Deep learning-based frameworks for the detection and classification of soniferous fish.

Area of Science:

• Marine biology
• Bioacoustics
• Artificial intelligence in ecology

Background:

• Passive acoustic monitoring (PAM) is a key tool for fish population assessment in natural environments.
• Soniferous fish vocalizations can overlap, creating challenges for accurate species identification and population counts.
• Deep learning offers advanced computational approaches for analyzing complex acoustic data.

Purpose of the Study:

• To compare the effectiveness of two deep learning frameworks for fish sound detection and classification.
• To evaluate performance metrics including accuracy, F1 scores, and inference speed.
• To assess the suitability of different deep learning approaches for fish population assessment using PAM.

Main Methods:

• Implemented a multi-label segmentation-based classification system (SegClas) using CNNs and LSTMs.
• Utilized an object detection (ObjDet) framework based on YOLO for sound event detection, classification, and counting.
• Tested both frameworks on vocalizations from Lusitanian toadfish, meagre, and weakfish in the Tagus Estuary.

Main Results:

• Both SegClas and ObjDet achieved high accuracy (>96%) and F1 scores (>87%) for species-level sound identification.
• ObjDet demonstrated slightly superior classification performance (F1 up to 92%) and enabled precise counting via bounding-box annotations.
• SegClas offered faster inference times and utilized segment-level labels, presenting a computationally efficient alternative.

Conclusions:

• Deep learning-based PAM frameworks are highly effective for identifying fish vocalizations, even with overlapping patterns and noise.
• Both ObjDet and SegClas offer distinct advantages for fish population assessment, catering to different ecological and operational requirements.
• These advanced analytical tools hold significant potential for enhancing fish population assessments through passive acoustic monitoring.