1 INTRODUCTION
One of the most frequently used methods to obtain a land cover map is the supervised classification of multispectral
LANDSAT images, but often the result is not enough accurate for most of the practical remote sensing applications.
Even using multitemporal images to capture the phenological evolution of the vegetation along the year, a high level of
uncertainty in some classes is often present.
Nevertheless it is very common to have several data sets from a given physical geographic area, either data acquired at
different moments or by different sensors, and from each one we may obtain a land use classification using different
techniques. Conceptually, each data source is better suited to extract certain characteristics, so it becomes necessary to
have a method to combine them getting the best from each one.
A supervised classification provides a likelihood distribution that tells us the assignment probability of a pixel to each
one of the legend classes. Using some test areas we can also obtain an individual measure of per cent classification
success ratio for each data source. The method we present here to merge two land use maps of the same geographic area
relies on the combination of the likelihood (assignment probabilities) and the classification success ratio of each data
source. Combining them, we obtain a new assignment probability distribution and, therefore, a new classification.