Evaluating high-resolution remote sensing data and machine learning for detecting strawberry guava (Psidium cattleyanum) and its biocontrol on Hawai'i island

Date

2025-05

Contributor

Instructor

Depositor

Speaker

Researcher

Consultant

Interviewer

Narrator

Transcriber

Annotator

Journal Title

Journal ISSN

Volume Title

Publisher

Volume

Number/Issue

Starting Page

Ending Page

Alternative Title

Abstract

Strawberry guava (Psidium cattleyanum) is the most abundant non-native tree species across the Hawaiian Islands. This invasive species can create nearly impenetrable, dense thickets across rugged terrain which makes it difficult to map its distribution. The Brazilian scale (Tectococcus ovatus) is a gall forming insect that is currently being deployed as the biological control to reduce the spread of strawberry guava. High resolution remote sensing technologies, including sensors attached to small unoccupied aircraft systems (sUAS) and helicopters, allow researchers to survey more land in a short amount of time with machine learning models to find invasive species and observe the effects of management actions. This project uses remote sensing data, including high resolution RGB aerial imagery, lidar data, and machine learning to detect strawberry guava and implements computer vision techniques to detect biocontrol spread. Helicopter and sUAS flights collected these data for a portion of ‘Ōla‘a Forest Reserve on Hawai‘i Island. Additionally, ground truth surveys were conducted over three 20 x 20 m plots where every tree was exhaustively surveyed. Images of strawberry guava in the canopy during different phenological stages and leaves with Brazilian scale galls present were annotated and used to train convolutional neural network (CNN)-based object detection models with the python library DeepForest. In parallel, forest structure metrics including canopy height, foliage height diversity, plant area index, and point density at z-values equal to four to six meters were derived from the lidar data and used to classify areas of strawberry guava via a random forest classification model. The CNN detector was able to detect strawberry guava in aerial imagery better during periods when the canopy underwent a red flush, or new leaf growth, (mAP = 0.85 (SD = 0.02), F1 score = 0.58 (SD = 0.03)) compared to a green leaf canopy (mAP = 0.79 (SD = 0.02), F1 score = 0.50 (SD = 0.02)). The random forest model using forest structural metrics derived from lidar performed better in terms of precision (F1 score=0.72) than both CNN models while the red phenology stage CNN had better recall. Plant area index closely followed by canopy height and foliage height diversity were the metrics in the model with the greatest mean decrease accuracy, or importance to the model’s performance. The models captured large stands of strawberry guava in both aerial imagery and lidar data, but struggled to capture individual strawberry guava tree stems. For the Brazilian scale, another CNN-based object detection model was developed and was able to detect galling on strawberry guava leaves (mAP = 0.57 (SD = 0.02), F1 score = 0.40 (SD = 0.02)), but low precision values indicated many false positive predictions due to rain, discoloration, blur, and lighting. Using these tools to identify strawberry guava is useful for selecting candidates for ongoing biocontrol deployments and could be applied at a larger scale to map the presence and absence strawberry guava in Hawai‘i. Monitoring the Brazilian scale progress in the canopy with aerial imagery and machine learning will help better determine whether its establishing in strawberry guava canopy and spreading over time to reduce tree spread into high elevation native forests.

Description

Keywords

Remote sensing, Environmental science, Computer science, biological control, convolutional neural network, invasive species, lidar, random forest, RGB aerial imagery

Citation

Extent

90 pages

Format

Geographic Location

Time Period

Related To

Related To (URI)

Table of Contents

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.

Rights Holder

Local Contexts

Email libraryada-l@lists.hawaii.edu if you need this content in ADA-compliant format.