This application illustrates the use of our fast hierarchical cluster analysis to describe similar geographical regions from remotely sensed satellite data.  Each region typically comprises a rectangular grid of 200x400 = 80,000 pixels, for which 44 attributes are collected.  We use Cluster Data in ClustanGraphics to construct a hierarchical classification tree for the 80,000 cells, and then map the hierarchical classification tree directly on to the geographical image (as below).

Once the data has been clustered hierarchically using cluster data, the tree is saved from ClustanGraphics to a text or Excel file.  It can then be read into a visualization program where the clusters are spatially mapped on to the original image as shown above.  The horizontal scroll bar is used to collapse or separate the clusters, thereby displaying many levels of sensitivity from coarse (2 clusters) to fine (e.g. 1000 clusters).  In the above image, NDVI colours are used to display an ecological classification of a region of East Africa at the 10 cluster level.

This method was then used to construct an ecological map of the whole of Africa as viewed from space, comprising nearly ½m pixels.  The hierarchical cluster analysis now permits a sequence of these images to be viewed at different levels of sensitivity, from ½m clusters down to 1, as a movie.  An example of an intermediate image from this sequence is shown right, revealing the principal biophysical regions within Africa.

Scientific reference: Boone, R.B., Galvin, K.A., Smith, N.M., and Lynn, S.J.  2000.  Generalizing El Nino effects upon Maasai livestock using hierarchical clusters of vegetation patterns.  Photogrammetric Engineering and Remote Sensing,  66:737-744.  Further details here

.