Fisher Discriminant Based Hierarchical Method for Forest Remotely Sensed Data Classification

Fisher discriminant analysis has been applied to obtain the confusion matrix, which was used to calculate the classification evaluation index, and hierarchical classification tree model was built in accordance with these index. Thus, Fisher hierarchical classification method based on Fisher discriminant analysis was proposed to forest remote sensed images. A part of sample data from forest resources inventory and Landsat8 remote sensed image data of Hangzhou city were used to test this method, furthermore the results was compared to those of maximum likelihood classification method, support vector machine and random forest classification method. Results showed that the total classification accuracy of Fisher based hierarchical classification is 79.45%, which is higher than that of maximum likelihood classification, support vector machine and random forest by 12.33%, 21.00% and 10.50%, respectively. Kappa coefficient was 0.756 8, which meant it was higher than those from maximum likelihood method, support vector machine and random forest whose values were 0.145 5, 0.256 4 and 0.126 4, respectively. In the hierarchical classification model based on Fisher discriminant analysis, the gradation of hierarchical classification tree are construction land→water→farmland→bamboo grove→broad-leaved forest→coniferous forest→conifer-broadleaf forest. In the hierarchical classification based on Fisher discrimination, a model can only distinguish 1 category.