Clutter statistical modeling for synthetic aperture radar(SAR) measurements in nonhomogeneous clutter environments is a complex and challenging task. In order to overcome the dependence of the Parzen window estimation in the non-parametric modeling algorithm for setting the window width and picking the optimal kernel function, a new K nearest neighbors optimized estimation algorithm is proposed. The proposed algorithm solves the problem of that the estimation is inaccurate or cannot be estimated due to the fixed neighbor number. Meanwhile, the algorithm neither requires any prior knowledge nor needs to set the window width and select the optimal kernel function. Compared with Parzen window estimation, K distribution and $G^0$ distribution, the proposed K-nearest neighbor optimization estimation algorithm can accurately model single-peak, multi-peak and even irregular image data, which is better than K distribution and $G^0$ distribution. And the processing of the first and last data of the image is better than the Parzen window estimation. The experimental results verify the accuracy, robustness and simplicity of the proposed method for background clutter modeling of SAR images, and the effectiveness of global constant false alarm rate target detection.