A confidence-aware contrastive deep learning network for coral reef fish sound classification

Vocalizations of coral reef fishes provide critical insights into population dynamics, community structure, and ecological processes, making passive acoustic monitoring an essential tool for coral reef conservation and management. However, automatic recognition of coral reef fish sounds remains highly challenging due to strong background noise, pronounced intra-class variability, and high inter-class similarity. To address these challenges, a contrastive reef-acoustic learning network (CoRAL-Net) is proposed for coral reef fish sound classification, in which a robust spectro-temporal feature extractor based on a convolutional neural network-multi-head self-attention (CNN-MSA) module is designed to process the time-frequency spectrograms. Furthermore, to improve triplet construction in contrastive learning, a confidence-aware sampling strategy is proposed to adaptively form triplets based on attention-derived uncertainty and embedding strength, thereby enhancing the representational capability of the CNN-MSA extractor through informative and class-balanced sample selection. This study validates CoRAL-Net on a predominantly field-collected dataset from coral reefs in Hainan Province, China, which contains three dominant types of fish sounds and natural noise recorded under challenging acoustic environments. Experimental results demonstrate that CoRAL-Net significantly outperforms the baseline methods in overall classification performance, highlighting the potential of this confidence-aware strategy to advance passive acoustic biodiversity monitoring in coral reef ecosystems.