Automated UAV and Satellite Image Analysis For Wildlife Monitoring.

Bowler, Ellen (2023) Automated UAV and Satellite Image Analysis For Wildlife Monitoring. Doctoral thesis, University of East Anglia.

Preview

PDF Download (96MB) | Preview

Abstract

Very high resolution satellites and unmanned aerial vehicles (UAVs) are revolutionising our ability to monitor wildlife, especially species in remote and inaccessible regions. However, given the rapid increase in data acquisition, computer-automated approaches are urgently needed to count wildlife in the resultant imagery. In this thesis, we investigate the application of convolutional neural networks (CNNs) to the task of detecting vulnerable seabird populations in satellite and UAV imagery. In our first application we train a U-Net CNN to detect wandering albatrosses in 31-cm resolution WorldView-3 satellite imagery. We compare results across four different island colonies using a leave-one-island-out cross validation, achieving a mean average precision (mAP) score of 0.669. By collecting new data on inter-observer variation in albatross counts, we show that our U-Net results fall within the range of human accuracy for two islands, with misclassifications at other sites being simple to filter manually. In our second application we detect Abbott’s boobies nesting in forest canopy, using UAV Structure from Motion (SfM) imagery. We focus on overcoming occlusion from branches by implementing a multi-view detection method. We first train a Faster R-CNN model to detect Abbott’s booby nest sites (mAP=0.518) and guano (mAP=0.472) in the 2D UAV images. We then project Faster R-CNN detections onto the 3D SfM model, cluster multi-view detections of the same objects using DBSCAN, and use cluster features to classify proposals into true and false positives (comparing logistic regression, support vector machine, and multilayer perceptron models). Our best-performing multi-view model successfully detects nest sites (mAP=0.604) and guano (mAP=0.574), and can be incorporated with expert review to greatly expedite analysis time. Both methods have immediate real-world application
for future surveys of the target species, allowing for more frequent, expansive, and lower-cost monitoring, vital for safeguarding populations in the long-term.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Chris White
Date Deposited:	28 Mar 2023 12:46
Last Modified:	28 Mar 2023 12:46
URI:	https://ueaeprints.uea.ac.uk/id/eprint/91697
DOI:

Downloads

Actions (login required)

View Item