Dominant Tree Species Classification Using GF-2 Images Based On Seasonal Characteristics

Remote sensing technology is widely used in forest monitoring.Accurately classifying different vegetation types and tree species can improve the accuracy of each monitoring indicator.In recent years,China has launched a series of domestic satellites such as "GF" and "ZY",which has enriched the domestic remote sensing satellite data and brought great convenience to the development of domestic remote sensing work.When classifying vegetation and tree species,if the seasonal characteristies of different vegetation types in different seasons are fully considered,the classification accuracy can be further improved.Therefore,GF-2 satellite images with both high spatial resolution and high temporal resolution are used as data in this study.This study takes Beijing Jiufeng Forest Park as the research area,and the vegetation types and tree species in this research area are effectively classified through pre-processing,segmentation experiment aimed to select the optimal segmentation parameters,feature analysis and the combination of artificial rules and classifiers.The results provide a basis for the development of related forestry work.The main findings are as follows:(1)The canopy spectral reflectance curves of different tree species(Pinus tabulaeformis,Lateral berlin,Quercus variabilis)in different seasons are measured,and the spectral reflectance of three tree species is different in green band and near infrared band.;the deciduous tree species(Quercus variabilis)have obvious changes in the spectral reflectance curve during leaf discoloration period,which shows that this period is conducive to tree species identification;the changes of spectral reflectance curve are similar to the changing trend of GF-2 images,proving that GF-2 images can reflect the seasonal changes of tree species.(2)Through the selection of reasonable range of initial values of segmentation scale,heterogeneity factor experiment,ESP segmentation scale evaluation tool analysis and spectral difference segmentation,the optimal segmentation parameters suitable for vegetation classification are selected.(3)Using the seasonal characteristics of the five images,feature selection and rule establishment are carried out.Three hierarchical classification methods using multi-temporal images combining artificial rules based on feature analysis and classifier were established,comparing with two hierarchical classification methods using single-temporal image.The classification overall accuracy results are as follows:the accuracy of multi-temporal hierarchical random forest method is highest,which is 91.9%;the accuracy of multi-temporal hierarchical CART decision tree method and multi-temporal hierarchical nearest neighbor method takes second place,which is 89.9%and 88.2%respectively;the accuracy of single-temporal hierarchical random forest method and single-temporal hierarchical CART decision tree method is slightly worse,which is 71.4%and 68%respectively.The results show that multi-temporal images can effectively improve the overall accuracy of vegetation and tree species classification.