You can't do this! You have to be clear in one thing that a single pixel reflectance varies with satellite to satellite based on number of factors. I.e., illumination variation between time as well as adjacency are major issues. you have to do your tracking based on feature to feature. If you clearly understand least mappable unit area in each resolution you can get the knowledge behind your query.
You can find the affine transformation between each successive image. So lets say you have three image I1,I2,I3 and you want to track pixel (i,j) in I1. First calculate the transformation between I1 I2 to get T1-2. Than apply the transformation T1-2 on the pixel (i,j) to get its location of this pixel in I2. lets say the new location of the pixel is (n,m) now find the transformation between I2 and I3 T2-3 and apply the traqnsfomration on the pixel (n,m) in I2 to get the location of the pixel in I3 lets call it (x,y. So you have your track of your pixle from (i,j) -> (n,m) -> (x,y)
I agree with the previous answers. I can add another suggestion: if you are using images from the same sensor (let say all are from TM-Landsat 5 satellite) or from different sensors on different satellites with the same spectral bands and resolution (let say TM-Landsat 5 and ETM-Landsat 7...or SPOT 4, 5, ...etc.) and you can do precise georreferencing of each image with ground control points, you resample each image using nearest neighbour (to avoid interpolation that change pixel values), and you will get a set of images that will be pixel to pixel corregistered, then you can track the pixel changes along time, doing a multitemporal analysis. By this way you are transforming each image to the same reference system without changing pixel values. If the georreferencing is precise (let say the RMS error of the adjustment is close to 0.2 of the pixel size) the corregistration probably will be good for your purposes.