Deep‑learning pipeline automates pinniped detection and body‑condition metrics from drone imagery

Arena Tukanova presented a project to automate pinniped abundance, demographic classification and body‑condition assessment using high‑resolution UAV imagery and deep‑learning segmentation. The method builds georeferenced orthomosaics and 3‑D haulout models, extracts 2‑D measurements via segmentation masks, and then reconstructs 3‑D morphology to improve size and condition estimates. On harbor seal datasets from British Columbia, overall detection probability was about 91% for RGB sensors and mean measurement error for automated measurements was roughly 6 cm compared with manual measurements.

Tukanova described thermal sensors as showing systematic bias (mean −20 cm in tests) and identified substrate and drone platform as factors affecting measurement error. The team plans to finalize 3‑D reconstructions, reduce terrain artifacts on rocky haulouts, validate across species (harbor, stellar sea lions, northern fur seals) and incorporate species identification for mixed haulouts. "We believe that near‑realtime abundance and haulout composition data will be a useful tool for rapid assessments and monitoring," Tukanova said.