Using Cluster Analysis to Improve the Selection of Training Statistics in Classifying Remotely Sensed Data