Training Convolutional Neural Networks Using An Automated Feedback Loop To Estimate The Population Of Avian Species
Date of Award
Master of Science (MS)
Using automated processes to detect wildlife in uncontrolled outdoor imagery in the field of wildlife ecology is challenging task. This is especially true in imagery provided by an Unmanned Aerial System (UAS), where the relative size of wildlife is small and visually similar to its background. In the UAS imagery collected by the Wildlife@Home project, the data is also extremely unbalanced, with less than 1% of area in the imagery being of wildlife. To tackle these challenges, the Wildlife@Home project has employed citizen scientists and trained experts to go through collected UAS imagery and classify it. Classified data are used as inputs to convolutional neural networks (CNNs) which seek to automatically mark which areas of the imagery contain wildlife. The output of the CNN is then passed to a blob counter which returns a population estimate for the image. A feedback loop was developed to help train the CNNs to better differentiate between the wildlife and the the visually similar background and deal with the disparate amount of wildlife training images versus background training images. When using the feedback loop and citizen scientist provided data, population estimates by the CNN and blob counter are within 3.93% of the manual count by the field biologists. When expert provided data is used the estimates are within 5.24%. This is improved from 150% and 88% error in previous work which did not employ a feedback loop for the citizen science and expert data, respectively. Citizen scientist data worked better than expert data in the current work potentially because a matching algorithm was used on the citizen scientist data but not the expert data.
Bowley, Connor Ryan, "Training Convolutional Neural Networks Using An Automated Feedback Loop To Estimate The Population Of Avian Species" (2017). Theses and Dissertations. 2173.