Please use this identifier to cite or link to this item:

Invasive species detection from RGB aerial imagery: investigating links between plant characteristics and transfer learning success

File Size Format  
Ta_hilo.hawaii_1418O_10215.pdf 1.88 MB Adobe PDF View/Open

Item Summary

Title:Invasive species detection from RGB aerial imagery: investigating links between plant characteristics and transfer learning success
Authors:Ta, Erica
Contributors:Perroy, Ryan L. (advisor)
Tropical Conservation Biology & Environmental Science (department)
Keywords:Conservation biology
Artificial intelligence
Remote sensing
deep learning
show 3 moreinvasive species
object detection
transfer learning
show less
Date Issued:Dec 2021
Publisher:University of Hawaii at Hilo
Abstract:Advancements in remote sensing techniques and deep learning applications like object detection have improved invasive species monitoring systems. Deep learning typically uses large training sets, up to millions of images, to consistently recognize targets, but generating these training sets may not be practical for incipient invasive species targets of interest. When large datasets are unavailable for training, one approach is to use transfer learning to overcome data limitations. The process applies knowledge learned from the source network (pre-trained on a task on which large datasets are available) to a target problem that has limited data samples. Here I examine how object detection performance for the following invasive species of interest in Hawaiʻi differs with the inclusion of cross-species transfer learning: Miconia (Miconia calvascens), Guinea grass (Megathrysus maximus), and four symptomatic visible classes of Rapid ‘Ōhi‘a death (ROD): red, brown, fine white, and skeleton. I also measured visual plant features of contrast, shape, size, and texture to understand how different plant morphologies provide easier or more challenging scenarios for plant object detection using aerial visible imagery. I found that 9 out of 30 transfer learning instances had significantly higher mean average precision (mAP) scores than instances without transfer learning (p < 0.00167, α = 0.00167 (0.05/30). Transfer learning was found to be most effective between red, brown, fine white, and skeleton ROD classes and least effective among miconia and guinea grass. The feature measurement of contrast was significantly correlated with source model mAP (R = 0.82, p = 0.045) whereas texture was strongly correlated (R = 0.77, p = 0.073), size (R = 0.54, p = 0.27) was moderately correlated, and circularity (R = -0.096, p = 0.86) was weakly correlated. My results indicate there are advantages in plant detections that utilize similar source and target candidates for transfer learning, in addition to incorporating source candidates whose image data exhibit higher contrast and higher texture measurements. Overall, this study may inform future workflows to detect plants from aerial imagery by demonstrating how available data can be best leveraged or repurposed through transfer learning to detect a plant target using limited available datasets.
Pages/Duration:62 pages
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.
Appears in Collections: Tropical Conservation Biology and Environmental Science
TCBES Theses

Please email if you need this content in ADA-compliant format.

Items in UH System Repository are protected by copyright, with all rights reserved, unless otherwise indicated.