Automated sea cage geometry detection from 3D multibeam sonar using machine learning

Reliable interpretation of acoustic observations in aquaculture systems requires accurate knowledge of the spatial structure being monitored. In flexible sea cages, geometric deformation under varying hydrodynamic conditions alters the spatial reference frame used to analyse sensor data and can lead to errors in automated monitoring and analysis. This study presents an automated, data-driven approach for detecting sea cage geometry directly from 3D multibeam sonar observations collected at a commercial salmon farm. The approach combines time-synchronised acoustic measurements with physics-based simulation-derived reference geometries and trains a 3D convolutional neural network to infer cage geometry from sonar data alone. The final implementation was developed in PyTorch using supervised training on matched sonar–geometry pairs. Validation results show strong agreement between predicted and reference cage geometries, with overlap-based similarity values in the range of approximately 0.93–0.94. These results demonstrate that cage geometry can be inferred directly from sonar observations without relying on external deformation sensors during operation. The proposed method provides a non-invasive and scalable approach for real-time cage-scale monitoring and supports downstream applications such as fish behaviour analysis, biomass interpretation, and precision aquaculture decision support.