I only have some generic insights that might help. Firstly you are analyzing a thermal image and the multi-layer co-occurance matrix that you have mentioned is actually known as the grey-level co-occurrence matrix (GLCM) and is a textural measure (image texture). Textural measures are spatial aspects of any imagery so basically what you are looking at is the spatial facet of the thermal image and what you should do is consider which GLCMs (whether entropy, mean, homogeneity, correlation, dissimilarity, or second moment) suits your analyses. You may not likely utilize all the GLCMs since some are highly correlated to certain phenomena e.g. entropy was used to in a study for evaluation of the complexity of tree stand structure (Lu et al, 2005). Therefore firstly you have to test each GLCM on several of the window sizes (kernels) e.g. 3x3, 5x5, 7x7, etc to see which of them effectively depicts the patterns you are trying to identify or match. After that you can easily match and evaluate the patterns. Finally, remember that a combined measure of all the GLCMs may not truly portray what you are trying to investigate so it is better to test each one to identify their appropriateness. I hope that will be helpful for you.

In general, if one speaks about the correlation, one has to take into account all the values of the array. The pattern is a sequence of features: one can build a linear correlation coefficient for A={011011011000} and B={ 011011011100}, but the correlation of the masked sequence A_masked={111111} and B_masked={111111} will be the same even if the B= { 111111111111}. I don't know the ENVI software, but these considerations are general, so I would not mask the image.

I only have some generic insights that might help. Firstly you are analyzing a thermal image and the multi-layer co-occurance matrix that you have mentioned is actually known as the grey-level co-occurrence matrix (GLCM) and is a textural measure (image texture). Textural measures are spatial aspects of any imagery so basically what you are looking at is the spatial facet of the thermal image and what you should do is consider which GLCMs (whether entropy, mean, homogeneity, correlation, dissimilarity, or second moment) suits your analyses. You may not likely utilize all the GLCMs since some are highly correlated to certain phenomena e.g. entropy was used to in a study for evaluation of the complexity of tree stand structure (Lu et al, 2005). Therefore firstly you have to test each GLCM on several of the window sizes (kernels) e.g. 3x3, 5x5, 7x7, etc to see which of them effectively depicts the patterns you are trying to identify or match. After that you can easily match and evaluate the patterns. Finally, remember that a combined measure of all the GLCMs may not truly portray what you are trying to investigate so it is better to test each one to identify their appropriateness. I hope that will be helpful for you.

thank you Glen i will study the cited paper. The "high correlation to certain phenomena" is what i'm looking for i would like to give a name to certain phenomena and to correlation, but in literature there aren't many papers with this aim. I'll try to do what you suggest me evaluating the results.

You're welcome Claudia and I think its a good idea to examine how GLCM correlation is used in the paper I cited. This paper gives at least a good description of how GLCMs are tested on several kernels (window sizes) to identify their suitability of correlation. Just FYI the paper is titled: Above-ground biomass estimation using Landsat TM data in the Brazilian Amazon.