Enhancing multi-frequency acoustic signal extraction of Antarctic krill Euphausia superba using U-Net convolutional neural network

Antarctic krill Euphausia superba is a key species in the Southern Ocean ecosystem and represents the target species for the largest fishery in the region in terms of catch. Traditional acoustic monitoring of krill is challenged by complex noise removal and semi-automated processing, which can introduce biases in krill biomass estimates. In this study, we used a deep learning approach based on the U-Net convolutional neural network to extract krill signals from Simrad EK60 scientific echosounder data using various combinations of acoustic frequencies. Experimental results demonstrated that this end-to-end deep learning method can effectively recognize and segment krill swarms in acoustic data. The model using triple frequencies (38, 70, and 120 kHz) performed best, achieving an accuracy of 99.18%, an F1-score of 97.91%, and an intersection-over-union of 95.90% for krill. The combination of different frequency acoustic data significantly impacted the accuracy of krill extraction, with the 120 kHz frequency yielding the most substantial influence on the extraction results. Compared to traditional methods, this approach is more automated, does not rely on high-performance acoustic equipment, and maintains high recognition accuracy in complex marine environments. This study highlights the robust extraction capabilities of deep learning methods for krill acoustic data features across different frequencies, providing new technical support for ecological monitoring of krill. Additionally, it serves as a promising reference for extracting the signals of krill swarms and potentially other marine organisms using deep learning approaches.