APPLYING OBJECT DETECTION TO MONITORING MARINE DEBRIS

dc.contributor.advisor Peterson, Michael
dc.contributor.author Sherwood, Leah
dc.contributor.department Tropical Conservation Biology & Environmental Science
dc.date.accessioned 2020-06-19T17:36:34Z
dc.date.available 2020-06-19T17:36:34Z
dc.date.issued 2020-05
dc.description.degree M.S.
dc.identifier.uri http://hdl.handle.net/10790/5298
dc.subject Environmental science
dc.subject Computer science
dc.subject beach surveys
dc.subject Darknet/YOLO
dc.subject deep learning
dc.subject marine debris
dc.subject object detection
dc.subject Plastic pollution
dc.title APPLYING OBJECT DETECTION TO MONITORING MARINE DEBRIS
dc.type Thesis
dcterms.abstract Most of what is known about the type and distribution of plastic marine debris has been learned from beach surveys conducted by hundreds of researchers and volunteers since the 1980s. However, beach surveys of plastic marine debris require significant manual labor and lack harmonization across survey sites. In this thesis, I demonstrate how object detection technology based on deep learning can be deployed to partially automate the manual labor required to conduct beach surveys and upload the survey results to a centralized marine debris database. To create a proof-of-concept implementation, I developed an object detection system for marine debris using Darknet, an open source framework for convolutional neural networks, and the detection algorithm YOLOv3. I trained the detector on nine object classes: bags, bottlecaps, bottles, buoys, containers, hagfish traps, nets, oyster spacers, and “other,” and achieved a mean average precision (the standard metric of accuracy in the object detection literature) of 52%. The best performing class was hagfish trap, with an average precision of 80%, and the worst performing class was “other,” with an average precision of 34%. Next, a team of UH Hilo undergraduates and I migrated the system to the Android smartphone platform using Tiny YOLO, a smaller version of YOLO that was developed for running on low-powered computing devices. I compared the performance of YOLOv3 and Tiny YOLO at different image sizes with and without transfer learning (pre-training). My results demonstrate that it is possible to deploy object detection technology at beach survey sites to identify and count marine debris objects in real time. The technology is also applicable to other scenarios such as monitoring for plastic marine debris underwater or on the ocean surface. Ultimately, I expect the technology to be deployed as part of a “human in the loop” system in which the object detection component interacts with the person performing the beach survey so that the system can continuously improve in accuracy as it is used in the field while reducing the time and human labor costs associated with beach debris surveys.
dcterms.extent 67 pages
dcterms.language en
dcterms.publisher University of Hawaii at Hilo
dcterms.rights All UHH dissertations and theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission from the copyright owner.
dcterms.type Text
local.identifier.alturi http://dissertations.umi.com/hilo.hawaii:10183
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
Sherwood_hilo.hawaii_1418O_10183.pdf
Size:
5.99 MB
Format:
Adobe Portable Document Format
Description: