Classification of Underwater Crown-of-Thorns Starfish Using Vision Transformer Networks

The population count of crown-of-thorns starfish (CoT) raises concern due to its regional and seasonal outbreaks resulting in unwanted damage to coral reefs, thus monitoring its spread is a vital responsibility to maintain balance in biodiversity. Conventional CoTS detection and monitoring techniques usually require specialized skills, expensive devices and time-consuming. Deep Learning-based image classifier, particularly the ViT-B16, has been developed for CoT detection but it is confirmed limited to inductive biases, lacks multi-scale feature extraction, and has difficulty in handling fine-grained details. With these emerging challenges, this study developed a comparative analysis of different variants of Vision Transformer (ViT) networks for classifying CoT and non-CoT species based on custom underwater images extracted from Kaggle. Nine ViT variants were configured, modelled and tested, namely, SimpleViT, NesT, LeViT, Convolutional ViT (CVT), T2TViT, CrossFormer, XCiT, RegionViT, and NaVit. The developed CVT model outperformed other ViT variants, achieving an accuracy of 82.84 % and F1-Score of 83.92%, making it the most recommendable and suitable choice for environmental monitoring tasks such as CoT classification. Additionally, CVT requires moderate computational resources as compared to other variants of ViTs which are known for significant hardware requirements. Thus, the findings of this study contribute to the ongoing efforts to monitor and mitigate the spread of CoT by correctly classifying CoT species, providing valuable insights for the actual deployment of ViT classifiers in marine pest management.