BuoyancyNet: a deep learning approach for assessing float buoyancy in mussel aquaculture.

Abstract

The aquaculture sector is vital to Aotearoa New Zealand's (NZ) economy, with greenshell mussel cultivation playing a leading role. As the industry expands into more exposed offshore environments, maintaining proper buoyancy in mussel farm structures becomes increasingly challenging. Buoyancy issues can result in significant product losses through sinking or mussel detachment, creating a critical need for scalable, automated monitoring solutions. This paper introduces BuoyancyNet, a novel deep learning-based method for predicting float buoyancy in mussel aquaculture. Using a labelled dataset of over 36,000 float images from mussel farms, the proposed approach leverages a vision transformer enhanced with one-dimensional convolutional layers to learn spatial relationships between consecutive floats along a float line. BuoyancyNet achieves a 3.5% improvement in multi-class classification accuracy over baseline models and demonstrates robust performance under diverse environmental conditions, including poor lighting and occlusions. By addressing challenges in large-scale buoyancy management, this approach provides a promising step toward more efficient and scalable monitoring solutions for NZ's mussel farming industry.