Data-Driven Regression Framework for Reliable Fish Growth Prediction in Aquaculture Farms

Abstract

Aquaculture output has grown by over 35% in the past decade, yet field reports indicate that nearly 20% of expected growth is lost each cycle due to insufficient monitoring and inaccurate estimation. Despite steady production increases, the absence of reliable automated prediction systems limits farm efficiency. Traditional methods, which involve manually capturing fish to estimate length and weight, are slow, inconsistent, and highly sensitive to operator skill. Frequent handling also disturbs the pond environment, stressing fish and affecting growth behavior. These challenges make continuous monitoring difficult, especially under rapidly changing environmental conditions. Real-world applications such as adaptive feeding, early detection of growth deviations, and dynamic environmental adjustments require real-time, accurate prediction support. Current systems largely rely on linear regression models, basic regression techniques, or manually engineered features from simple sensors, which struggle to capture the nonlinear patterns inherent in aquaculture. To address this, we propose ConvETR, a regression framework combining a Convolutional Neural Network (CNN) with an Extra Trees Regressor (ETR), designed to process continuous visual sensor streams rather than static images. The CNN extracts meaningful temporal–spatial features, capturing density variations, movement behavior, and subtle water-column cues related to growth, while the ETR stabilizes predictions by handling noise and nonlinear relationships. This integrated, non-intrusive approach improves accuracy, enables real-time decisions, and offers a scalable solution for modern IoT-enabled aquaculture systems requiring reliable fish-growth regression.