SealNet: A fully-automated pack-ice seal detection pipeline for sub-meter satellite imagery

• Published in: Remote sensing of environment 2020-03, Vol.239, p.111617, Article 111617
• Main Authors: Gonçalves, B.C., Spitzbart, B., Lynch, H.J.
• Format: Article
• Language: English
• Subjects: Algorithms; APIS; Artificial neural networks; CCAMLR; Centroids; Classification; Climate change; Counting; Crabeater seal; Deep learning; Environmental changes; Euphausia superba; Fishing; Image detection; Image segmentation; Krill; Learning algorithms; Leptonychotes weddellii; Lobodon carcinophaga; Machine learning; Marine ecosystems; Neural networks; Object detection; Object recognition; Oceans; Phocidae; Pipelines; Polar environments; Predators; Satellite imagery; Sea ice; Seals; Seals (animals); Segmentation; Very high resolution; Weddell seal
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2019.111617

Antarctic pack-ice seals, a group of four species of true seals (Phocidae), play a pivotal role in the Southern Ocean foodweb as wide-ranging predators of Antarctic krill (Euphausia superba). Due to their circumpolar distribution and the remoteness and vastness of their habitat, little is known about their population sizes. Estimating pack-ice seal population sizes and trends is key to understanding how the Southern Ocean ecosystem will react to threats such as climate change driven sea ice loss and krill fishing. We present a functional pack-ice seal detection pipeline using Worldview-3 imagery and a Convolutional Neural Network that counts and locates seal centroids. We propose a new CNN architecture that detects objects by combining semantic segmentation heatmaps with binary classification and counting by regression. Our pipeline locates over 30% of seals, when compared to consensus counts from human experts, and reduces the time required for seal detection by 95% (assuming just a single GPU). While larger training sets and continued algorithm development will no doubt improve classification accuracy, our pipeline, which can be easily adapted for other large-bodied animals visible in sub-meter satellite imagery, demonstrates the potential for machine learning to vastly expand our capacity for regular pack-ice seal surveys and, in doing so, will contribute to ongoing international efforts to monitor pack-ice seals. •First automated system for surveying seals using satellite imagery•Finds 30% of seals while only generating less than 2 false-positives per correct detection•Over 10× faster than an experienced human observer using a single GPU