Unsupervised Selective Labeling for Semi‐Supervised Aerial Imagery Recognition
Ecology and Evolution · 2026-11 · 16 (11)
DOI: 10.1002/ece3.73943 ↗Kurz gefasst
Fachartikel in der Zeitschrift Ecology and Evolution, online erschienen am 18.11.2026. Verfasst von Zilin Wang und 9 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: U.S. Geological Survey.
Abstract (Originalsprache)
Traditional aerial wildlife monitoring surveys rely on rapid visual identification and counting by expert biologists from moving aircraft. Modern camera systems mounted on survey aircraft are now being deployed to collect large datasets of imagery that provide a persistent, reusable snapshot of wildlife populations. Manual analysis of these extensive datasets can overwhelm the staffing of natural resource agencies, especially for large‐scale surveys. Deep learning systems are being investigated as a potential solution to automate analysis, but significant human effort is still required to curate a labeled dataset for model training. To address this challenge, we propose Semi‐Supervised Aerial Imagery Recognition (referred to as SSAR), a framework that minimizes annotation effort while maintaining classification performance. The key of SSAR is to select the best subset of images for annotation while leveraging the hidden information from the remaining unlabeled data. Concretely, SSAR integrates three components: (1) unsupervised feature learning tailored to aerial imagery to capture ecologically meaningful visual patterns, (2) unsupervised selective labeling that prioritizes representative and diverse images for annotation, and (3) semi‐supervised classification that incorporates the remaining unlabeled images during training. We evaluate SSAR on wildlife recognition case studies using object‐centric image crops from high‐resolution aerial imagery collected over the United States Atlantic Outer Continental Shelf (OCS) and the Great Lakes region, comprising 57,265 images across 11 coarse categories. Using only 6.4% labeled data, SSAR matches the fully supervised macro accuracy of 85.0% achieved with 100% labeled data and further surpasses it by reaching 86.8% with 12.8% labeled data. These results demonstrate that SSAR enables accurate and scalable aerial wildlife recognition with an order‐of‐magnitude reduction in manual annotation.
Förderung
- U.S. Geological Survey
Zitieren
APA
Wang, Z., Yu, S. X., Landolt, K. L., Koneff, M. D., Pickens, B. A., Schuster, S., Fara, L. J., Murphy, A. C., Dieck, J., & White, T. P. (2026). Unsupervised Selective Labeling for Semi‐Supervised Aerial Imagery Recognition. Ecology and Evolution, 16(11). https://doi.org/10.1002/ece3.73943
Vancouver
Wang Z, Yu SX, Landolt KL, Koneff MD, Pickens BA, Schuster S, et al. Unsupervised Selective Labeling for Semi‐Supervised Aerial Imagery Recognition. Ecology and Evolution. 2026;16(11). doi:10.1002/ece3.73943
BibTeX
@article{Wang2026,
author = {Wang, Zilin and Yu, Stella X. and Landolt, Kyle L. and Koneff, Mark D. and Pickens, Bradley A. and Schuster, Sierra and Fara, Luke J. and Murphy, Aaron C. and Dieck, Jennifer and White, Timothy P.},
title = {Unsupervised Selective Labeling for Semi‐Supervised Aerial Imagery Recognition},
journal = {Ecology and Evolution},
year = {2026},
volume = {16},
number = {11},
doi = {10.1002/ece3.73943}
}
EUVIMED