Date of Award

Summer 8-22-2020

Level of Access

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Forest Resources

Advisor

Peter Nelson

Second Committee Member

Daniel Hayes

Third Committee Member

Ryan Hanavan

Additional Committee Members

Aaron Weiskittel

Shawn Fraver

Abstract

Water management and irrigation practices are persistent challenges for many agricultural systems. Changing seasonal and weather patterns impose a greater need for understanding crop deficiencies and excesses (e.g. water, sunlight, nutrients) for optimal growth while allocating proper resources for prompt response. The wild blueberry industry is at heightened susceptibility due to its unique growing conditions and uncultivated nature. Early detection of stress in agricultural fields can prompt management responses to mitigate detrimental conditions including drought and disease. Remote sensing has provided timely and reliable information covering large spatial extents, while novel applications in hyperspectral data and imaging spectroscopy have shown potential in early stress detection. We assess airborne spectral data accompanied by ground sampled water potential over three developmental stages of wild blueberries to accurately detect water content. Airborne scans of spectral data were collected three times throughout the 2019 summer in Deblois, Maine. Data were collected over two adjacent fields, one irrigated and one nonirrigated. Ground sampled data were collected in tandem to the UAV collection. The ground sampled data over the irrigated and non-irrigated fields guided digital sampling from the imagery to act as training for our models. Using methods in machine learning and statistical analysis, we related hyperspectral reflectance measurements to different water potential levels in blueberry plant leaves to decipher vegetation signals both spatially and temporally through utilizing the capacity of imaging spectroscopy. Models were developed to determine irrigation status and water potential. Seven models were assessed in this study with four used to process six hyperspectral cube images for analysis. These images were classified as irrigated or non-irrigated and estimated water potential levels. Our global water potential model had an R2 of 0.62. Models for the water potential predictions were verified with a validation dataset. Forest insect and disease pests have a significant impact on the well-being of individual trees and forest stands, affecting ecosystem processes and potentially human health. Dispersing through 35 states within only 17 years (USDA, 2020), the effect of emerald ash borer (Agrilus Planipennis Fairmaire) (EAB) in the United States has been particularly severe and devastating. Early detection of stress in forests can prompt management responses to mitigate detrimental conditions including drought and disease as well as pest outbreaks. Remote sensing has provided timely and reliable information covering large spatial extents, while novel applications in hyperspectral data and imaging spectroscopy have shown potential in early stress detection. We build on previous work by assessing airborne spectral data, and health classifications of EAB infested ash trees in aims to accurately detect stress. Airborne scans of spectral data were collected within three days in late July 2019 over three sites in southern New Hampshire. Ground sampled data were collected in November 2019 and include sampled ash classified on a scale of 1-5 (1=healthy, no major branch morality, 5=dead). The ground sampled data of different health classifications guided digital sampling from the imagery to act as training and validation for our models. Using methods in machine learning and statistical analysis, we related reflectance measurements to different classifications of ash tree health to understand tree stress signals while utilizing the capacity of remote sensing. Models were developed to classify health in ash trees impacted by EAB. The first entailed a shadow classifier, followed by one for health. Eighteen cube images contained ground sampled data and were processed with the two models, then further buffered. Pixel classification for each buffer sample was calculated. The health classifier model was used on a validation test set and had an prediction accuracy of 76.1%.

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share