Using variance analysis to differentiate similar classes in hyperspectral imagery

Several phenomena in hyperspectral imagery affect the ability to differentiate pixels belonging to similar classes. Sensor noise, pixel mixing at various spatial resolutions, bit depth reduction from conversion to reflectance, shadows, specular reflectance and class specific properties all complicate class differentiability. This paper presents a statistical testing method to aid in the separation of similar classes. This hypothesis test with variance as the parameter of interest is performed on synthetic and real hyperspectral imagery. The results show that statistical tests based on the model of the spectral presented here can sucessfully identify samples containing only a single class and those containing multiple classes.