SAR speckle filtering is an important research subject. An ideal algorithm should smooth the Speckle without blurring texture and fine detail. A new speckle filtering method

using selective threshold in directionalet domain

is proposed. Image was decomposed and transformed in 12 directions by using multi-directional frames of directionalet. According to the characteristics of different length pixel sequences

the pixel sequences are divided into three groups and dealt with different threshold method respectively by Gaussian noise statistical distribution interval. To retain many fine details in the image

the median of high frequency coefficients variance is used as noise variance estimate value in every scale. A real SAR image is used in the experimentation. The result shows that the proposed method outperforms several classical speckle reduction algorithms in visual quality. ENL and ESI are enhanced by 97 and 0.15.