关键词: 组合分类器, 高分五号, 高分六号, 树种识别, 面向对象, 特征选择

Abstract:

Objective: In view of the problems that the recognition accuracy of remote sensing tree species under the complex stratospheric structure and high forest density of subtropical forest is not high, and that the performances of different classification algorithms to the recognition of different tree species are various, the method of multi-classor combined tree species recognition of high spectral resolution combined with high spatial resolution data was investigated to promote the in-depth application of multi-source data in the field of forest resource survey and monitoring. Method: Taken Huangfu Mountain National Forest Park as research area, and assisted with data of GF-6 PMS and GF-5 AHSI, digital elevation model(DEM) and forest resource survey data, the effective tree species recognition algorithm of multi-classor adaptive for object multi-source data, which is under the conditions of subtropical natural secondary forests, complex coronary structures and high forest density, was built in this study. First of all, the GF-6 PMS data was divided according to object-oriented multi-scale and combined with the field survey data by using the graph cutting algorithm(GC) to select the sample. Then, it was carried out to use GF-5 AHSI data to extract 26 vegetation index features(VIF), to use GF-6 PMS data to extract 128 parties of each band texture features(TEF), and combined with 304 bands of GF-5 AHSI with removal of bad band as spectral features(SF) and 3 DEM-built terrain factors as terrain features(TRF, factor features according to various inter-class separability were selected and 10 recognition schemes based on characteristic selection results were built. Finally, the weight adaptive voting combination classifier(WACC) using near-category(KNN), support vector machine(SVM), Bayesian classification(Bayesian), classification regression tree(CART), and random forest(RF) was constructed to recognize tree species based on 10 classification schemes and to verify the accuracy of tree species identification. Result: Under the linear discriminant analysis model, the discriminant accuracy was identified to tend to be stable when the factor of spectral characteristics, texture features and vegetation index characteristic were increased to 28, 12 and 10, respectively. The spectral characteristic factors with better recognition ability for the tree species were mainly concentrated in the red light and near-infrared bands, the texture feature factors mainly concentrated on mean, entropy and angular second-order moment, while the vegetation index features were mainly concentrated on the index of characterization greenness, carbon attenuation, and coronary aquifer aquifers. The overall accuracy of tree species recognition in the weight adaptive classifier combination algorithm was 87.51%, and the Kappa coefficient of 0.854 was better than that of the single classifier algorithm. The overall accuracy obtained by different characteristic factor schemes under each classification algorithm was as follows: SF+TEF+VIF > SF+TEF+VIF+TRF > SF+TEF+TRF > SF+VIF > SF+VIF+TRF > SF+TEF > SF > SF+TRF > VIF > TEF. Conclusion: he weight adaptive classifier combination algorithm under the combination of SF-TEF-VIF features could effectively identify tree species in subtropical natural secondary forests, and enjoy good recognition accuracy and credibility, and the data of GF-5 AHIS and GF-6 PMS would present a good application potential and prospect in the fields of forest resource survey and monitoring.

Key words: combinatorial classifier, GF-5, GF-6, tree species identification, object oriented, feature selection