Machine Learning-Based Classification of On-Bottom Oyster Farms Using a Mechanical Scanning Sonar

Oysters play a vital ecological role and serve as natural filters and ecosystem engineers. Sensing technology advances, particularly in environmental monitoring, can offer a promising solution for monitoring oyster reefs and supporting sustainable oyster farming. The ability to classify the seabed and oyster crop inventory would enable precision oyster farming, improve production, and contribute to the health of marine ecosystems. As a first attempt towards precision oyster farming, we propose a novel monitoring approach that utilizes a low-cost, single beam mechanical scanning sonar device empowered with machine-learning algorithms for accurate classification of seabed and oyster inventory conditions. Lab experiments were conducted with controlled seabed/substrate types, density of live oysters, and different oyster sizes. Two different learning-based methods (Convolution Neural Network (CNN) and Dense Neural Network (DNN)) are proposed for classification and their performance with respect to different frequencies is investigated. The lab experimental results demonstrate the proposed CNN algorithm performs the best at 500 kHz, achieving high total accuracies (number of successfully classified entries across all classes) of 86.5%, 87.5%, and 93.8% for classifying substrate types, oyster density, and oyster sizes, respectively. Moreover, overlaying images of three different frequencies improves the accuracy of the density classification by 9.5%. Furthermore, using the DNN with vector inputs that have increased dataset size allows for further improvements of 3.6%, 5.1%, and 7.3% when compared to the CNN. We also validate our proposed approach using data collected from an active oyster farm, demonstrating its potential for real-world applications.