| Forest vegetation is the main body of forest resource.Accurate identification of forest vegetation types can lay the foundation for the research and utilization of forest resources.With the development of remote sensing technology and the emergence of high spatial resolution data tree classification has been feasible.However,traditional optical remote sensing can only describe the horizontal pattern of ground objects,so it was very difficult to identify individual tree species.Light detection and ranging(LiDAR)has great potential and advantages in the classification of individual tree species.This study takes two 100m× 100m square plots of the Maoershan Experimental Forest Farm of Northeast Forestry University as the research object.Firstly,the LiDAR data was preprocessed to obtain the canopy height model(CHM),and then the original CHM was optimized;Then,the individual tree canopy segmentation was performed by region-based hierarchical cross-section analysis(RHCSA),and then the accuracy test was performed to obtain a one-to-one matching canopy.The multi-spectral remote sensing CCD(charge coupled device)image and the airborne LiDAR were used as data sources.Based on the CCD image,21 features of spectral features and textures are extracted respectively and based on the airborne LiDAR data,34 features of height,intensity and canopy size were extracted.Then the random forest method was used to select features and to obtain optimized features;Two kinds of non-parametric classifiers,random forest(RF)and support vector machine(SVM)were used to classification.Combined with different data sources and different features,12 classification schemes were used to identify individual tree species,and the overall accuracy(OA),user accuracy(UA)and producer accuracy(PA)were used for accuracy evaluation and comparing the results of classification.This study draws the following conculusions.(1)Using random forest for feature selection when only LiDAR data was used,the number of features was reduced from 34 to 12;when only multi-spectral CCD image was used,the number of features was reduced from 21 features to 11;when combining LiDAR data and multi-spectral CCD image,the number of features was reduced from 55 to 11.The result showed that no matter which data was used for the experiment,the results after feature seletion by random forest were better than those without feature seletion by random forest.In addition,features were selected based on random forest internal calculation,when using the support vector machine as a classifier,the classification accuracy can still be improved.As a result,using random forest to select features can improve classification accuracy,and the results were reliable,and can be applied to different data sources and multiple classifiers.(2)Random forest can derive the importance of features by calculating the mean derease accuracy.Experiments showed that spectral features have the greatest contribution to individual tree species classification,regardless of whether LiDAR data was used or not.Texture features,height and intensity features extracted by LiDAR data also contributed significantly to classification.Furthermore,based on the results of the two plots,the intensity features were more stable than the height features.(3)The overall accuracy of the classification using multispectral CCD images and LiDAR data was significantly improved compared to the use of a single data source,and the overall accuracy was increased by 20.8%and 7.8%,respectively,compared to the use of multispectral CCD images and LiDAR data,respectively.Kappa coefficient has also improved.So,combining multi-source remote sensing data can improve the accuracy of individual tree species classification. |
PDF Full Text Request