A Temporal–Spatial Deep Learning Framework for Intelligent Fish Growth Modelling Using CNN-Driven Ensemble Regression

Abstract

Aquaculture production has increased over the last decade; however, field observations suggest that nearly 20% of potential yield is lost in each cycle due to inadequate monitoring and imprecise growth estimation. Although overall production continues to rise, the lack of dependable automated prediction systems restricts operational efficiency in fish farming. Conventional practices, which involve manually handling fish to measure length and weight, are time-consuming, inconsistent, and dependent on operator expertise. Moreover, frequent handling disrupts the aquatic environment, induces stress in fish, and negatively influences their growth patterns. These limitations hinder continuous monitoring, particularly in environments with rapidly fluctuating conditions. Practical aquaculture operations, including adaptive feeding strategies, early identification of growth anomalies, and dynamic environmental control, require accurate and real-time predictive capabilities. Existing solutions predominantly rely on Linear Regression (LR), Ridge Regression (RR), and Lasso Regression (Lasso), which are insufficient for modeling complex, non-linear interactions in aquaculture data. To overcome these challenges, this study introduces ConvETR, a hybrid regression framework that integrates a Convolutional Neural Network (CNN) with an Extra Trees Regressor (ETR). The CNN extracts meaningful temporal and spatial features from continuous sensor inputs, capturing behavioral and environmental variations, while the ETR enhances prediction stability by effectively handling noise and non-linearity. This non-intrusive and scalable approach enables accurate real-time predictions, improves decision-making, and supports efficient and sustainable aquaculture management.