It is my understanding that a grid of pixels are gathered to represent a sample of "good" surface pixels found within the scene, From this sample of pixels a series of sums are calculated to obtain the covariance matrix for the three channels. Then the covariance and the correlation matrices are computed. The eigenvectors and eigenvalues described by the correlation matrix are then computed. The stretching vector (or normalized vector) is formed by taking the reciprocal of the square root of each value in the eigenvalue vector. The final transform is then composed of the the matrix of the eigenvectors and the stretching vector. This final transform is then what is applied to the data. I hope this is what you were looking for. Ryan

Thanks Ryan for the explanation but I am really looking for a reference article or a book in order to understand what was really implemented in ENVI, and use the reference(s) in publications when needed. The basics of the technique is in principle as described in the tip but then why should it result similar, not the same? I guess I will follow Gillespie et al. 1986 or the given tip and do it by following the steps described. Nilda

I'm afraid I haven't been able to track down which original reference was used by the ENVI developers when designing ENVI's Decorrelation Stretch tool. But it seems to be essentially what was described by Gillespie et al. 1986. - Peg Exelis VIS