Improving Aquaculture Systems using AI: Employing predictive models for Biomass Estimation on Sonar Images.

Accurate biomass estimation is a major concern in the aquaculture industry due to its role in the efficient operations of fish farms. In this paper, we propose the application of machine learning and deep learning techniques on sound navigation and ranging (Sonar) readings of tanks to predict fish biomass under both clear and murky water conditions. While previous works have proposed similar approaches, they generally face two operational challenges. First, typical setups consider RGB or infrared cameras, which are strongly influenced by water conditions and limit their application, for RGB cameras the light penetration is severely affected by water turbidity, while infrared is strongly absorbed by water. Second, modern fish farming installations such as recirculating aquaculture systems (RAS) operate high-density or super high-density fish tanks, which introduce additional challenges such as noise in sensor readings and occlusions. Our method addresses these issues by (i) leveraging Sonar technology which is less susceptible to variations in water conditions and performs well under both clear and murky water(turbid water); and (ii) designing a custom loss function to reduce the effect of noise which can result in overestimation, and occlusions which can lead to the underestimation in the prediction of fish biomass. We achieve an overall root mean squared error (RMSE) of around 5 for both clear and murky water using both machine learning and deep learning approaches, which is a reasonable value for our dataset. The custom loss function with additional penalties and constraints improves the RMSE and R-2 performance over our preliminary results. The assessment was performed on data collected in an actual operational environment, comprising minimally configured RAS tanks at Blue Aqua International, which is an aquaculture technology provider and a fish farm that intends to develop and commercialize a product for automated biomass estimation.