An Integrated Deep-Ensemble Framework for Non-Invasive Fish Growth Prediction in Precision Aquaponics

Abstract

Sustainable aquaponics management in India’s rapidly evolving fisheries sector necessitates the adoption of high-precision, non-invasive growth monitoring technologies. Traditional methodologies for estimating fish length and weight rely on manual sampling or static IoT thresholds, which often result in significant physiological stress to aquatic specimens and poor predictive accuracy due to the non-linear complexity of water quality parameters. This research proposes an intelligent predictive framework centered on a novel hybrid architecture, ConvETR. The methodology integrates a 1D Convolutional Neural Network (CNN) for feature extraction with an Extra Trees Regressor (ETR) to stabilize prediction variance. To validate the system, a high-frequency sensor dataset, capturing variables such as Temperature, pH, Dissolved Oxygen, and Ammonia was utilized to compare ConvETR against baseline models, including Linear Regression (LR), Lasso Regression (Lasso), and Ridge Regression (RR). Experimental results demonstrate that the ConvETR model significantly outperforms traditional approaches, achieving an R2 score of 0.9975 for fish length and 0.9957 for fish weight prediction. Furthermore, the system is deployed within a secure Graphical User Interface (GUI) utilizing a Redis-backed authentication layer to ensure data integrity and role-based access control. By reducing the RMSE for weight prediction from 27.94 (traditional) to 2.97 (proposed), this study provides a robust, scalable solution for "Aquaculture 4.0." The ConvETR framework effectively eliminates handling-related mortality and optimizes resource allocation, offering a pivotal advancement for climate-resilient and automated aquaponic farm management in South India and beyond.