Classification of heterogeneous solids using infrared hyperspectral imaging.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Heterogeneous powder mixtures present classification challenges due to variable component spatial arrangements in hyperspectral images.
• Accurate classification requires methods that capture sample heterogeneity despite non-congruent image data.

Purpose of the Study:

• To develop and evaluate a robust method for classifying heterogeneous powder mixtures using FT-IR hyperspectral imaging.
• To address the issue of non-congruent image data by creating representative feature vectors for each sample.

Main Methods:

• Exploration of dimension reduction techniques to generate feature vectors from FT-IR hyperspectral images.
• Classification using Discriminant Analysis (DA) and Soft Independent Modeling of Class Analogy (SIMCA).
• Development of a median-interquartile range "super-spectrum" as a feature vector to represent image heterogeneity.

Main Results:

• The "super-spectrum" feature vector combined with Principal Component Analysis (PCA) and DA achieved 87.5% accuracy on training data (leave-one-out cross-validation).
• This approach demonstrated 100% accuracy on an independent test set.
• Compared to single-point spectra (52.5% training, 72% test accuracy), the "super-spectrum" method showed superior performance.

Conclusions:

• The median-interquartile range "super-spectrum" is an effective feature vector for classifying heterogeneous samples in FT-IR hyperspectral imaging.
• This chemometric approach significantly enhances classification accuracy for complex powder mixtures compared to traditional spectral analysis.
• The developed method offers a reliable solution for analyzing non-congruent hyperspectral image data.