基于WorldView-2数据的城市典型绿化树种分类

发布时间：2018-01-08 12:39

本文关键词：基于WorldView-2数据的城市典型绿化树种分类　出处：《内蒙古农业大学》2016年博士论文　论文类型：学位论文

【摘要】：通过遥感技术识别树种是尚未解决的难题之一,也是广大学者关注的焦点问题之一。目前基于高分辨率影像及辅助数据的树种分类已经取得了一定的成果,但依然存在许多诸如侧重影像信息维度窄、影像特征构建与筛选不科学、分类器休斯现象未解决等问题。本研究以呼和浩特市WorldView-2影像为数据源,经影像预处理,确定分类树种,构建影像高维光谱指数集合、纹理特征集合,基于最大似然的递归特征消除(MLC-RFE)选择重要变量,规避最大似然的休斯现象,获取树种分类的最优光谱指数子集、纹理特征子集。充分结合影像光谱波段、光谱指数、纹理等特征类型,利用最大似然对组合数据进行分类,以支持向量机的分类结果作为参照,实验结果取得了较好的分类精度。主要研究结果如下：(1)NDVI影像中蓝色屋顶、绿色塑胶操场2类地物与植被具有相似的NDVI特性,为城市植被提取造成干扰,但三者在WorldView-2八个波段中的光谱曲线存在较大差异,通过波谱角分类可实现三者完全分离,精准获取到城市植被部分的影像。(2)采用最大似然对针叶树、阔叶树与草类进行分类,利用8月份WorldView-2影像分类的总体精度为93.9871%, Kappa系数为0.9098,利用2月份QuickBird影像分类的总体精度为96.6667%, Kappa系数为0.9500,结果表明特殊时相数据源的选择更有利于针叶树、阔叶树、草类的识别。(3)基于WorldView-2光谱波段的树种分类中,最大似然对完整8波段分类的总体精度较传统4波段高10.7231%, Kappa系数高0.1253；支持向量机对完整8波段分类的总体精度较传统4波段高9.9183%,Kappa系数高0.1158,表明WorldView-2新增的海岸蓝色、黄、红边、近红外2波段在树种分类中具有重要的作用。(4)基于27维光谱指数的树种分类中,NDVI6、FDI2、NREB是树种分类中最重要的3个光谱指数；NDVI6、FDI2、NREB、ARVI、NDVI5、NDVI2、GRVI、NYR、NDVI1, IPVI、NPCI、R/RE、NDVI3、NIRNDVI、SAVI、NDVI7、NIR/GREEN、TA578、TA678是树种分类中最优光谱指数子集的19个成员；SL57、SL67、NDVI4、SL58、RVI、EVI、OSAVI、 SL56是导致最大似然发生休斯现象的8个光谱指数。(5)本研究新建的5个光谱指数SL57、SL67、SL58、TA578、TA678在MLC-RFE变量选择中,SL58在第5轮次中被消除,SL57、SL67在第7轮次中被消除,TA578、TA678在第8轮次中才被消除,在第7轮次消除结束后获得了最优光谱指数子集,所以TA578、TA678是最优光谱指数子集的成员,表明TA578、TA678在基于光谱指数的树种分类中具有重要的作用,同时也说明树种光谱曲线的面积指数优于斜率指数。(6)基于24个纹理特征的树种分类中,MEA-PC1、MEA-PC2、 EA-PC3是树种分类中最重要的3个纹理特征；MEA-PC1、MEA-PC2、MEA-PC3、 ENT-PC2、ENT-PC1、DIS-PC2、SM-PC1、VAR-PC2、HOM-PC3、COR-PC1、 OR-PC3、CON-PC2、CON-PC1、VAR-PC3、 DIS-PC1、ENT-PC3是树种分类中最优纹理特征子集的16个成员；HOM-PC2、SM-PC2、 CON-PC3、HOM-PC1、DIS-PC3、COR-PC2、VAR-PC1、SM-PC3是导致最大似然发生休斯现象的8个纹理特征。(7)27维光谱指数分类的总体精度为72.4616%, Kappa系数为0.6787,较最优光谱指数子集分类的总体精度(75.3962%)低2.9346%, Kappa系数(0.7126)低0.0339,表明在高维光谱指数分类中,最大似然存在着轻微的休斯现象；24个纹理特征分类的总体精度为40.5151%%, Kappa系数为0.3031,较最优纹理特征子集分类的总体精度(81.1664%)低40.6513%, Kappa系数(0.7799)低0.4768,表明在高维纹理特征分类中,最大似然存在着严重的休斯现象。(8)本研究中,支持向量机分类的最高总体精度为84.6335%, Kappa系数为0.8204,从所有的分类中可以看出它对数据维数的增加不敏感,可以有效挖掘各个特征的有用信息,分类性能较稳定。最大似然分类的最高总体精度为87.5310%,Kappa系数为0.8543,它对数据维数的增加较敏感,高维数据中会发生休斯现象,不能充分挖掘各个特征的有用信息,分类性能不稳定。本研究构建的MLC-RFE消除了对最大似然分类精度的提高具有抑制作用的特征,规避了最大似然的休斯现象,使其在高维特征分类中的分类性能得到极大地提高,取得比支持向量机更高的分类精度。(9)树种分类中,基于主成分的最高总体精度为63.9752%, Kappa系数0.5789;基于光谱波段的最高总体精度为74.0713%, Kappa系数0.6974；基于光谱指数的最高总体精度为75.3962%, Kappa系数0.7126；基于纹理特征的最高总体精度为81.1664%, Kappa系数0.7799；在光谱指数结合光谱波段中,最高总体精度为73.4274%, Kappa系数0.6900；在纹理结合光谱波段与主成分中,最高总体精度为86.3918%, Kappa系数0.8410；在纹理结合光谱指数与主成分中,最高总体精度为87.4319%, Kappa系数0.8532；在纹理结合光谱指数、光谱波段、主成分的混合特征中,最高总体精度为87.5310%, Kappa系数0.8543。除光谱指数结合光谱波段不能提高分类的总体精度与Kappa系数外,其余的特征组合类型均取得比单纯基于主成分、光谱波段、光谱指数、纹理特征要高的总体精度与Kappa系数,表明树种分类中有效结合各特征类型,可以取得更好的分类结果。
[Abstract]:One of the problems yet to be solved through remote sensing technology to identify species, is one of the focus of attention of scholars. At present, tree image classification and auxiliary data based on high resolution has achieved certain results, but there were still many problems such as focus on the image information dimension narrow, image feature construction and selection is not scientific, the classifier does not solve the problem of Hughes phenomenon etc. This study takes Hohhot WorldView-2 image as data source, through image preprocessing, determine the classification tree, build image high-dimensional spectral index set, texture feature set, recursive feature elimination based on the maximum likelihood (MLC-RFE) to select important variables, avoid the maximum likelihood of the Hughes phenomenon, the optimal spectral index subset obtaining species classification and the texture feature subset. Combined with spectral band, spectral index, texture feature type, using the maximum likelihood combination of data For classification, the support vector machine classification results as a reference, the experimental results achieved better classification accuracy. The main results are as follows: (1) NDVI image in the blue roof, green plastic playground 2 objects and vegetation with NDVI is similar to the characteristics of city vegetation extraction caused interference, but there is a big difference in spectral curve three in the eight WorldView-2 band, the spectral angle classification can achieve complete separation between the three, part of the city to obtain accurate vegetation image. (2) the maximum likelihood of conifer, broadleaf trees and grasses are classified, the overall classification accuracy of WorldView-2 image of August was 93.9871%, Kappa coefficient is 0.9098, the overall the classification accuracy of QuickBird image of February is 96.6667%, Kappa coefficient was 0.9500. The results show that the selection of more special phase data source for conifer, broadleaf trees, grasses identification based on W (3). Species classification of orldView-2 spectral bands, the overall accuracy of maximum likelihood classification than the traditional 8 full band 4 band high 10.7231%, high Kappa coefficient 0.1253; support vector machine to complete the overall accuracy of classification than the traditional 8 band 4 band high 9.9183%, high Kappa coefficient 0.1158, show new WorldView-2 coast blue, yellow, red the 2 side, near infrared band plays an important role in species classification. (4) based on tree species classification of 27 dimensional spectral index, NDVI6, FDI2, NREB are the 3 most important species in the spectral index; NDVI6, FDI2, NREB, ARVI, NDVI5, NDVI2, GRVI, NYR, NDVI1, IPVI NPCI, R/RE, NDVI3, NIRNDVI, SAVI, NDVI7, NIR/GREEN, TA578, TA678, 19 members of the species classification index of optimal spectral subset; SL57, SL67, NDVI4, SL58, RVI, EVI, OSAVI, SL56 is the result of 8 spectral index of maximum likelihood Hughes phenomenon (5) in this study. The new 5 A spectral index SL57, SL67, SL58, TA578, TA678 and MLC-RFE in the selection of variables, SL58 is eliminated, SL57 in the fifth inning, SL67 is eliminated, TA578 in the seventh inning, TA678 was eliminated in the eighth round, seventh rounds of elimination in the optimal spectral index subset, after the end of TA578, TA678 is the optimal subset of the members, the spectral index showed that TA578 and TA678 play an important role in species classification based on spectral index, but also shows the species spectral curve area index is better than the slope index. (6) based on tree species classification of 24 texture features, MEA-PC1, MEA-PC2, EA-PC3 is the 3 most important species in texture feature the classification of MEA-PC1, MEA-PC2, MEA-PC3; ENT-PC2, ENT-PC1, DIS-PC2, SM-PC1, VAR-PC2, HOM-PC3, COR-PC1, OR-PC3, CON-PC2, CON-PC1, VAR-PC3, DIS-PC1, ENT-PC3 is the 16 member optimal texture feature subset in the classification of species; HOM-PC2, SM-P C2, CON-PC3, HOM-PC1, DIS-PC3, COR-PC2, VAR-PC1, SM-PC3 are 8 texture features of the maximum likelihood Hughes phenomenon caused. (7) the overall accuracy of 27 dimensional spectral classification index is 72.4616%, Kappa coefficient was 0.6787, compared with the overall accuracy of the optimal spectral index subset classification (75.3962%) 2.9346%, Kappa (coefficient of 0.7126 0.0339) low, show that in high dimensional spectral index classification, maximum likelihood there is a slight Hughes phenomenon; the overall accuracy of the 24 texture feature classification 40.5151%%, Kappa coefficient is 0.3031, the overall accuracy compared with the optimal texture feature subset classification (81.1664%) 40.6513%, Kappa coefficient (0.7799) shows that in the 0.4768 low, high dimensional texture classification, maximum likelihood is a serious phenomenon of Hughes. (8) in this study, the highest overall accuracy of support vector machine classification is 84.6335%, Kappa coefficient is 0.8204, from which all can be seen in the classification of it Not sensitive to the increase of data dimensionality, useful information can effectively tap the various features, the 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 data dimension increases, Hughes phenomenon in the high dimensional data, useful information can not fully excavate various features, classification the performance is not stable. The construction of the MLC-RFE elimination characteristics have inhibitory effect on the maximum likelihood classification to improve the accuracy of the maximum likelihood, to avoid the Hughes phenomenon, the classification in high dimensional feature classification can be greatly improved, higher classification accuracy than support vector machine (9 species). In classification, the highest overall accuracy based on principal component was 63.9752%, Kappa coefficient is 0.5789; the highest overall accuracy based on band 74.0713%, Kappa coefficient is 0.6974; the highest total based on spectral index Body precision was 75.3962%, Kappa coefficient is 0.7126; the highest overall accuracy based on texture feature is 81.1664%, Kappa coefficient is 0.7799; in the spectral index and spectral bands, the highest overall accuracy is 73.4274%, Kappa coefficient is 0.6900; in texture and spectral bands with principal components, the highest overall accuracy is 86.3918%, Kappa coefficient is 0.8410; in the combination of texture the spectral index and principal components, the highest overall accuracy is 87.4319%, Kappa coefficient is 0.8532; combined with the spectral index, the texture spectrum band, mixing characteristics of principal components, the highest overall accuracy is 87.5310%, Kappa coefficient 0.8543. in spectral index and spectral bands can improve the overall accuracy and Kappa coefficient classification, types and characteristics of the rest of the combination are better than only based on principal component spectral bands, spectral index, texture features to the overall accuracy and Kappa coefficient is high, the tree species classification Better classification results can be obtained by effectively combining each characteristic type.

【学位授予单位】：内蒙古农业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：S731.2;S771.8

【参考文献】