Comparison of Gaussian mixture and linear mixture models for classification of hyperspectral data

Use of hyperspectral data for military and civilian applications has spawned a number of techniques for automated and semi-automated characterization of spectral data. Characterization of spectral data according to linear mixture models and stochastic models has been used for classification of terrain and for enabling detection based on these data. Application of these techniques to hyperspectral data has presented a number of technical and practical challenges. The authors present a comparison of two fundamentally different models that are used to characterize and perform classification on spectral data: (1) Gaussian mixture and (2) linear mixture models. The characterization of hyperspectral data by each of these models is analyzed theoretically and empirically.