A new supervised weighted LLE method based on the Fisher projection was proposed. This method utilized the Fisher projection distance to replace the sample’s geodesic distance, and the importance score of each sample was obtained based on this distance, then the importance scores were added into the cost function of LLE. This method can overcome the disadvantage of traditional LLE, an unsupervised learning algorithm which cannot solve the classification problem very well, and can exploit the category information better and reduce the influence of noise points at the same time. The experimental results based on the real-world plant leaf databases show its mean accuracy of recognition is up to 92.36%.