A new approach to unmixed spectral areas determination in remote sensing classification

In this paper we propose a new methodology to determine homogeneous spectral training areas. A set of spectral signatures for heterogeneous training areas are provided by a classical classification method. Then, a set of threshold limits has been established through a statistical study to separate spectral signatures into different subclasses which initially seem to belong to the same class. That procedure have been integrated in a computer tool (CASUS), which allows determining automatically the homogeneous spectrally training areas in the scene. The training areas determined by CASUS can improve noteworthy image classification traditional methods.