Dimensionality reduction approach was proposed based on anisotropic kernel diffusion map to extract the features of poplar leaves, in which the kernel parameters were adjusted adaptively. In order to improve the accuracy and efficiency, singularity points were removed and features normalization method was employed to obtain the robust features. The maximum margin criterion method was utilized to obtain anisotropic kernel parameter by gradient descent method. The results show that the anisotropic kernel diffusion map has good performance on efficiency for poplar leaves compared with LE, LTSA and PCA. The comparisons of classification experiments have been conducted, by using SVM (support vector machine) classifier to recognize the water shortage of poplar leaves, and the results validate the accuracy and stability of the proposed method.