A Study on Fish Image Classification via Multi-Model Ensemble and Data Augmentation: The Case of Philippine Ports

Efficient classification of fishery resources plays a crucial role in reducing poverty, ensuring food security, and minimizing economic losses in developing countries. However, traditional manual sorting methods are inefficient and costly, leading to a decline in fish freshness and resource wastage. To address this challenge, this study proposes a deep learning framework based on multi-model fusion and data augmentation, using fish image classification at Philippine ports as the application scenario. We trained and evaluated the performance of ResNet34, VGG16, and Vision Transformer (ViT) under three conditions: the full dataset, small-sample subset, and augmented data. Experimental results demonstrate that dataset size has a decisive impact on model performance. On the full dataset (13,304 images), ResNet34 and VGG16 achieved classification accuracies of 98.12 % and 97.61 %, respectively, significantly outperforming ViT (90.97 %). Under the small-sample condition (10 % subset, 1,330 images), ViT's performance dropped sharply to 41.76 %, while ResNet34 maintained robust accuracy at 82.84%. Data augmentation techniques - including geometric transformations, noise injection, and color jitter - improved ViT's accuracy by 8 %, but had minimal effect on ResNet34 (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+0.3 \%$</tex>), indicating ViT's higher dependency on data volume.