Application of AMIS-optimized vision transformer in identifying disease in Nile Tilapia

Abstract

Efficient health monitoring in Nile tilapia aquaculture is critical due to the substantial economic losses from diseases, underlining the necessity for innovative monitoring solutions. This study introduces an advanced, automated health monitoring system known as the “Automated System for Identifying Disease in Nile Tilapia (AS-ID-NT),” which incorporates a heterogeneous ensemble deep learning model using the Artificial Multiple Intelligence System (AMIS) as the decision fusion strategy (HE-DLM-AMIS). This system enhances the accuracy and efficiency of disease detection in Nile tilapia. The research utilized two specially curated video datasets, NT-1 and NT-2, each consisting of short videos lasting between 3–10 s, showcasing various behaviors of Nile tilapia in controlled environments. These datasets were critical for training and validating the ensemble model. Comparative analysis reveals that the HE-DLM-AMIS embedded in AS-ID-NT achieves superior performance, with an accuracy of 92.48% in detecting health issues in tilapia. This system outperforms both single model configurations, such as the 3D Convolutional Neural Network and Vision Transformer (ViT-large), which recorded accuracies of 84.64% and 85.7% respectively, and homogeneous ensemble models like ViT-large-Ho and ConvLSTM-Ho, which achieved accuracies of 88.49% and 86.84% respectively. AS-ID-NT provides a non-invasive, continuous, and automated solution for timely intervention, successfully identifying both healthy and unhealthy (infected and environmentally stressed) fish. This system not only demonstrates the potential of advanced AI and machine learning techniques in enhancing aquaculture management but also promotes sustainable practices and food security by maintaining healthier fish populations and supporting the economic viability of tilapia farms.