Very High Resolution Remote Imagery Forest Classification Based On Moving Window's Fourier Transform

The high spatial resolution remote sensing image has plentiful texture information,which may facilitate the recognition of land covers,and this is the reason why it is one of the hot spots in remote sensing technology application nowadays.It is well known that Fourier transform is a powerful tool for processing and analyzing texture information.In this paper,Fourier transform was used to extract the texture features of the forest as the entry point to classify the forest.The aim was to explore ways to improve the classification accuracy of tree species when only the texture features participate in the classification.Fourier transform based on window scale can effectively avoid the defect that the Fourier transform can not highlight local information.In this paper,Worldview panchromatic band data of 0.5 m spacial resolution image was used as data source,and One-dimensional(1D)and Two-dimensional(2D)Fourier transforms are performed on the basis of the moving window to extract One-dimensional and Two-dimensiona texture feature vectors for the 9 land covers in the study area.different classification methods were used to classify forests based on eigenvectors to find appropriate moving window.A total of 21 square windows with odd side lengths from 3×3 to 43×43 were tested.Texture features generated by each side window all using Fisher Discriminant Method(FDM),Random Forest(RF),Support Vector Machine(SVM),Included Angle Cosine(IAC)and Correlation Coefficient(CC),etc.5 classification methods were used to perform the classification,and compute classification accuracy.The study of forest classification is mainly based on two levels:(1)Broad-leaved forest,Phyllostachys praecox forest,Pinus massoniana forest,mao bamboo forest,chinese fir forest,mixed conifer-broadleaf forest,road,farmland and water etc.9 class(full class)were used to extract the distribution information of each tree species.(2)Only 6 tree species are classified(excluding 3 non-forest),and the best classification window is determined based on the accuracy;based on the best classification window,the tree species are combined into one class—forest,and then the three other non-forest class;extract forest distribution information;remove non-forest area,and finally extract the distribution information of each tree species.Conclusion: Based on Fourier transform of moving window to construct high-resolution remote sensing image features texture features,in a suitable window size,the forest may be extracted with high precision and the forest tree species are more accurately classified.1?Full-class classification: For 1D features and 2D features under the best window Random Forest is superior to the other 4 classification methods;while the 2D features are superior to the 1D features,the total accuracy is 88.69%,and Kappa coefficient is 0.863 4.2?Forest classification: For 1D features and 2D features Fisher discriminant performance under the best window is better than the other 4 classification methods;while once again 2D features classification result is better than the 1D one,the total accuracy is 84.86%,and Kappa coefficient is 0.814 9.3?Forest and non-forest Classification: Fisher discriminant method is also the best under the best window of forest;and the 1D features classification result is better than the 2D features,the total accuracy is 99.91%,and Kappa coefficient is 0.998 2.