Abstract: Cucumber leaf images are collected during a nonstandard environment and color, shape, texture and surf characteristics are extracted by interactive segmentation. In order to organize the high dimensional characteristics of cucumber leaf image, the method of approximate nearest neighbor detection algorithm based on a Kd-tree is proposed and overcomes the inefficiency of the high dimensionality vector detection. Subspace data structure and data variance characteristics vector of cucumber disease image is calculated by the Kd-tree recursively generated and a number of pixel values closer to the adjacent area are detected because analyze connecting regional, so a nearest neighbor chain is constituted which can detect the diseases of image. Using cucumber leaves, stems, roots, fruits, seedling diseases image as experimental data sets, experimental results show that the nearest chain disease detection method based on Kd-tree is higher image precision and recall ratio.