A novel error compensated prediction tree method used for lossless compression of multispectral imagery is proposed in this paper. This method combines prediction tree and adaptive linear prediction techniques

so the spatial redundancy

spectral statistical redundancy and spectral structural redundancy in multispectral imagery are all exploited. For each pixel of the imagery

the prediction tree based on interband structural similarity is used to remove the spatial redundancy and a prediction error is created. Then this error is compensated by an interband linear adaptive predictor

which is constructed according to the spectral statistical redundancy. Our method removes the spectral structural and statistic redundancies as well as the spatial redundancy

so the better compression results can be obtained. The construction of adaptive predictor for each pixel will introduce prohibitive computational complexity to the algorithm. Therefore a modified algorithm based on local stability of imagery is also designed to reduce the computational complexity. The experimental results from practical multispectral images have shown that our method is better than the original prediction tree one.