Abstract: Locally linear embedding (LLE) is a classical and effective manifold learning method, which can project the original samples into a low dimensional space by preserving the least reconstructed weights among the neighbor points. But LLE is very sensitive to noisy points and outliers. In order to improve the robust of LLE, a weighted LLE (WLLE) algorithm was proposed in this paper. The importance score of each point was obtained by the heat kernel function. The importance scores were then added into the cost function of WLLE. The undesirable effect resulted by the noisy points and outliers on the embedding result can be largely reduced. The experimental results based on the real-world plant leaf databases show the effectiveness and feasible of the proposed method.