Classification of Fish Species in South Sumatra Based on Underwater Imagery Using ResNet-50

Abstract

Fish species classification in underwater ecosystems posed a significant challenge, particularly due to poor lighting that affected the quality of underwater images and decreased the accuracy of species identification. This study aimed to improve the accuracy of fish species classification in South Sumatra based on underwater images by utilizing the Super-Resolution Generative Adversarial Network (SRGAN) to enhance image quality and ResNet-50 for species classification. The research employed a Dell XPS 13 9310 device with an Intel Core i7 processor and 16GB of RAM for model training. Fish image data were collected from Google Images and YouTube according to predefined fish species, including Oreochromis mossambicus (Mujair), Oreochromis niloticus (Nila), Johnius trachycephalus (Gulamah), Eleutheronema tetradactylum (Senangin), and Chanos chanos (Bandeng). The data was divided into 70% for training, 15% for validation, and 15% for testing. The experimental results showed that the developed model achieved a training accuracy of 94.10%, validation accuracy of 88.25%, and testing accuracy of 84.68%. This research contributed to the field of underwater image classification and can be applied to conservation and monitoring of fish species in aquatic ecosystems.