Robust shrimp disease detection using multi-model convolutional neural networks-based ensemble strategies

Abstract

Viral shrimp diseases pose serious threats to aquaculture production and public health. The lack of effective treatments for these viral infections highlights the urgent need for the development of early and accurate detection methods. Convolutional neural networks (CNNs) have emerged as a promising solution for the non-destructive identification of shrimp diseases. However, individual CNN models may have limitations in accurately classifying these diseases. To address this issue, combining the outputs of multiple CNN models using ensemble learning approaches can be advantageous. In this context, this study aims to classify shrimp diseases using multiple CNN models and ensemble learning strategies. Beta normalization, hard voting, and weighted ensemble learning approaches were employed in the study. The experiments were conducted on a publicly available dataset. In the study, 11 different pre-trained CNN models were used, and their performance was evaluated using 5-fold cross-validation. The results showed that the MobileNet model achieved the highest individual performance, with an average accuracy of 0.919 ± 0.001. This model was followed by DenseNet169, DenseNet121, and DenseNet201 in terms of accuracy rates. The weighted learning strategy (WM-3) using these four models achieved an average accuracy of 0.973 ± 0.004. Additionally, the Gradient-weighted Class Activation Mapping (Grad-CAM) method was used to evaluate the decision-making mechanisms of these models. Statistical evaluations were performed using the Wilcoxon Signed-Rank test and Cohen's d effect size analysis. These findings indicate that utilizing ensemble strategies with a combination of heterogeneous CNN models can significantly improve the accuracy of shrimp disease classification compared to individual CNN models.