Leveraging TensorFlow data generators and U-Net architecture for scalable classification of acoustic fish categories

• Acoustic fish echoes identification and classification. • The use of TensorFlow data generators for scalable training. • A deep learning-based U-Net architecture for pixel-wise classification. • An F1-score of 0.7266 on test dataset. Small pelagic species play a crucial role in Morocco’s fishery sector, contributing significantly to total catches and production. Their contribution is also vital in ensuring food security, employment, and supporting the country’s economy. Accurately estimating their biomass and abundance, as well as monitoring their spatial and temporal variations, are essential for sustainable management. To achieve this, multiple research vessels equipped with echosounders collect acoustic and environmental data along the Moroccan continental shelf each year. Currently, experienced operators manually process and analyze this data to classify fish species and estimate their biomasses, but this approach is time-consuming and subject to individual interpretation. Furthermore, isolating fish species echoes from bottom echoes has always been challenging, as bottom echoes can significantly inflate fish biomass estimates. In this study, we propose a scalable pipeline that combines TensorFlow data generators with a deep learning-based U-Net architecture to classify acoustic categories from echograms. While the initial aim was to classify among all the available acoustic categories, the heavy class imbalance led us to merge them into two broader classes: ‘Probable-Fish’ and ‘Background’, which includes seabed, plankton, zooplankton, and open water. The model achieved promising results, with a classification score performance of an F1-score of 0.7266. The proposed pipeline yields results comparable to previous studies, while requiring less computational overhead, offering improved memory optimization, and enabling scalable training on medium-capacity machines.