Hyperspectral image classification

Hyperspectral images are comprised of a large number of finely divided spectral bands, from 10s to 100s of such bands. The following analysis utilises a hyperspectral dataset produced by the NASA AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) project. As noted previously, hyperspectral imagery is become increasingly available and is essentially the same as multi-band imagery but with far more bands provided. It has particular application in geological work, both terrestrial and for planetary/lunar data analysis.

The dataset we use in this subsection is comprised of 224 contiguous spectral bands in the range 400nm (nanometres) to approximately 2500 nm at 10 nm intervals. The raster image files cover 512x614 cells of 20m squares and relate to a mining region in western Nevada, USA (hence c.10km by 12km in area). This dataset is available for download via the data section of the AVIRIS website and from the TNTMips web site as one of their tutorial datasets. It is discussed in more detail at the USGS spectral lab site: http://speclab.cr.usgs.gov, from which reference spectra may be obtained.

The layers of raster datasets of this type are spatially matched and highly correlated. This means that storage of such data as 224 raster objects would be extremely inefficient and hence of much more compact (compressed) 3D object variant is stored by TNTMips which they describe as a hypercube. Figure 4‑23 shows two alternative 2D representations of the region, where 3 spectral layers have been chosen to represent the red/green/blue (RGB) components of the image for display purposes. This provides a strong visual impression of the data, but each image is based on just 3 of the 224 layers and depends for its interpretation upon the choice of the three spectral bands and colour scheme applied. There are a number of alternative representations for datasets of this type. The 3D hypercube described above can be displayed and layers selected systematically or as desired (Figure 4‑24). This is essentially a visualisation tool that enables closer inspection of the data and highlights the correlated nature of the information, but does not provide a direct route to analysis. Several software packages support analysis of this type of dataset, including Idrisi (the HYPERSAM and HYPERMIN facilities) and the TNTMips spectral angle mapping facility, which we have used in the current example.

Figure 4‑23 2D map of Cuprite mining district, Western Nevada, USA


Bands: R=207, G=123, B=36	Bands: R=12, G=6, B=1

Figure 4‑24 3D hypercube visualisation of Cuprite mining district, Western Nevada, USA

Each pixel of the 3D hypercube has a spectral value for the ordered set of frequencies (i.e. across the depth of the cube). This ordered set can be viewed as a graph or spectrogram and can be compared with reference spectrograms (e.g. ground survey reference samples) or using laboratory sets obtained from prior research for different surface materials and minerals. Similarity of (normalised) spectral plots to the image spectral plot provides a means of classifying individual pixels and hence the region as a whole. Laboratory-generated reference spectra are available from various sources, including the USGS spectral analysis laboratory cited above. It should be noted, however, that laboratory-generated reference spectra will differ from remotely-sensed spectra owing to many factors, including non-uniform atmospheric distortions. Special facilities (e.g. the hyperspectral PROFILE facility in Idrisi) are available to identify and, where necessary, eliminate those bands that are subject to particularly high atmospheric distortion. As noted above, the analysis of hyperspectral images can be undertaken by regarding each cell or pixel as a single point in 224-dimensional space. The recorded value for each spectral band is then the coordinate value for this pixel for the selected dimension. The single point in n-space can be seen as a vector from the spectral origin, 0, with the full pixel set (512x614 pixels) being seen as clouds of points in n-space. The direction of these vectors is more important than their magnitude, since the latter are affected by average brightness. Simple spectral angle mapping (SAM) consists of comparing a known (reference) spectrum, which has its own characteristic m-dimensional vector, with those in the sample image. Pixels that have an angular proximity to the reference vector below a user-specified value are assigned to that class (cf. Table 4‑6, minimum distribution angle method). Every pixel in the source hypercube can be compared with the reference vectors in m-space and the angle between the pairs of vectors computed. The resulting set of scalar values can then be mapped as a form of “angular distance” raster (fuzzy, or soft class assignment) or just those values that are below a given angular distance threshold can be mapped (absolute, or hard class assignment). Figure 4‑25 shows an absolute (hard) class assignment for a single spectrum overlaid on Figure 4‑23B above (blue area).

Figure 4‑25 Single class assignment from spectral angle analysis