New Publication: Deep learning–based frameworks

Manuel Vieira
Sep 2
2 min read

Our latest paper "Deep Learning–Based Frameworks for the Detection and Classification of Soniferous Fish" was published in the Journal of the Acoustical Society of America (Read the article here).

About the Study Passive acoustic monitoring (PAM) is increasingly recognized as a powerful tool for studying fish populations in their natural environments. This study compared two deep learning–based frameworks for detecting, classifying, and counting sounds produced by soniferous fish in the Tagus Estuary, Portugal:

SegClas: A multi-label segmentation-based classification system combining convolutional neural networks and long short-term memory networks.
ObjDet: An object detection approach using a YOLO-based model for species classification and vocalization counting.

Key Findings:

Both methods achieved >96% accuracy and F1 scores above 87%, even under noisy conditions.
ObjDet offered slightly higher performance and more precise vocalization counts, while SegClas was computationally faster and efficient.

Looking Ahead As highlighted in the paper, ongoing refinement of these frameworks — including adaptive thresholding and hybrid modeling strategies — will help minimize misclassifications and extend their use in diverse ecological contexts. Integrated into ecosystem management, deep learning–based PAM could eventually deliver real-time insights from field stations, enabling earlier detection of ecological threats and contributing to the protection and sustainable management of marine environments.

These methods are simple to adjust and re-train, and with proper validation, they can be utilized in other contexts. In our lab, we aim to apply these techniques to study the Tagus estuary and other soundscapes.

Ref: Huang, Z., Ochs, D., Amorim, M., Fonseca, P. J., Goel, M., Nunes, N. J., Vieira, M. & Lopes, M. (2025). Deep learning–based frameworks for the detection and classification of soniferous fish. The Journal of the Acoustical Society of America, 158(2), 1060-1071. https://doi.org/10.1121/10.0038800