Traditional learning-based boundary extraction algorithms classify each pixel edge separately and then get boundaries from the local decisions of a classifier. However, we propose a supervised learning method for boundary extraction by using edgelets as boundary elements. First, we extract edgelets by clustering probabilities of boundary. Second, we use features of edgelets to train a classifier that determines whether an edgelet belongs to a boundary. The classifier is trained by utilizing edgelet features, including local appearance, multiscale features, and global scene features such as saliency maps. Finally, we use the classifier to decide the probability that the edgelet belongs to the boundary. The experimental results in the Berkeley Segmentation Dataset demonstrate that our algorithm can improve the performance of boundary extraction.