Spectral Unmixing and Minimum Volumn Constraint in Disentangled Representation Learning with Vision Transformer