When the number of labeled samples is limited, Laplacian SVM needs as many as possible unlabeled samples to improve the performance of classification. However, when the number of unlabeled samples is large, the required time and space complexity would be unacceptable. In order to apply it to largescale classification problems like SAR image segmentation, a new method for SAR image segmentation named as improved Laplacian support vector machine algorithm (Improved Laplacian SVM) was proposed. Watershed algorithm was first used to decompose the original image into several small prototype blocks, and image features of each small prototype blocks were extracted as training samples. Then an improved Laplacian SVM algorithm was proposed to classify data sets. The proposed method was verified on three SAR images. The experiments show that the method not only improves the accuracy of segmentation but also greatly reduces the running time of Laplacian SVM algorithm for image segmentation.