PCA is performed on the difference image and the first three PCs are selected.
The PCs are divided into three classes either using EM algorithm or any unsupervised classification algorithm such as k-means.
Pixels belonging to the second cluster : no-change pixels in the ICM.
These pixels are used for further analysis.
P. Gong, “Change detection using principal component analysis and fuzzy set theory.” Can. J. Remote Sens ., vol. 19, pp. 22–29, 1993.
Multispectral change detection using IR-MAD and initial change mask.
Experiments with ICM Landsat ETM+ images over Juelich, Germany taken in May, June and August, 2001. Major changes due to agricultural regions. Only the human settlement areas and mining area remain unchanged.
MAD is invariant to linear/affine transformations.
Let P , Q be the PCA transformation images of the image pair F , G .
The MAD variates, M i , can be expressed in terms of P , Q as
M i =U i -V i = c i T P - d i T P ,
where c i , d i eigenvectors of the generalized eigenvalue problem.
Main assumption: components with low eigenvalues are noise components
=> Do not contribute to the correlation of the no-change pixels
=> Do not contribute to the top MAD variates
PCs with low eigenvalues are then ignored , hence facilitating hyperspectral change detection.
Effect of feature reduction in the multispectral case Comparison of the slopes for radiometric correction using manually selected no-change pixels and automatically selected no-change pixels using different inputs.