An improved minimum cut model is presented considering that the minimum cut criteria favors cutting small sets of isolated nodes,then equivalence relation between the improved minimum cut model and weighted kernel k-means is researched,and the influence of different similarity functions on the results of segmentation are also analysed. And based on these,a multiscale image segmentation method based on graph weighted kernel k-means is proposed,this method avoids calculating graph spectral,which is a key step when using graph cut model to segment images,also,it avoids selecting kernel matrix,which is important to the weighted kernel k-means,finally it realizes multiscale image segmentation. The analysis of anti-noise and experimental results on a number of optical images show the effectiveness of this method.