Hyperspectral NIR image analysis

Burger, James (2006) Hyperspectral NIR image analysis. Doctoral diss. Unit of Biomass Technology and Chemistry, SLU. Acta Universitatis agriculturae Sueciae vol. 2006:60.

Full text available as:

PDF (Thesis) 12526 Kb

Abstract

Hyperspectral images add a new dimension to the field of spectroscopy, specifically spatial resolution. In addition to the identification and quantification of bulk constituents provided by integrating type spectrometers, hyperspectral images provide a means of accurately quantifying and locating constituent variation within the field of view of the camera.

Hyperspectral images provide a massive quantity of data, and as with NIR spectroscopy, multivariate chemometrics tools must be utilized to appropriately extract accurate information. This thesis looked at techniques to clean and modify or condition the raw spectral data to improve the prediction results of regression techniques such as PLS. It was found that extra diagnostic tools for regression models could be based on image data. A new metric based on a combination of prediction bias and variance was proposed for determining the number of latent variables.

Data set conditioning was based on several approaches. Sets of standard reference materials were used to improve conversion of data counts into percent reflectance units and to provide instrument standardization. A multi-step approach to outlier detection was formulated that incorporated thresholding tests for excessive data values, combined with tests based on Euclidean distance measurements and angle cosines between spectra. Finally, various spectral pretreatments or filters were considered to complete the spectral cleaning and modification process.

Results from the application of multivariate analysis techniques to this optimally conditioned data were presented. Data visualization tools included histograms and spatial mapping of constituent concentration predictions, colorization of score plots, and false color image presentations of combinations of score images or prediction maps.

The use of these data exploration, correction, and regression tools was demonstrated by the systematic analysis of increasingly complex data samples. Carefully designed laboratory samples were used to examine the theoretical limitations of prediction of chemical content and correction for physical properties including the dependencies of diffuse light scattering effects on particle size. Sample sets of cheese and wood pellets were used to demonstrate the overall utility of proper data conditioning in the application of hyperspectral NIR imaging to more difficult real-world problems.