A Comparative Evaluation of Convolutional and Transformer Architectures for Underwater Coral Reef Classification

Abstract

Coral reefs are among the most biodiverse and most rapidly declining marine ecosystems, and automated classification of reef imagery underpins biodiversity assessment, bleaching surveillance and conservation planning. Seven widely used transfer-learning backbones EfficientNetB0, ResNet50, ResNet101, ResNet152V2, VGG16, VGG19 and Xception are evaluated under a fixed train, validation and test partition, and four modern architectures ResNet50, EfficientNet-B3, ViT-Base and DenseNet121 under stratified 13-fold cross-validation with out-of-fold aggregation; ResNet50 appears in both, allowing the sensitivity of a single architecture to the evaluation protocol to be observed directly. Overall accuracy is a weak guide to practical utility every architecture recovers the morphologically distinctive classes almost perfectly while failing on the same three visually ambiguous classes. Second, those failures are only partially correlated across architectures a class recovered at 0.84 recall by one model is recovered at 0.13 by another indicating substantial headroom for ensemble methods and identifying precisely which classes future work should target.