摘要
The least squares approaches are widely used in remote sensing image analysis to solve the linear mixture model. They assume the spectra of the endmemebers are known and fixed vectors for linear unmixing. But it is clearly shown from the spectral libraries that one material has various spectra. Therefore, total least square has been proposed to have the robustness to accommodate those variations and achieve minimum error. In this study, we apply two constraints on the estimated abundance in total least square: sum-to-one and nonnegative constraints. These two constraints ensure the sum of all estimated abundance is one and no abundance fraction is less than zero. The performance comparison with regular least square approaches is conducted with a hyperspectral image scene.