Options
March 30, 2023
Master Thesis
Title
Investigating Machine Learning Techniques to Improve Hyperspectral Data Classification Performances through Added Multi-Modal Features
Abstract
Currently, one of the biggest challenges in the agricultural field is the identification
and classification of plant health-related problems, specifically stress caused
by drought. Consequently, analyzing the ground field-based hyperspectral data
covering the visible, near-infrared and mid-infrared ranges of the electromagnetic
spectrumprovides us with valuable information about the plant’s physiological
conditions even before any significant changes in its physical appearance (shape
or colour) are observed. Consequently, hyperspectral technology has become increasingly
crucial for emerging agricultural applications. The thesis focus on how
these emerging technologies (hyperspectral field-based sensing and machine
learning) are being utilized to solve these challenging issues. The hyperspectral
sensing applications, as well as the machine learning various techniques, have
helped a lot in investigating the plant’s health, classification of the vegetation
or crop types, plant’s biophysical detection (e.g., biomass), disease and stress
detection (e.g., drought) and also nutrition components detection (e.g., Nitrogen).
Themain focus of themaster’s thesis is integrating the hyperspectral sensed
data with these multiple data resources (e.g., photosynthesis traits, metabolic
traits and water uptake traits). Following that, producing more consistent and
valuable information than the hyperspectral sensed data provided independently,
leading to improving the classification of stressed plants caused by drought. In
this way, we can increase crop health by improving the accuracy of the built
machine learning algorithm by providing it with more information about the
drought-stressed vs treated barley plants. In other words, we improve the machine
learning algorithm’s accuracy by combining the physiological laboratory
measurements with the field-based hyperspectral data of barley plants. Using
means of feature engineering (e.g., Sequential Forward Selection and Sequential
Backward Elimination) to choose the added features as well as feature extraction
techniques(PCA,t-SNE) to overcome the curse of dimensionality of the hyperspectral
imaging data and accurately represent it. Initially, the thesis research
involved performing a preprocessing phase to ensure the samples’ quality and
scaling the dataset before executing any predictions. Additionally, selecting the
best-performing machine learning models based on the group cross-validation
(Leave-one-out) method results. Moreover, evaluating the performance of the
model’s predictions after adding these measurements, compared to the model
with the monomodal hyperspectral data for the classification task.
and classification of plant health-related problems, specifically stress caused
by drought. Consequently, analyzing the ground field-based hyperspectral data
covering the visible, near-infrared and mid-infrared ranges of the electromagnetic
spectrumprovides us with valuable information about the plant’s physiological
conditions even before any significant changes in its physical appearance (shape
or colour) are observed. Consequently, hyperspectral technology has become increasingly
crucial for emerging agricultural applications. The thesis focus on how
these emerging technologies (hyperspectral field-based sensing and machine
learning) are being utilized to solve these challenging issues. The hyperspectral
sensing applications, as well as the machine learning various techniques, have
helped a lot in investigating the plant’s health, classification of the vegetation
or crop types, plant’s biophysical detection (e.g., biomass), disease and stress
detection (e.g., drought) and also nutrition components detection (e.g., Nitrogen).
Themain focus of themaster’s thesis is integrating the hyperspectral sensed
data with these multiple data resources (e.g., photosynthesis traits, metabolic
traits and water uptake traits). Following that, producing more consistent and
valuable information than the hyperspectral sensed data provided independently,
leading to improving the classification of stressed plants caused by drought. In
this way, we can increase crop health by improving the accuracy of the built
machine learning algorithm by providing it with more information about the
drought-stressed vs treated barley plants. In other words, we improve the machine
learning algorithm’s accuracy by combining the physiological laboratory
measurements with the field-based hyperspectral data of barley plants. Using
means of feature engineering (e.g., Sequential Forward Selection and Sequential
Backward Elimination) to choose the added features as well as feature extraction
techniques(PCA,t-SNE) to overcome the curse of dimensionality of the hyperspectral
imaging data and accurately represent it. Initially, the thesis research
involved performing a preprocessing phase to ensure the samples’ quality and
scaling the dataset before executing any predictions. Additionally, selecting the
best-performing machine learning models based on the group cross-validation
(Leave-one-out) method results. Moreover, evaluating the performance of the
model’s predictions after adding these measurements, compared to the model
with the monomodal hyperspectral data for the classification task.
Thesis Note
Magdeburg, Univ., Master Thesis, 2023
Author(s)
Language
English