<![CDATA[Sustainability of marine resources and management of aquatic ecosystems depend on accurate fish classification. CNNs have proven successful in image classification tasks; however, they often face the problem of limited data variation. The purpose of this study was to examine how data augmentation affects the training accuracy of CNN models for fish species classification. Two scenarios were studied: the first scenario involved training without data augmentation, and the second scenario involved training with data augmentation. In both scenarios, a custom CNN architecture for ten epochs was used. Experimental results showed that using data augmentation with the configuration used actually caused the model performance to deteriorate. Loss values on both datasets increased, with training accuracy dropping from 76.08% to 63.81%, and validation accuracy also dropping from 91.13% to 84.55%. Overly aggressive augmentation parameters or insufficient training time for the introduced data variation could have caused this decline. Interestingly, validation accuracy was consistently higher than training accuracy in both situations, indicating that certain datasets have specific features. This study emphasizes the importance of carefully optimizing augmentation parameters and training duration to maximize the benefits of data augmentation in image classification.]]>
Copyrights © 2025
