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

Area of Science:

• Aquaculture technology
• Marine engineering
• Artificial intelligence in agriculture

Background:

• Greenshell mussel farming is crucial to New Zealand's economy.
• Offshore expansion presents buoyancy challenges for mussel farms.
• Current monitoring methods are insufficient for large-scale operations.

Purpose of the Study:

• To introduce BuoyancyNet, a deep learning model for predicting mussel farm float buoyancy.
• To develop a scalable and automated solution for buoyancy management.
• To improve the efficiency and reduce losses in offshore mussel aquaculture.

Main Methods:

• Utilized a dataset of over 36,000 labelled float images from mussel farms.
• Developed a novel deep learning approach using a vision transformer enhanced with 1D convolutional layers.
• Trained the model to learn spatial relationships between consecutive floats.

Main Results:

• BuoyancyNet achieved a 3.5% improvement in multi-class classification accuracy compared to baseline models.
• The model demonstrated robust performance in diverse environmental conditions, including low light and occlusions.
• Successfully addressed challenges in large-scale buoyancy monitoring for mussel farms.

Conclusions:

• BuoyancyNet offers a promising, efficient, and scalable solution for monitoring mussel farm buoyancy.
• The deep learning approach can significantly reduce product losses due to buoyancy issues.
• This technology supports the sustainable expansion of New Zealand's offshore mussel aquaculture industry.