This paper presents a new wavelet transform coding algorithm for complex Synthetic Aperture Radar (SAR) image compression. It uses a improved recursive set-partitioning procedure to compress the large blocks which are abundant in zero bits, and uses entropy coding based on adaptive context prediction model to compress the small blocks which contains non-zero bits. The experiment results showed that the new algorithm gained better compression efficiency than SPECK. We also compared with common complex SAR image compression algorithm in references using professional parameters in SAR, such as Means Phase Errors (MPE), complex spatial correlation efficient, the new algorithm gets better performances too.