Trustworthy Coral Bleaching Classification with Convolutional and Transformer Networks

Automated classification of healthy and bleached coral in underwater photographs can scale reef monitoring, but accuracy reported on public image collections is inflated when near-duplicate frames are shared between training and evaluation partitions. This study introduces a reusable leakage-mitigated, scene-aware benchmarking protocol and applies it to the Coral Reefs Images collection of 10,382 photographs. A single average hash flagged a same-class near-duplicate candidate in the original training partition for 418 of 720 native validation and test images (58.1%), against approximately 29% under a stricter multi-hash criterion. Scene grouping, group-stratified partitioning, and an embedding-based purge yielded 1,710 test images from 112 scene groups. Because the purge relocates individual images, the partitions are leakage-mitigated but not strictly scene-disjoint. Seven convolutional, transformer, and hybrid architectures, each tuned with 5 to 9 completed Optuna trials and trained with three seeds, reached mean test accuracy of 93.0% to 95.6% and ROC-AUC of 0.969 to 0.984. Scene-cluster bootstrap intervals for accuracy were 8 to 12 percentage points wide, and only one of 21 architecture pairs remained significant after joint Holm correction, whereas McNemar tests, which ignore scene clustering, were overconfident. The wide intervals indicate substantial uncertainty and limited evidence for distinguishing small differences, so the data do not support a reliable architecture ranking. A noise and blur augmentation arm, evaluated for four architectures, raised mean accuracy by 4.5 and 4.1 percentage points under strong noise and heavy blur (descriptive, untested) at a clean cost of 0.56 percentage points. Grad-CAM maps achieved lower deletion AUC and higher insertion AUC than smooth random maps for all models, which shows relative map utility and not biological faithfulness. Corner and caption occlusion changed few predictions, but matching dropout during training means this does not establish the absence of logo or caption shortcuts. An independent audit still found neighbors above 0.95 cosine similarity for 10.9% of test images, so the reported accuracies are within-collection estimates that may be optimistic for independently collected reef imagery.