Abstract:
WorldView-1 panchromatic band data of 0.5 m spatial resolution image of east Lin’an district of Hangzhou, Zhejiang province in April 2008 was used as data source, and two-dimensional Fourier transform based on moving windows was carried out to produce a texture feature vector, and different classification methods were used to classify forests based on feature vectors to find an appropriate moving window size.A total of 21 square windows with odd side lengths from 3×3 to 43×43 were tested.Texture features generated by each side were classified by Fisher discriminant, random forest(RF), support vector machine(SVM), included angle cosine(IAC) and correlation coefficient(CC), and classification accuracy wascomputed.Based on the forest classification accuracy, the optimal windows corresponding to the five classification methods were 41×41, 41×41, 23×23, 39×39, and 39×39.Under the optimal window, five classification methods had an accuracyof 95% to distinguish forests from non-forests, and the order of the total classification accuracy was as follows: Fisher discriminant> RF > SVM > CC > IAC, the Fisher's discriminant method hada total accuracy and a Kappa coefficient of 99.81% and 0.9963.Forest tree species werefurther classified, the total accuracy was Fisher discriminant> RF > SVM > CC> IAC, Fisher discriminant method hada total accuracy and Kappa coefficient of 84.86% and 0.8149.The results showed that under the optimal window, Fisher discriminant method was the best classification method.