In this paper, we are interested in designing lossless coders for a class of multispectral images. More precisely, we consider quincunx sampled images such as those encountered in new generation satellite imaging systems. Our approach exploits both the spatial and the spectral correlations existing in these images by applying block-adaptive predictors. A clustering algorithm is applied to the spectral bands of the original image in order to compute optimal predictors within each group of blocks. Simulation tests carried out on natural multicomponent images show that the proposed adaptive interband differential prediction outperforms state-of-art lossless coders.