SealNet

SealNet performs automated facial recognition to identify individual harbor seals (Phoca vitulina) from photographs for long-term ecological and population monitoring.

Key Features:

• Automated Photo Identification: Detects, aligns, and crops (chips) seal faces from photographs to generate standardized inputs for identification.
• Deep Convolutional Neural Network (CNN): Employs a deep CNN architecture tailored for small datasets, capable of classifying individuals in scenarios such as ~100 seals with five photos per seal.
• High Accuracy: Achieved 88% Rank-1 and 96% Rank-5 closed-set identification accuracy in Casco Bay, Maine (2019–2020) and outperformed PrimNet adapted from primates.
• Validation Dataset: Validated on a dataset of 1,752 images representing 408 individual seals collected from haul-out sites.

Scientific Applications:

• Population Monitoring: Enables non-invasive individual identification to support long-term population size and demographic studies of harbor seals (Phoca vitulina).
• Behavior and Habitat Use: Facilitates longitudinal tracking of individual movements, haul-out site fidelity, and behavioral observations without physical tagging.
• Conservation and Management: Generates individual-level data to inform conservation strategies and management decisions for coastal marine mammal populations.

Methodology:

Images collected at haul-out sites are processed by detecting, aligning, and cropping seal faces and then input to a deep convolutional neural network for closed-set identification; performance was evaluated on 1,752 images of 408 individuals.