Typical Urban Greening Tree Species Classification Based On WorldView-2

Using remote sensing technology to identify tree species is one of the problems that not be solved, is also one of the focal issues concerned by the scholars. Currently, tree species classification has made certain achievements based on high resolution images and auxiliary data, but there are still exist many problems, such as focusing image information dimensions narrow, image features construction and selection not scientific and classifier Hughes phenomenon unresolved. In this study, take Huhhot WorldView-2 image as data source, after image preprocessing, determine the classification tree species, build image high dimensional spectral indices set and texture features set, based on maximum likelihood with recursive feature elimination choose important variables, to avoid Hughes phenomenon of the maximum likelihood classification, get the best spectral index and texture feature subset for tree species classification. Fully integrated image spectral band, spectral index, texture and the type of other characteristics, tree species classification of mixed data using maximum likelihood, the support vector machine classification results are used as reference, the experimental results have achieved good classification accuracy. The main results are as follows:(1) The blue roof, green plastic playground two objects in NDVI are similar to vegetation, cause interference for extraction urban vegetation, but they are have a big difference between spectral curves in WorldView-2’s eight bands, through the spectral angle classification can be completely separated them, urban vegetation images can be effectively extracted.(2) Distinguish conifer tree, broadleaf tree and grass using maximum likelihood, the overall accuracy is 93.9871%, Kappa coefficient is 0.9098 used WorldView-2 image of August, the overall accuracy is 96.6667%, the Kappa coefficient is 0.9500 used QuickBird image of February,the results showed that the special phase data source select is more advantageous to identification coniferous tree, broadleaf tree and grass.(3) Tree species classification based on WorldView-2 spectral bands, used maximum likelihood, the overall accuracy of complete 8 bands is higher 10.7231% than the traditional 4 bands, Kappa coefficient is also higher 0.1253; used support vector machine, the overall accuracy of complete 8 bands is higher 9.9183% than the traditional 4 bands, Kappa coefficient is also higher 0.1158, Showed that WorldView-2 new coastal blue, yellow, red edge, near infrared band 2 plays an important role in the classification of tree species.(4) Tree species classification based on 27-dimensional spectral indices, NDVI6、 FDI2> NREB are the most three important spectral indices in tree species classification; NDVI6、FDI2、NREB、ARVI、NDVI5、NDVI2、GRVI、NYR、NDVI1、 IPVI、NPC、R/RE、NDVI3、NIRNDV、SAVI、NDVI7、NIR/GREE、TA578、 TA678 are the nineteen member of the optimal spectral index subset in tree species classification; SL57、SL67、NDVI4、SL58、RVI、EVI、OSAVI、SL56 are eight spectral indices caused maximum likelihood occurred Hughes phenomenon.(5) In this study, the new five spectral indices SL57、SL67、SL58、TA578、TA678 in MLC-RFE variable selection, SL58 is eliminated in fifth rounds, SL57% SL67 are eliminated in seventh rounds, TA578、TA678 are eliminated in eighth rounds, the seventh rounds elimination ended, optimal spectral index subsets is obtained, so TA578、TA678 are the member of the optimal subset of the spectral index, showed that TA578、TA678 play an important role in the classification of tree species based on spectral index, also showed that the tree species spectral curve of area index is better than the slope index.(6) Tree species classification based on the 24 texture features, MEA-PC1、 MEA-PC2、MEA-PC3 are the most three important texture features in tree species classification; MEA-PC1、MEA-PC2、MEA-PC3. ENT-PC2. ENT-PC1、DIS-PC2. SM-PC1、VAR-PC2、HOM-PC3、COR-PC1、COR-PC3、CON-PC2、CON-PC1、 VAR-PC3、DIS-PC1、ENT-PC3 are the sixteen member of the optimal texture feature subset in tree species classification; HOM-PC2、SM-PC2、CON-PC3、HOM-PC1、 DIS-PC3、COR-PC2. VAR-PC1、SM-PC3 are eight texture features caused maximum likelihood occurred Hughes phenomenon.(7) The overall accuracy of 27 dimensional spectral indices in tree species classification is 72.4616%, Kappa coefficient is 0.6787, compared with the optimal spectral index subset classification, the overall accuracy (75.3962%) lower 2.9346%, Kappa coefficient (0.7126) lower 0.0339, showed that in high dimensional spectral indices classification, maximum likelihood exist a slight Hughes phenomenon; The overall accuracy of 24 dimensional textural features in tree species classification is 40.5151%%, Kappa coefficient is 0.3031, compared with the optimal textural feature subset classification, the overall accuracy (81.1664%) lower 40.6513%, Kappa coefficient (0.7799) lower 0.4768, showed that in high dimensional textural features classification, maximum likelihood exist a serious Hughes phenomenon.(8) In this study, the highest overall accuracy of support vector machine classification is 84.6335%, Kappa coefficient is 0.8204, it can be seen in the classification of the data, it is not sensitive to the increase of dimensions, can be effectively excavate the useful information of the various features, classification performance is stable. The highest overall accuracy of maximum likelihood classification is 87.5310%, Kappa coefficient is 0.8543, it is sensitive to the increase of the data dimensions, in high dimensional data the Hughes phenomenon will happen, useful information can not be fully excavate from each feature, the classification performance is not stable. The MLC-RFE constructed by this study, it eliminates the features that obstacle maximum likelihood increase accuracy, avoid the Hughes phenomenon in maximum likelihood classification, the classification performance has been greatly improved in high dimensional feature set, achieve higher classification accuracy than support vector machine.(9) In tree species classification, based on principal components; spectral bands; spectral indices; texture features; spectral indices combined with spectral bands; texture features combined spectral bands and the principal components; ture features combined spectral indices and the principal components; the mixing features of the texture features, spectral bands,spectral indices and the principal components, the highest overall accuracy (Kappa coefficient) are 63.9752% (0.5789); 74.0713% (0.6974); 75.3962% (0.7126); 81.1664% (0.7799); 73.4274%(0.6900); 86.3918% (0.8410); 87.4319%(0.8532); 87.5310% (0.8543), respectively. In addition to classification of spectral bands combining spectral indices does not improve the overall accuracy and Kappa coefficient, the type of others feature combination have a higher overall accuracy and Kappa coefficient than only based on principal components, spectral bands, spectral indices, texture features. Showed that in tree species classification, effectively combining various features, it can be obtain a better classification results.