Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Semi-automated detection of eagle nests: an application of very high-resolution image data and advanced image analyses to wildlife surveys

Andrew, M.E. and Shephard, J.M. (2017) Semi-automated detection of eagle nests: an application of very high-resolution image data and advanced image analyses to wildlife surveys. Remote Sensing in Ecology and Conservation, 3 (2). pp. 66-80.

PDF - Published Version
Download (1MB) | Preview
Free to read:
*No subscription required


Very high-resolution (VHR) image data, including from unmanned aerial vehicle (UAV) platforms, are increasingly acquired for wildlife surveys. Animals or structures they build (e.g. nests) can be photointerpreted from these images, however, automated detection is required for more efficient surveys. We developed semi-automated analyses to map white-bellied sea eagle (Haliaeetus leucogaster) nests in VHR aerial photographs of the Houtman Abrolhos Islands, Western Australia, an important breeding site for many seabird species. Nest detection is complicated by high environmental heterogeneity at the scale of nests (~1–2 m), the presence of many features that resemble nests and the variability of nest size, shape and context. Finally, the rarity of nests limits the availability of training data. These challenges are not unique to wildlife surveys and we show how they can be overcome by an innovative integration of object-based image analyses (OBIA) and the powerful machine learning one-class classifier Maxent. Maxent classifications using features characterizing object texture, geometry and neighborhood, along with limited object color information, successfully identified over 90% of high quality nests (most weathered and unusually shaped nests were also detected, but at a slightly lower rate) and labeled <2% of objects as candidate nests. Although this overestimates the occurrence of nests, the results can be visually screened to rule out all but the most likely nests in a process that is simpler and more efficient than manual photointerpretation of the full image. Our study shows that semi-automated image analyses for wildlife surveys are achievable. Furthermore, the developed strategies have broad relevance to image processing applications that seek to detect rare features differing only subtly from a heterogeneous background, including remote sensing of archeological remains. We also highlight solutions to maximize the use of imperfect or uncalibrated image data, such as some UAV-based imagery and the growing body of VHR imagery available in Google Earth and other virtual globes.

Item Type: Journal Article
Murdoch Affiliation(s): School of Veterinary and Life Sciences
Publisher: Wiley
Copyright: © 2017 The Authors.
Item Control Page Item Control Page


Downloads per month over past year