Unraveling Acoustic Signal Patterns in Fisheries Through DINO-Based Self-Supervised Learning

Acoustic surveys provide important data for fisheries management. During the surveys, ship-mounted echo sounders send acoustic signals into the water and measure the strength of the reflection, so-called backscatter. Acoustic data, the collection of those backscatters, are annotated by manual processes which are resource-intensive and time-consuming. In addition, during these annotation processes, only target fish species are typically labeled, leaving other structures such as zooplankton layers unlabeled. Recently introduced supervised methods help with the annotation process. Those methods, however, still rely on the annotations which are incomplete and hard to get. Our main objective in this research is to develop a robust and annotation-free methodology to advance the analysis of fisheries acoustic data. We use the self-supervised method based on DINO (Self-Distillation with No Labels) which was introduced for computer vision tasks. The DINO model learns image representations from global and local views of the image and creates the corresponding feature embeddings that can be used for subsequent tasks. This is achieved using two networks, a teacher and a student, with the same network architecture. They generate normalized K-dimensional features using temperature SoftMax and are compared through cross-entropy loss. Gradients flow only through the student network, and the teacher’s parameters are updated using an exponential moving average of the student’s parameters. We have systematically adapted the DINO model to handle acoustic data by substituting images with fixed-sized acoustic patches. Additionally, we have refined data augmentation techniques to align with the unique characteristics of acoustic data. The resulting embeddings were then analyzed using a basic k-NN algorithm, utilizing pre-existing labels. The classification algorithm resulted in a high precision of patch classification. Additionally, when clustering the embeddings corresponding to the unlabeled portion of the acoustic data, distinct structures within the data were observed to form cohesive clusters. This implies that the self-supervised method can identify unique energy patterns in different regions of the acoustic data and could be used to assist with acoustic data exploration and annotations. Our findings highlight the importance of emerging self-supervised techniques in the signal processing domain, specifically for acoustic data in fisheries research. This methodology fills the research gap of self-supervised learning application in the fisheries acoustics domain from a mathematical signal and image processing viewpoint.