Abstract

A leading cause for mortality in the pine forests of western North America, the mountain pine beetle, has impacted over 400,000 acres of ponderosa pine forest in the Black Hills of South Dakota since 1996. Methods aimed at earlier detection, prior to visual manifestation of a mountain pine beetle damage in the tree crown, have not been successful because of the overlap and variability of spectral response between the initial stages of attack (green-attacked) and non-attacked tree crowns. Needle-level reflectance spectra was measured from green-attack and non-attack ponderosa pine trees in early spring following an infestation and analyzed using a multi-statistical approach to determine which spectral features best discriminate green-attack needles. Green-attack reflectance was significantly higher than non-attack from 424-717 nm and 1151-2400 nm. Bands in the shortwave-infrared had increased measures of separation between classes compared to visible and near-infrared bands. Peaks in separation related to moisture absorption features, from 1451-1540 nm and 1973-2103 nm, and pigment absorption features from 462-520 nm and 663-689 nm, were consistently observed over multiple statistical analyses. While these features show promise for operational canopy-level detection, it is unknown if they can be scaled up due to large within-class variability and spectral overlap between classes.

To examine the potential for canopy-level detection, in-situ training data was collected for green-attack and non-attack trees from known locations within the Black Hills at a similar time a WorldView-2 image was acquired of the study area. Along with eight WV-2 bands, all possible normalized two-band indices were calculated to examine the suitability of WV-2 data for detecting green-attack damage. The performance of three different classifiers, logistic regression, linear discriminant analysis, Random Forest, was evaluated. Normalized two-band indices using a combination of a near-infrared band and visible band increased separation compared to single WV-2 bands. Random Forest classifiers using the eight WV-2 bands as predictors yielded an independently validated accuracy of 70.6%. Compared to non-attack, green-attack class accuracies were lower, likely due to the high within-class variance and spectral overlap between classes observed. Even with these limitations, the methods presented offer improvements over existing green-attack detection methods.

Advisor

Martin Mitchell

First Committee Member

Fei Yuan

Second Committee Member

Christopher Ruhland

Date of Degree

2016

Language

english

Document Type

Thesis

Degree

Master of Science (MS)

Department

Geography

College

Social and Behavioral Sciences

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS