I want to try several "backgrounds" for change detection on bitemporal images (satellite images of the same place but in different moments in time). Images consist of 5 bands. I have python implementations of the MAD transform (which maximizes the correlation of linear combinations of the bands of the two images), PC transform (which maximizes the variance of each linear combination of the images and then we would look at the substraction of these) and a simple algebraic method which solely consists of substracting the images, one band at a time. As far as it goes, everything looks great but I want to measure the performance of the algorithms. Since I am also interested in spatially coherent change I would maybe do this for 1 row like explained before but then for 2 rows: 111111 111111 222222 222222 etc
and so on for 3,4,5,.. rows... Does this sound completely absurd? Also maybe some images with salt and pepper noise, and then images with random squares of 2x2 pixels, 3x3, etc. Even if this would't be ridiculous (please tell me) how would I reconcile this with the idea of having 5 bands, after all the algorithms are designed to incorporate the information of several bands. Sorry for the long read, hope I didn't bore you too much. Greetings!