Abstract:

The adoption of remote sensing techniques in plant science enables noninvasive or minimally invasive measurement, which is also time and labor saving when compared to traditional field measurements. In this thesis, a method to distinguish switchgrass genotypes with the analysis of remotely-sensed hyperspectral imagery is proposed.

A processing protocol for hyperspectral imagery including preprocessing, ortho-mosaic generation, hyperspectral endmember detection and unmixing, and pixel-level genotype classification is presented. Machine learning algorithms for material detection, dimensionality reduction, and classification are proposed and discussed in detail. The impact of dimensionality reduction on classification is examined in depth.

The challenges in distinguishing between genotypes of the same species arise from the fact that the spectra of different genotypes from the same species are very similar with overlapping distribution. Furthermore, high between-band correlation in hyperspectral data suggests that dimensionality reduction of hyperspectral images may be useful. Whether dimensionality reduction can help in discriminating between spectra of genotypes of the same species is investigated. Three dimensionality reduction methods are described. The first method, principal component analysis, reduces dimensionality by linearly transforming the data to a space with lower dimensionality such that the maximal data variance is preserved. The other two methods reduce the dimensionality by finding a new feature space where the class separability is maximized. The second method, linear discriminant analysis, reduces dimensionality by linear projection. The third method, the Siamese network, reduces dimensionality by a non-linear mapping from the original space.

K-nearest neighbor (KNN) is used as the multi-class classifier in this work. KNN is one of the simplest non-parametric supervised classification algorithms which assumes that similar data points exist in proximity. The significance of reducing dimensionality by non-linear methods in improving the classification performance is manifested.

Links:

Citation:

H. Sheng, “Switchgrass Genotype Classification using Hyperspectral Imagery,” M.Sc. Thesis, Gainesville, FL, 2019. 

@MastersThesis{Sheng2019SwitchgrassHudanyunThesis,
author = {Hudanyun Sheng},
title = {Switchgrass Genotype Classification using Hyperspectral Imagery},
school = {Univ. of Florida},
year = {2019},
address = {Gainesville, FL},
month = {October},
}