Speaker
Description
Analysis of passive acoustic monitoring (PAM) data is a cost-effective method to study whale presence. The resulting recordings have a strong imbalance between short call events and long periods of background noise, making automated whale call detection challenging. Also, the annotation of the whale calls is often imprecise, which complicates model training. State-of-the-art methods convert audio clips to spectrograms and use computer vision models, such as You Only Look Once (YOLO), for the detection and classification of whale calls.
We base our research on the BioDCASE[1] challenge dataset and reference method. The dataset contains around 188,000 spectrograms generated from 1,880 hours of acoustic recordings, containing blue and fin whale calls. The method uses YOLO11 and has a reported F1-score of 0.44 (Recall = 0.32, Precision = 0.62), which is not sufficient for operational use. Therefore, we evaluate different ways to improve the performance of detection models: edge label filtering, which removes the labels from the spectrogram edges; bounding box inflation, which enlarges the annotations along the time and frequency axes; additional reannotation of the data; and testing different YOLO architectures. For this, we reannotated 1,000 spectrograms randomly sampled from the dataset and fine-tuned the model by continuing to train the last five layers of the baseline model. We compare this to fine-tuning on the original labels and YOLO11s with other YOLO models. Instead of using the original BioDCASE data split, which assigns entire site-year subsets to either the training or the test set, we split each subset separately into both to avoid a batch effect in the experiments.
Although fine-tuning and edge label filtering have shown promising preliminary results, the more accurate experiments are still ongoing, and the results will be presented at the Symposium.
[1] BioDCASE Challenge 2026 Task 2: https://biodcase.github.io/challenge2026/task2