Using Machine Learning with Scanning Sonar Data and Artificial Targets for Shrimp Biomass Estimation

Isap, Hamzah and Bush, Alex (2025) Using Machine Learning with Scanning Sonar Data and Artificial Targets for Shrimp Biomass Estimation. Masters thesis, Lancaster University.

Text (2025IsapMSc)
2025IsapMSc.pdf - Published Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (1MB)

Abstract

Smart aquaculture is a data-driven approach to optimise operations and is a valuable practice for shrimp farmers to upscale sustainably. Acoustic telemetry is generally regarded as the most reliable form of data acquisition to obtain desired stock information, such as biomass and abundance. Current studies in the field deploy high-grade scientific sonars and their data to train sophisticated models, overlooking the financial viability. In contrast, this study explores the potential for basic-specification single-beam scanning sonars to construct acoustic datasets for model training. We propose separate methods for using machine learning to predict two stock measurements: school density and abundance, using artificial targets in a sample area. To model school density, a monofilament net containing a varied density of standardised uniform air-filled spheres produces echo traces, which an optimised neural network categorises to an overall accuracy of 90.78%. To model shrimp abundance, artificial targets modelling shrimp are presented to capture abundance with active material and orientation variables. We collect averaged echograms of the tank containing a variable abundance of suspended targets. We then deploy a variation of echo-integration where the sum of digital signals for each beam position is processed as features. Optimised Gaussian process regression models are the best-performing models in predicting the number of targets in the tank. Training models on different population ranges found the maximum error around 10%, with the best model demonstrating an MAE of 1.36 (2.7%). Models fit data with an R-squared upwards of 0.98. The proposed methods demonstrate the promising potential of low-cost sonar implementation within the aquaculture industry.