Nonlinear unmixing of mixed pixels in remote sensing imagery are commonly conducted with neural networks (NN) models

which

however

lacks physically based interpretation as linear models. This paper the difficulties of knowing patterns of pixel mixture and error distribution to some extent. This paper proposes to use mean square error

bivariate distribution function

confidential error

and synthetically mixture complexity techniques to analyze and evaluate the pixel mixture

which can provide insights in some features of nonlinear mixture models. Experiment with MODIS associated with ETM+ data demonstrates that the visualized method can obtain the decomposition characterization of PPLN model efficiently. In addition

visualized assessment shows that PPLN provides higher accuracy compared with the BP neural network. The overall unmixing error decreases from 0.182 8 to 0.171 7 in terms of RMSE

improved by 6.5%. The experiment also demonstrates that urban and sparse vegetation are the potential occurrence places where pixels are severely mixed.