CNN Architectures for Fish Disease Classification in Aquaculture
Jyotirmoy Basumatary
Department of Electronics and Communication Engineering, Assam University, Silchar, Assam, India.
G. J. Venkata Siva
Department of Electronics and Communication Engineering, Assam University, Silchar, Assam, India.
Raj Kumar Harijan
Department of Electronics and Communication Engineering, Assam University, Silchar, Assam, India.
Abhishek Bhattacharjee
Department of Electronics and Communication Engineering, Assam University, Silchar, Assam, India.
Mariam Jabali Laskar *
Department of Electronics and Communication Engineering, Assam University, Silchar, Assam, India.
*Author to whom correspondence should be addressed.
Abstract
Fish diseases can compromise aquaculture production through reduced growth, deteriorating fish quality and increased mortality. Conventional disease identification often depends on time-consuming manual inspection. This study compared convolutional neural network (CNN) architectures containing two, three and four convolutional layers for classifying five fish disease categories and healthy fish. A dataset of 1,600 images was assembled from existing repositories and selected websites. The images were resized to 224 × 224 pixels and divided into training, validation and testing subsets at an 80:10:10 ratio. The models were trained for 100 epochs using a batch size of 32, with rotation and contrast augmentation applied to the training images. Model performance was assessed through classification accuracy, training and validation curves, and class-level evaluation measures. Across five executions, average test accuracies were 91.13%, 95.73% and 95.27% for the two-, three- and four-layer CNNs, respectively. The three-layer CNN showed comparatively stable training and validation behaviour, whereas the other architectures displayed less stable generalisation or indications of overfitting. Two additional baseline models were evaluated on the same dataset: fine-tuned VGG16 achieved 96.41% accuracy, and a VGG16–Random Forest hybrid achieved 93.56%. Among the three CNN architectures examined, the three-layer model provided the highest average accuracy and a favourable balance between classification performance and architectural depth on the dataset used.
Keywords: Convolutional neural network, fish disease detection, aquaculture, convolutional layer, pooling layer