Abstract：In order to explore the effects of training sample numbers on the classification accuracy of remote sensing image in wetland, the object-oriented classification was put forward based on the statistical theory. GF-2 images were used as data source and the basis of constructing the standard training samples. The applicability and effectiveness of KNN algorithms and SVM algorithms ( i. e. linear functions, polynomial functions and radial basis functions) in wetland classification were analyzed. The minimum number of training samples was obtained by classified statistics and analyzing the extraction of wetland elements. The method was used in analysis of Tianehu park of Sanmenxia city, Henan province. Training samples from threefold to tenfold of bands was extracted. For each sample, 3 times of experiments were carried out and 96 classification maps were obtained. The results show that: when the samples are sixfold number of bands, the overall classification accuracy is 85% and kappa coefficient is 0. 8, which meets the accuracy requirement for classification. Meanwhile, the number of training samples based on the object-oriented is less than that of pixel-based supervised classification. The SVM classification algorithm is more accurate than KNN classification algorithm for wetland classification.

Keyword：training samples;wetland;accuracy of remote sensing classification;object-oriented;