaleVAD-BPN: Improving Baleen Whale Call Detection with Boundary Proposal Networks and Post-processing Optimisation

Christiaan M. Geldenhuys, Günther Tonitz, Thomas R. Niesler

TL;DR

This paper tackles false positives and minority-class detection in baleen whale call SED by introducing WhaleVAD-BPN, a boundary proposal network that gates backbone outputs using intermediate features to refine temporal localization. It couples this architectural augmentation with two principled post-processing hyperparameter search strategies (forward and backward) to optimize frame- and event-level decisions, yielding substantial gains over a baseline system. On the BioDCASE Task 2 data, the approach achieves a cross-validated development F1-score of 0.475, representing notable improvements in precision and minority-class detection while preserving recall. The work demonstrates the practical impact of integrating a lightweight boundary proposal module with systematic hyperparameter optimisation for robust, data-efficient baleen whale call detection in challenging acoustic environments.

Abstract

While recent sound event detection (SED) systems can identify baleen whale calls in marine audio, challenges related to false positive and minority-class detection persist. We propose the boundary proposal network (BPN), which extends an existing lightweight SED system. The BPN is inspired by work in image object detection and aims to reduce the number of false positive detections. It achieves this by using intermediate latent representations computed within the backbone classification model to gate the final output. When added to an existing SED system, the BPN achieves a 16.8 % absolute increase in precision, as well as 21.3 % and 9.4 % improvements in the F1-score for minority-class d-calls and bp-calls, respectively. We further consider two approaches to the selection of post-processing hyperparameters: a forward-search and a backward-search. By separately optimising event-level and frame-level hyperparameters, these two approaches lead to considerable performance improvements over parameters selected using empirical methods. The complete WhaleVAD-BPN system achieves a cross-validated development F1-score of 0.475, which is a 9.8 % absolute improvement over the baseline.

WhaleVAD-BPN: Improving Baleen Whale Call Detection with Boundary Proposal Networks and Post-processing Optimisation

TL;DR

This paper tackles false positives and minority-class detection in baleen whale call SED by introducing WhaleVAD-BPN, a boundary proposal network that gates backbone outputs using intermediate features to refine temporal localization. It couples this architectural augmentation with two principled post-processing hyperparameter search strategies (forward and backward) to optimize frame- and event-level decisions, yielding substantial gains over a baseline system. On the BioDCASE Task 2 data, the approach achieves a cross-validated development F1-score of 0.475, representing notable improvements in precision and minority-class detection while preserving recall. The work demonstrates the practical impact of integrating a lightweight boundary proposal module with systematic hyperparameter optimisation for robust, data-efficient baleen whale call detection in challenging acoustic environments.

Abstract

While recent sound event detection (SED) systems can identify baleen whale calls in marine audio, challenges related to false positive and minority-class detection persist. We propose the boundary proposal network (BPN), which extends an existing lightweight SED system. The BPN is inspired by work in image object detection and aims to reduce the number of false positive detections. It achieves this by using intermediate latent representations computed within the backbone classification model to gate the final output. When added to an existing SED system, the BPN achieves a 16.8 % absolute increase in precision, as well as 21.3 % and 9.4 % improvements in the F1-score for minority-class d-calls and bp-calls, respectively. We further consider two approaches to the selection of post-processing hyperparameters: a forward-search and a backward-search. By separately optimising event-level and frame-level hyperparameters, these two approaches lead to considerable performance improvements over parameters selected using empirical methods. The complete WhaleVAD-BPN system achieves a cross-validated development F1-score of 0.475, which is a 9.8 % absolute improvement over the baseline.
Paper Structure (25 sections, 1 equation, 4 figures, 7 tables)

This paper contains 25 sections, 1 equation, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Baleen whale call activity detection
  4. Proposal networks
  5. Post-processing
  6. Frame-level techniques
  7. Event-level techniques
  8. Literature Review
  9. Detection and classification of baleen whale calls
  10. Post-processing in bioacoustic SED
  11. Data
  12. Experimental Structure
  13. Architecture modifications
  14. Boundary proposal network
  15. Intermediate projection head
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Spectrogram representations of exemplar baleen whale call types JeanLabadye2025BioDCASE2025Task2DevDataset. Bounding boxes indicate presence of particular call type, provided by domain expert annotations in dataset. Figures (a) – (d) correspond to blue whale vocalisations, and (e) – (g) correspond to fin whale vocalisations.
  • Figure 2: Illustration of the system we propose for improved whale call boundary detection. The system is divided into four sections: pre-processing, feature extraction, classification, and boundary proposal. The output tensor dimensionality is shown below each module in the system.
  • Figure 3: Illustration of the post-processing applied to the call probabilities produced by a per-frame sed model. (\ref{['fig:post-processing:overview']}) An overview of the steps employed during both frame-level (purple) and event-level (green) post-processing. (\ref{['fig:post-processing:forward']} and \ref{['fig:post-processing:backward']}) The forward- and backward-search approaches to the selection of post-processing hyperparameters, respectively. Both optimisation methods consist of two stages, where some parts of the network remain fixed (opaque dashes) while hyperparameters for the remainder are searched to maximise the F1-score over all development folds.
  • Figure 4: Precision-recall curves for each call type and for each of the three site-year development sets. Both the baseline [ (dashed)] and the proposed [ (solid)] models are shown.