Feature-based cluster selection framework for binary classification on imbalanced acoustic data

Underwater acoustics is an important component of fisheries management. Echograms, generated from acoustic backscatter data collected via echosounders, provide information on the distribution and abundance of marine life. Acoustic target classification (ATC) categorizes backscatters by assigning them into specific groups, such as fish or seabed. Traditionally, ATC requires extensive manual annotation, which is both time-consuming and prone to inconsistency. Automated approaches often rely on labeled data, but obtaining precise annotations remains challenging. This study introduces a method combining self-supervised learning (SSL) and clustering to perform high-resolution analysis of acoustic data. SSL features are extracted from echograms and clustered to classify pixels in a binary task. To address class imbalance, an over-clustering strategy is applied, followed by a greedy cluster selection algorithm guided by labeled data to maximize the F 1 score. Aggregated probabilities for each pixel are calculated based on the selected clusters and thresholded to assign class labels. This feature-based approach is compared to a version using untreated data instead of features, and to a fully supervised method. Results show that the feature-based approach outperforms the raw data-based version and achieves performance close to the supervised model. Self-supervised data representations enable the training of simple yet effective downstream classification models.