Exploration of datasets of hyperspectral images

Hyperspectral images of size usually greater than 50 MB can be easily acquired in very short times, generally without the need of sample pretreatment. While Multivariate Image Analysis (MIA) tools can be efficiently used for the exploration of single hyperspectral images or of groups composed by a limited number (say up to 10) of merged images, the exploration of datasets composed by a large number (>10) of images is less straightforward. However, a representative sampling of a large number of specimens is frequently required to correctly estimate both intra- and inter-sample variability. This implies the acquisition of datasets composed by a large number of hyperspectral images and of several GB in size, especially in those cases where only one or a few samples can be included in a single image scene. In this context, the exploration of the dataset by applying MIA to single images or to subgroups of merged images does not allow to gain a global overview of the entire dataset variability and to properly highlight the possible presence of outliers, clusters and/or trends. A fast procedure which can be adopted to deal with this issue consists in computing the average spectrum of each image, to build a matrix of average spectra of the analyzed hyperspectral images. Although this approach leads sometimes to satisfactory results (especially when dealing with homogeneous materials), the information related to spatial variability is lost, and the hyperspectral image data are turned into “common” (i.e., not spatially resolved) spectral data. By averaging spectra, for example, the useful information related to the presence of a defect localized in a relatively narrow image area could be diluted within the massive amount of other “well behaving” pixels, becoming no longer detectable. Aiming to develop a fast and easy-to-use tool able to facilitate the exploration of large datasets of hyperspectral images while maintaining both spectral- and spatial-related information of the original images, we have proposed an approach which consists in automatically converting each hyperspectral image into a signal named hyperspectrogram [1]. Essentially, the hyperspectrogram can be viewed as a fingerprint containing the relevant information brought by the original hyperspectral image, and is composed by a first part accounting for the spatial information and by a second part accounting for the spectral information. By representing each image with a vector of few hundreds of points, this procedure enables to compare simultaneously up to hundreds of images by means of common multivariate analysis methods, such as PCA. In order to facilitate the exploration of datasets of hyperspectral images through hyperspectrograms, we have recently developed a Matlab Graphical User Interface (GUI), which easily allows calculation and visualization of hyperspectrograms, exploration of the dataset and visualization of the features of interest contained within each single sample directly in the original image domain.