Loading...

Improving the Multivariate Classification Results of Hyperspectral Imaging Data by Considering Spatial Information Using Deep Learning Methods

Chogan, Hannaneh | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57417 (03)
  4. University: Sharif University of Technology
  5. Department: Chemistry
  6. Advisor(s): Parastar Shahri, Hadi
  7. Abstract:
  8. Hyperspectral imaging (HSI) has gained attention because it offers many benefits: it requires minimal sample preparation, is non-destructive, highly accurate, and provides detailed spatial information. However, the large amount of data it generates leads to high computational costs. To handle this and extract useful information from HSI data, various machine learning and chemometrics methods have been developed in recent years. This study aims to use a U-shaped deep neural network (U-Net) to classify hyperspectral imaging data in the 400-1000 nm wavelength range. The goal is to authenticate and detect adulterations in saffron, a valuable product. This method preserves the spatial information of the data, improving classification accuracy. It is being applied for the first time in food analysis and fraud detection. One challenge is obtaining pixel labels, which is done automatically using principal component analysis (PCA) to improve efficiency in verifying food authenticity. The study focuses on detecting three types of saffron adulterations: safflower, saffron styles, and calendula. The method also pinpoints the exact location of the adulteration by classifying pixels. Initially, the model was trained on mean spectrum images. To enhance performance, the 500 to 600 nm wavelength range was selected form hypercubes, and two-dimensional images were extracted. The model was then retested with various test data, including independent saffron samples and saffron with poly-crystal packages, to evaluate its generalizability. The results were impressive: image classification achieved a perfect accuracy of 1.00, and pixel classification ranged from 0.93 to 1.00, demonstrating the model's effectiveness. Finally, this model was compared with traditional models like partial least squares-discriminant analysis (PLS-DA), support vector machine (SVM), and artificial neural network (ANN). These traditional models had accuracies of 0 and 0.5 when tested with independent data and saffron with a poly-crystal packages, respectively, showing they were less reliable than the U-Net model. The U-Net not only classifies pixels but also identifies the precise location of adulterations. The results highlight the power of deep learning methods in analyzing hyperspectral data, especially when considering the spatial aspect
  9. Keywords:
  10. Hyperspectral Imaging ; Food Authenticity ; Chemometrics Method ; Machine Learning ; Deep Learning ; Principal Component Analysis (PCA)

 Digital Object List

 Bookmark

No TOC