Context-Dependent Deep Learning for Bird Recognition in Drone Survey Imagery
July 2020 - December 2021
- University of Maine
Surveys of colonial nesting birds typically are conducted with a combination of ground-based surveys and by counting birds recorded on aerial imagery captured from planes and unoccupied aerial vehicles (UAVs or drones). Manually interpreting this imagery is time-intensive. There are many emerging applications of artificial intelligence to increase efficiency and accuracy of interpreting digital imagery, but these also can be time-intensive and require substantial computer resources. We are exploring effects of context of the focal object in training the deep learning system, to decrease network training time through improved computational complexity (by reducing dimensionality, feature number, and image size), while improving accuracy. We expect this approach will facilitate leveraging knowledge from human experts to reduce the training time and augment the performance of the image analysis system, which is generally difficult to do with deep learning systems.