Species Classification
Upload a whale photograph or underwater audio recording. Our trained models will identify the species and show you similar look-alikes to double-check the result.
Photo Classification
EfficientNet-B4 · 8 species · Happywhale-trained
Audio Classification
XGBoost / CNN · 8 species · 4s segments
About the Photo Model
Architecture: EfficientNet-B4 fine-tuned from ImageNet weights. 380×380 input, differential learning rates (1e-4 head, 1e-5 backbone), cosine annealing scheduler.
Training data: ~20K filtered images from the Happywhale Kaggle dataset across 7 target species + "other cetacean" rejection class.
Species: Right whale, humpback, fin, blue, minke, sei, killer whale, other cetacean.
Visual features used: Fluke patterns, dorsal fin shape, callosities, jaw coloring, saddle patches, flipper bands — learned automatically from mixed body views.
About the Audio Model
Architecture: XGBoost on 64 acoustic features (97.9% accuracy) or CNN (ResNet18) on mel spectrograms (99.3% accuracy). Audio is segmented into 4-second windows with 2-second hop.
Features: 20 MFCCs (mean + std), spectral centroid/bandwidth/rolloff/ flatness, spectral contrast (7 bands), ZCR, RMS energy, dominant frequency, temporal envelope statistics.
Training data: 452 audio files from Watkins Marine Mammal Sound Database + 3 Zenodo datasets → 10,185 segments after segmentation.
Species: Right whale, humpback, fin, blue, sperm, minke, sei, killer whale.