Interactive exploratory analysis of large hyper spectral images

Interactive exploration of hyperspectral images is very important part of analysis and, often, the first step most of researchers do. Usually, it employs PCA (or similar) decomposition of an image and brushing — selection of pixels simultaneously on score density plot and on channel or score images. By excluding the selected points and consequently creating a new model for the remaining data it is possible to get rid of outliers and background or to make a local model for a particular group of pixels. Interactivity of this procedure is quite important and, therefore, it requires a powerful computer with enough RAM to operate.

However, in case of large images (from hundreds of thousands of pixels and above), the interactive exploration becomes more and more tricky as the decomposition algorithms (such as e.g. SVD and NIPALS) are getting very slow. One of the ways to solve this issue is to use probabilistic approach, which allows to reduce the dimension of dataset needed for computing PCA loadings — the slowest part of the algorithms.

In this presentation we are going to demonstrate benefits of the probabilistic approach and compare it with conventional algorithms in terms of speed and accuracy of PCA decomposition using real life examples.