According to the spatial and spectral structural characteristics of multi spectral image data

this paper present a concept of Multi Pattern Prediction : given a principle

any pixel in an image can be predicted by any prediction function selected from an alternative function set in order to decorrelate the image more efficiently ; at the same time

taking the advantage of spectral structural correlation

the spectral adjacent pixels are decorrelated by the same prediction function

so the additional cost of storage in Multi Pattern Prediction would be reduced smartly. We present the Minimum Entropy Principle as the theoretic principle to select the predition function. And we get an equivalent principle named as Maximal Frequency of Minimum-Error Standard. Experiments on TM images show that this method can decorrelate images much more efficiently and lead to higher compression ratios.