Publication: Supervised data-driven analysis of hyperspectral imaging for early detection of water stress in plants
Loading...
Date
2024-02-01
Authors
Lin Jian Wen
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Understanding of plant’s biophysical information is crucial for advancement
in plant phenotyping and precision agriculture, where both ensure the sustainability
of crops. Using hyperspectral technology, abundant plant’s information can be
obtained in just one image which presented as a cube containing spectral and spatial
information. Hyperspectral imaging (HSI) captures wide range of spectrum, from
visible to near-infrared, allows different of plant traits to be monitored from
hyperspectral (HS) images. However, there are challenges to be tackled when
analysing HS image in reality. In this study, three objectives were achieved to solve
common challenges that exist in different stages of an HSI analysis. First, due to the
plant’s uneven surface and different inclination angles, illumination effect occurs and
add noise to the received spectrum. To recover the spectral signature, this study
investigated three normalization methods, Standard Normal Variate (SNV), Least
Absolute Deviations (L1) and Least Squares (L2) to quantify their performance in
noise reduction. Second, hundreds of bands were captured in an HS image and it
leads to curse of dimensionality. The performance in retaining variance were studied
on four spectral dimension reduction methods which are Principal Component
Analysis (PCA), Linear Discriminant Analysis (LDA), Analysis of Variance
(ANOVA) and Sequential Forward Selection (SFS). The third objective of this study
explored the possibility of deep learning model, 1D-CNN in early water stress
detection. A comparison with PCA-MLP model was conducted and its performance
in water stress detection was quantified. The findings in this study showed that SNV normalization outperformed L1 and L2 methods in mitigating illumination effect.
Additionally, when it comes to reducing spectral dimensions, both PCA and SFS
with 10 bands achieved better results in terms of retaining variance of original data
compared to ANOVA and LDA. However, PCA exhibited greater efficiency and
lower computational requirements. In the final experiment, the efficiency of 1D-
CNN in detecting early water stress and water potential prediction was demonstrated
and the performance was comparable to that of PCA-MLP. The efficiency of the
illumination correction and the spectral dimension reduction presentend in this study
establishes a promising foundation for the practical application of hyperspectral
imaging in detecting early water stress in plants for improving agricultural
monitoring