Land use / land cover maps are usually created using some form of minimum distance
classification algorithm, which selects classes based on a minimized distance between
the spectral characteristics of each pixel to a set of clusters in multidimensional
spectral space.  This leads to a single class assignment per pixel.  We have been
investigating the utility of retaining information about the distances in spectral
space of each pixel from both the cluster that lead to the class assignment, as well
as to the “second-closest” cluster, in order to map classification confidence per
pixel.  Initial experiments made extensive use of the R statistics package and the
GRASS/R link, whereas current developments are focussed on a GRASS module, to handle
large datasets in a reasonable time frame.

Land use classification maps are often made by classifying remotely-sensed imagery
with some form of a minimum distance algorithm.  This procedure computes the distance
between a pixel’s unique spectral signature and a set of clusters within
n-dimensional space that represent discrete land cover categories.  Each pixel
receives a label corresponding to the closest predefined cluster.  Repeating this
process for each pixel leads to a classified map, which reflects the most probable
(maximum likelihood) case, given a set of spectral measurements.

One limitation to this approach occurs when pixels have virtually identical distances
to two or more clusters.  Especially if the distances are large, the pixel may not
clearly belong to any single category, and may represent mixed land cover.  Without
further information, such pixel classifications remain suspect.

We have proposed that retention of the distance to the second closest cluster can
shed light on the confidence in category assignment.  In previous studies, we
presented several examples of how such additional information might enhance accuracy
assessments and improve classification confidence.  Calculating multi-spectral
distances to cluster centroids (an approximation of the Mahalanobis distance) allows
comparison of all potential class assignments.  This provides a measure of relative
confidence in the actual classification at the level of individual pixels, and
highlights easily confusable classes.

We have developed software to calculate the distances of each pixel to all possible
clusters, and store the distances to (and corresponding class assignments of) the
closest and second closest clusters.  The algorithm was originally designed in R,
using the GRASS/R link for handling spatial data, and a forest inventory dataset
provided by the Canadian Forest Service.  While that system proved effective for
demonstrating the concept, the processing time for anything larger than a small area
was excessive - e.g. processing a LANDSAT scene was definitely not practical.  The
algorithm has been converted into a new GRASS module in C.  We present results using
this module on a test dataset, and welcome discussion of future possibilities.