Abstract: [Objective] Object-based classification method provides a new way for classifying remote sensing image of high spatial resolution because it can synthetically use spectral feature, geometric feature, texture feature, and so on. However, increase in the number of features will lead to the emergence of the "dimension disaster", complicate operation, and decline in speed. It also leads to a decrease in classification accuracy under the limited training samples. In order to solve the problem in the feature space selection during object-based classification, an optimal features selection method based on ReliefF was proposed in this study.[Method] The method gave a weight to each feature according to the separability of features of training samples, and selected the features which are important to samples classification through analyzing the correlation between features. Taking Shanchuan town in Anji County in Zhejiang Province as the study area, 370 object samples for eight classes were selected. A total of 42 object features were selected, including mean and standard deviation of normalized difference vegetation index(NDVI), each band of SPOT5 image and their gray level co-occurrence matrix textures(GLCM). Based on the optimal features selected from the 42 object features using the ReliefF algorithm, nearest neighbor algorithm in object-based classification method was used to extract the distribution of the bamboo forest in the study area. Moso bamboo forest information extracted by object-based classification based on optimal features was compared with the results from classification and regression tree(CART) decision tree algorithm in object-based classification method, under the same segmentation parameters and training samples.[Result] 1) By using the ReliefF algorithm, the classification accuracy of bamboo forest samples has been greatly improved. The accuracy of moso bamboo samples was increased from 68% to 88%. Mean object value of red band as well as green band, mean component of GLCM in red band, entropy component of GLCM in red band, and mean object value of NDVI were the five optimal object features. The user's and producer's accuracies achieved 97% and 95%, respectively; 2) Both the user's and producer's accuracies of moso bamboo forest were lower when using CART decision tree than those using nearest neighbour(NN)classification based on optimal features, and the main reason was attributed to the serious confusion among moso bamboo forest, deciduous as well as conifer. [Conclusion] ReliefF algorithm focus on the discrimination ability of features, and using the features selected by ReliefF algorithm were prior to other related researches, which gave insight into the object-based classification of forest resources in the remotely sensed technique.

Key words: moso bamboo forest, ReliefF algorithm, optimal features, object-based, SPOT5 imagery