In SAR image most significant information of texture and region boundaries often appear in the middle and high frequency channels. Wavelet packet (WP) analysis decomposes the upper frequency bands as well as the lower ones. So it can achieve more accurate representation of the medium and high frequency information of SAR images. According to the " self-similarity " of the same subband, every subband is divided into subblocks of different importance and different size. Thus proper bit allocation and quantization can be implemented in subblocks. Meanwhile one can resort to image segmentation to achieve proper compression ratio in different image regions. The experimental results show the proposed method can suit the energy specialty of the SAR image. It can obtain higher PSNR and increase the image interpretability.