Taxonomically Grounded Explainable AI for Automated Identification of Six Coastal Jellyfish Species Using a Convolutional Neural Network

Abstract

Jellyfish blooms are increasing in frequency and economic impact worldwide, yet field identification of species remains a bottleneck for citizen-science monitoring and beach safety, because species differ sharply in ecological role and sting severity but are visually similar to non-specialists. We built a ground-truth taxonomic diagnostic-feature map and an ecological/medical-relevance table for six coastal species (moon, barrel, blue, compass, lion's mane and mauve stinger jellyfish) from the marine-biology literature, independently of any model. A dataset of 900 field and archival images (150 per species) was used to train a logistic-regression baseline and a ResNet50 convolutional neural network (CNN) under identical stratified train/validation/test splits (636/132/132), and Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize the CNN's attention on held-out test images. The pixel-level baseline reached 58.3% test accuracy (macro-F1 = 0.587), while the transfer-learned CNN reached 98.5% accuracy (macro-F1 = 0.985), consistent with accuracies reported for comparable jellyfish-classification CNNs in the literature. Direct inspection of all twenty-four Grad-CAM overlays showed attention consistently anchored on the expected diagnostic feature (tentacle mass, pigmented bell, bell-surface warts, or gonad pattern) for lion's mane, compass, mauve stinger and moon jellyfish, and largely so for barrel jellyfish, but for blue jellyfish the species responsible for both of the CNN's held-out test-set errors — half of the reviewed overlays showed attention drawn to background water texture or a source-image watermark rather than the animal. A biologically grounded explainability protocol, comparing Grad-CAM attention against independently curated taxonomic criteria rather than accuracy alone, is necessary before automated jellyfish identification tools are deployed in citizen-science or public-safety contexts; here it both confirmed the CNN's biological validity for most species and uncovered a concrete, non-morphological failure mode for one, which accuracy alone would have concealed.