Three-dimensional block-matching (BM3D) algorithm can effectively suppress the noise in stationary signal. However, it is not feasible for the speckle noise in synthetic aperture radar (SAR) image with random characteristics due to the single 3D transform threshold and the local neighborhood for searching similar blocks. We propose a BM3D algorithm based on K-Mean clustering for SAR image denoising. First, we calculate the feature vector according to the mean, variance, and poor value, and estimate noise variance of each image block. The adaptive 3D transform threshold will be determined through the estimated noise variance. Second, we can find similar image blocks of reference image block in the corresponding class of image blocks, and can find global similar image blocks quickly. The experiments demonstrate that the proposed algorithm achieves better visual effect and and higher peak signal to noise ratio than the BM3D algorithm and non-local mean algorithm.