University of North Dakota
Using Citizen Scientists To Inform Machine Learning Algorithms To Automate The Detection Of Species In Ecological Imagery
Abstract
dc:description.abstractModern data collection techniques used by ecologists has created a deluge of data that is becoming increasingly difficult to store, filter, and analyze in an efficient and timely manner. In just two summers, over 65,000 unmanned aerial system (UAS) images were collected, comprising the several terabytes (TB) of data that was reviewed by citizen scientists to generate inputs for machine learning algorithms. Uncontrolled conditions and the small size of target species relative to the background further increase the difficulty of manually cataloging the images. To assist with locating and identifying snow geese in the UAS images, a citizen science web portal was created as part of Wildlife@Home. It is demonstrated that aggregate citizen scientist observations are similar in quality to observations made by trained experts and can be used to train convolutional neural networks (CNN) to automate the detection of species in the imagery. Using a dataset comprising of the aggregate observations produces consistently better results than datasets consisting of observations from a single altitude, indicating that more numerous but slightly variable observations is preferable to more consistent but less numerous observations. The framework developed requires system administrators to manually run scripts to populate the database with new images; however, this framework can be extended to allow researchers to create their own projects, upload new images, and download data for CNN training.
Degree
thesis:*- Name thesis:degree_name
- Master of Science (MS)
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Computer Science
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mattingly, Marshall
- Contributors dc:contributor
-
- Travis Desell
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://commons.und.edu/theses/2281