EN中文

Journal of Geo-information Science >

2018 , Vol. 20 >Issue 12: 1777 - 1786

DOI: https://doi.org/10.12082/dqxxkx.2018.180310

Orginal Article

The Classification of Urban Greening Tree Species Based on Feature Selection of Random Forest

  • WEN Xiaole , 1, 2, 3 ,
  • ZHONG Ao 1, 2 ,
  • HU Xiujuan , 1, 2, 3, *
Expand
  • 1. College of Environment and Resources,Fuzhou University, Fuzhou 350116, China
  • 2. Institute of remote Sensing Information Engineering, Fuzhou University, Fuzhou 350116, China
  • 3. Fujian Provincial Key Laboratory of Remote Sensing of Soil Erosion and Disaster Protection, Fuzhou University, Fuzhou 350116, China
*Corresponding author: HU Xiujuan, E-mail:

Received date: 2018-07-03

  Online published: 2018-12-20

Copyright

《地球信息科学学报》编辑部 所有
Fold

Abstract

Since Urban forests played important roles in improving air, water and land quality, absorbing and mitigating carbon dioxide and many pollutants, mitigating urban heat island and reducing storm water runoff, its monitoring is a major issue for urban planners. It is of great significance to obtain the tree species timely and precisely in urban planning and green space management. At present, urban forest tree species mapping has benefitted from advances in remote sensing techniques. Using an object-oriented method combing spectral, textural, indicial and geometric features from high-resolution WorldView-2 imagery, this paper aimed to carry out the classification of seven main tree species in Fuzhou university, including Banyan (Ficus microcarpa), Mango(Mangifera indica L.), Camphor tree (Cinnamomum camphora), Bishop wood (Bischofia polycarpa), Chinese orchid tree(Bauhinia purpurea L.), Weeping fig (Ficus benjamina L.), and Kapok tree (Bombax malabaricum DC.). A random forest method was employed to determine the feature selection in this study. When eliminating 20 percent of the total features, the in situ validation results showed that the overall accuracy reached a highest value of 74.95% with Kappa coefficient of 0.67 when using 34 features for classification, which including 15 spectral features, 6 textural features, 8 indicial features and 5 geometric features, and the feature of mean spectral was the most significant, however, the standard deviation of each band is less important. The results also revealed that the feature selection of random forest could reduce or avoid the data redundancy and Hughes phenomenon, and thus could improve the classification accuracy of same type tree species. Moreover, the four additional bands of WorldView-2 imagery, especially the yellow and red edge band, and their composite indexes showed a higher importance in classification, which also indicates that these bands have great application prospects in vegetation remote sensing, especially in tree species classification.

Key words： WorldView-2; object-oriented; random forest; feature selection; tree species classification

Cite this article

WEN Xiaole , ZHONG Ao , HU Xiujuan . The Classification of Urban Greening Tree Species Based on Feature Selection of Random Forest[J]. Journal of Geo-information Science, 2018 , 20(12) : 1777 -1786 . DOI: 10.12082/dqxxkx.2018.180310

1 引言

城市绿化在减缓城市热岛效应及提高城市环境效益方面作用突出,其中树种多样性也是描述生态环境质量的重要参数之一[1]。随着对城市系统认识的逐步深入,人们越来越意识到树种分类信息对研究城市气候、制定城市规划具有不可或缺的作用。树种信息也为城市的规划管理者提供了监测、控制与评估城市绿化的主要依据[2,3]。及时、准确地获取城市树种信息对于改善城市环境、推进城市可持续发展至关重要[4]。
目前获取准确的树种信息有一定的难度,现有的树种调查方法大致分为实地调查和遥感解译 2种。基于实地调查的树种分类只适用于较小的区域范围,需要投入大量的人力物力且耗时巨大[5]。近年来,遥感解译识别树种信息正成为研究热点,遥感数据源多为多光谱、高光谱、LiDAR以及SAR等。其中,有部分学者利用雷达以及高光谱数据进行树种分类的研究[6,7,8],但成本高昂以及应用的地域局限性,限制了其应用的前景。多光谱数据（如IKONOS及Sentinel-2A等）都在树种分类研究中得到了一定的应用[9,10],但由于IKONOS波段数量较少、Sentinel-2A空间分辨率较低等原因,在城市树种分类的研究中也有一定的局限性。而WorldView-2数据在保持高空间分辨率的同时加入了4个新增波段,在满足空间分辨率需求的同时也满足光谱分辨率的需求,因此本文选用该影像进行研究。
现有的树种分类研究多针对农田以及森林地区,而针对城市地区的树种分类研究相对较少。由于景观美化的需求以及植物配置多样化的原因,城市绿化是复杂多变的,相对较高的局部空间变化为遥感自动分类研究带来了一定的困难[11]。仅依靠光谱曲线无法区分不同的绿化树种,而面向对象的方法可以将影像分割成均匀区域并且通过一系列特征如光谱、纹理和几何等来提取该区域的信息[12],从而解决这个问题。如Immitzer等[13]对奥地利布尔根兰州的10种乔木树种进行分类,使用基于像元的方法分类总精度为69%,当使用面向对象的方法时,精度得到提高。机器学习算法在进行大量且复杂的遥感数据分析时表现出卓越的性能,如Pu等[14]利用LDA和CART方法对美国坦帕市的7种乔木树种进行分类,总精度分别为65.61%和67.22%,表明非参数机器学习算法在树种分类上的性能要优于参数机器学习算法。随机森林（Random Forest,RF）是一种基于决策树分类的非参数机器学习算法,只需较少的训练数据即可获得较高的分类精度,已有部分学者将其应用于城市树种分类的研究当中,并取得了较好的分类结果,如Naidoo等[15]结合LiDAR与高光谱数据对克鲁格国家公园8种热带稀树草原树种进行研究,得出RF是在异质性较高的环境中,对树种分类最适用的方法;李丹等[16]对首都师范大学和北京师范大学周边地区分别进行优势乔木树种分类,利用RF算法分类精度分别为75.8%和65.3%。由于大量的特征数据集不但会造成数据的冗余,还容易产生休斯效应[17],因此利用RF中的特征重要性分析,对相关特征进行筛选,并分析得出适合的特征以提高分类精度。
有鉴于此,本文通过面向对象的方法对城市绿化树种进行特征提取,通过RF中的特征重要性分析对提取的光谱、纹理、指数和几何等特征进行排序、筛选与分析,以达到提高分类精度的目的,并评估不同特征对分类结果的影响。

2 研究区概况

本文以福建省福州市福州大学旗山校区为研究区（图1）,其地理坐标为26°03′~26°05′ N,119°11′~119°13′ E,属于典型的亚热带季风气候,年平均气温为20~25 ℃。选取校区内人工绿化较为完善的地块（面积约0.6 km2）作为实验区（图1中黑色多边形区域）进行主要绿化乔木的判别与分类。
Fig. 1 A location map of the study area in Fuzhou

图1 研究区在福州市的地理位置

3 数据源及技术方法

3.1 遥感数据源

本文选用的WorldView-2卫星影像获取时间为2017年9月25日10时58分,影像无云且质量好,平均侧摆角为11.7°。WorldView-2卫星由DigitalGlobe公司发射,是具有8个多光谱波段的商用高分辨率卫星（传感器）,幅宽为16.4 km,重访周期平均为1.1 d。详细的光谱和空间信息见表1。
Tab. 1 The spectral and spatial information of WorldView-2 imagery

表1 WorldView-2影像的光谱和空间信息

波段名 波长/nm 空间分辨率/m
海岸波段 400~450 2.00
蓝光波段 450~510 2.00
绿光波段 510~580 2.00
黄光波段 585~625 2.00
红光波段 630~690 2.00
红边波段 705~745 2.00
近红外1波段 770~895 2.00
近红外2波段 860~1040 2.00
全色波段 450~800 0.50

3.2 遥感数据预处理

对影像进行辐射定标以及FLAASH大气校正等,将其亮度值（DN）转换为传感器处反射率（At-sensor Reflectance）,用来消除大气引起的影响。影像融合可以使丰富的纹理和色彩信息相结合,充分利用高分辨率遥感影像的信息内容。本次使用Gram-Schmidt Spectral Sharpening方法对影像进行融合,该方法不受波段数量的限制并且可以较好地保持原影像的光谱和纹理信息[18]。

3.3 技术方法

3.3.1 样本数据选取
样本数据的获取均通过实地调查进行,沿着研究区主要道路进行绿化树种的记录,排除数量较少以及被其他冠层遮挡的树种最终将研究树种分为7类,分别为榕树（Ficus microcarpa）、杧果（Mangifera indica L.）、香樟（Cinnamomum camphora）、重阳木（Bischofia polycarpa）、羊蹄甲（Bauhinia purpurea L.）、垂叶榕（Ficus benjamina L.）以及木棉（Bombax malabaricum DC.）,其中榕树包括黄葛树（Ficus virens Aiton）和高山榕（Ficus altissima）。各树种训练样本如图2所示。
Fig. 2 Locations of training data for different tree species

图2 各树种的训练样本位置

3.3.2 影像分割
影像分割使用eCognition软件中的多尺度分割算法与光谱差异分割算法。多尺度分割由对象的尺度、形状、颜色、紧凑性以及平滑性所决定,颜色代表的是光谱的均匀性,而形状则考虑物体的几何地貌特征。光谱差异分割通常应用在多尺度分割之后,通过合并对象来增强分割结果。在分割时加入土壤调节植被指数（SAVI）[19]有助于植被的研究[20]。分割时将SAVI指数正规化后保持与8个多光谱波段值域相同,其公式表示为式（1）。
NSAVIi=（SAVIi-SAVImin）/（SAVImax-SAVImin） （1）
式中：NSAVIi为正规化后的SAVI值;SAVIi为SAVI在像元i的值;SAVImax和SAVImin分别为SAVI的最大值和最小值。
3.3.3 随机森林
Leo Breiman于2001年提出了随机森林算法[21],是一种基于分类和回归树的多决策树分类器。每一棵决策树执行Bootstrap抽样（一种有放回的抽样算法）之后基于样本的袋外数据误差OOB（out-of-bag）error来估计计算误差。RF不考虑每个节点上的所有变量来确定最佳的分割阈值,而是使用原始特征集的随机子集,因此产生了大量的非相关决策树。RF具有不需要变量遵循特定的统计分布、训练样本少、对过度拟合不太敏感以及可以对特征重要性进行排序等优点,适用于城市树种分类的研究。本次研究中决策树的数量设为500,每个节点的特征数使用默认的数值即特征总数的平方根。
3.3.4 特征提取与选择
对于融合后的WorldView-2影像而言,高空间分辨率以及8个波段可以提供更加多样的信息。结合面向对象的方法对各对象进行特征的提取（表2）,本研究主要使用光谱、纹理、指数以及几何这4类特征。
Tab. 2 Image-object （IO） features extracted from WorldView-2 imagery

表2 从WorldView-2影像对象中提取的特征

特征类型 特征名称 描述或公式
光谱特征 平均值 1-8波段的光谱平均值
标准差 1-8波段的光谱标准差
指数特征 NDI61 ρRed-Edge-ρCoastalρRed-Edge+ρCoastal
NDI84 ρNIR2-ρYellowρNIR2+ρYellow
NDI86 ρNIR2-ρRed-EdgeρNIR2+ρRed-Edge
NDI65 ρRed-Edge-ρRedρRed-Edge+ρRed
NDI74 ρNIR1-ρYellowρNIR1+ρYellow
NDI85 ρNIR2-ρRedρNIR2+ρRed
NDVI ρNIR1-ρRedρNIR1+ρRed
SAVI (ρNIR1-ρRed)(1+L)ρNIR1+ρRed+L,其中L=0.5
纹理特征 GLCM Mean ∑i,j=0N-1Pi,jN2
GLCM Std.dev ∑i,j=0N-1Pi,j(i,j-μi,j)
GLCM Homogeneity ∑i,j=0N-1Pi,j1+(i-j)2
GLCM Contrast ∑i,j=0N-1Pi,j(i-j)2
GLCM Dissimilarity ∑i,j=0N-1Pi,j|i-j|
GLCM Entropy ∑i,j=0N-1Pi,j(-lnPi,j)
GLCM Angular second moment ∑i,j=0N-1(Pi,j)(i,μi)(j-μj)(σi2)(σj2)
GLCM Correlation ∑i,j=0N-1(Pi,j)2
GLDV Entropy ∑k=0N-1Vk(-lnVk)
GLDV Mean ∑k=0N-1k(Vk)
GLDV Contrast ∑k=0N-1Vkk2
GLDV Angular second moment ∑k=0N-1Vk2
几何特征 Compactness P 4×π×面积周长
Compactness 对象的紧致程度
Shape index 对象的光滑程度
Roundness 对象与椭圆的相似程度
Border index 对象的不规则程度
Number of edges 对象边的数量

注：其中i是行号;j是列号;Pij是单元格i,j中的归一化值;N是行数或者列数;μi, j是GLCM的平均值;σi, j是GLCM的标准差;Vk是单元格i,j矩阵中的值,k=1, 2, …, n

（1）光谱特征主要指8个波段的平均值以及标准差,共产生16种光谱特征。
（2）纹理特征使用Haralick等[22]于1973年提出的灰度共生矩阵来描述。灰度共生矩阵（Gray-Level Co-occurrence Matrix,GLCM）提供了影像中像元与整体影像及像元与像元之间的空间关系,用2个位置像元的联合条件概率密度来表示纹理。灰度差异矢量（Gray-Level Difference Vector,GLDV）是GLCM对角线的总和,它计算相邻像元间的绝对差异。共产生12个纹理特征。如果对8个波段分别计算则产生96个特征,计算量巨大并产生信息冗余。主成份分析（PCA）可以将多个变量通过多维数据压缩技术选出少数重要变量,将多维的信息集中到少数的几个特征分量上,一般主要信息集中在前2个分量中[23]。因此,通过PCA对影像进行分析,选用占据了90.46%特征贡献率的第一主成分影像来提取纹理特征,共8个GLCM特征和4个GLDV特征。
（3）指数特征是指利用指数对影像进行特征提取。比值型指数通常选取反射差异较大的两个波段,反射强的波段置于分子,通过比值运算强化差异。利用WorldView-2影像新增的4个波段,将树木反射率差异较大的波段选出构建比值型指数并进行归一化处理。因此,本文创建了6个比值型指数（Normalized Difference Index,NDI）进行研究,分别用NDI61、NDI65、NDI84、NDI74、NDI85以及NDI86表示,其中NDImn中m代表反射率较高波段,n代表反射率较低波段。此外,本文还选用SAVI以及归一化差值植被指数（NDVI）进行提取,共得到8个指数特征。
（4）几何特征基于对象的形状获取信息,根据构成对象的像元计算特征值。共选取6个与绿化乔木树种相关的几何特征[14,20]。各特征描述或计算公式见表2。
特征选择选取了特征变量中重要性较高的特征变量。特征选取的方法是利用RF中的特征重要性分析,通过大量的决策树构建RF,计算每一棵决策树的OOB误差,之后对所有OOB数据中某特征值加入干扰并再次计算OOB误差,如果OOB误差增加幅度越大即精准度下降越多则说明其重要性越高。通过多次迭代,在每一次迭代中排除特征重要性排名靠后的20%左右的特征并重复这个过程。

4 结果与分析

4.1 影像分割结果

对影像设置不同的分割尺度（50-100）进行多尺度分割。根据研究表明对颜色设置较高的权重有助于优化分割结果,因此将颜色权重设置较高（0.6、0.7、0.8、0.9）,形状与颜色权重之和为1。紧致度以及光滑度的权重各自为0.5。之后对影像进行光谱差异分割,可以提高分类效率[20]。最后设定设置分割尺度为75,形状权重为0.1,光谱差异分割的阈值为200,图3为不同分割尺度下的结果,其中图3（a）为原始影像（RGB：753）,图3（e）为最终选择的分割影像。图3（d）-（f）分别为50、75以及100分割尺度下的分割结果,可以很明显的看出100分割尺度下绿化乔木连成一片,对比矩形区域可以看出50分割尺度下对象过于细碎。图3（b）、（e）为分割尺度75时,形状权重分别为0.2和0.1的分割结果,椭圆区域显示权重为0.2时有部分乔木连成一片。图3（e）、（c）为加入SAVI指数和不加入SAVI指数的分割结果,可以看出加入SAVI指数后分割效果明显加强。
Fig. 3 Segmentation results

图3 影像分割的结果

4.2 特征选择结果

特征数量与分类总精度的关系如图4所示,为了验证特征选择的有效性,本文采用两种非参数机器学习算法,RF与支持向量机（Support Vector Machine,SVM）同时进行实验。结果表明,并非特征数量越多精度越高,无论是RF还是SVM,特征选择都有助于分类精度的提高,且RF在本次研究中的表现总是优于SVM。
Fig. 4 Relationship between the number of features and overall accuracy

图4 特征数量与分类总精度的关系

对特征选择中淘汰的特征以及精度最高的一次分类中的特征重要性的排名（图5）进行分析可知,黄光波段的光谱平均值特征在特征重要性中排名第一,这是由于研究区内存在叶色黄绿的羊蹄甲和香樟,2类共占研究区乔木树种的37.71%,叶绿素含量降低和类胡萝卜素的增多会导致叶色发黄,在光谱变化上表现为黄光附近反射率升高[24]。因此,树种在黄光波段的可分性增强;几乎所有的指数特征和光谱平均值特征都在前50%,说明这2个特征对城市绿化乔木树种分类有着重要的作用。由WorldView-2新增的4个波段构成的指数特征的重要性也大部分位于前40%,并且优于SAVI,这表明海岸、黄、红边和近红外2波段对树种分类的研究具有帮助。GLCM Angular second moment、Number of edges分别位于第2、3名,这表明纹理和几何特征对树种的分类研究也有着一定的帮助。各波段的标准差的重要性较低,排名基本都在后20%。
Fig. 5 Importance Ranking of Seleted Features

图5 选取的特征重要性排序

被淘汰的8个特征分别为4个GLDV、1个标准差特征（Standard Deviation Yellow）、1个几何特征（Border Index）以及2个GLCM特征（GLCM Dissimilarity和GLCM Correlation）,其中,GLDV的重要性最低,在第一次的选择中就全部被淘汰。对于GLCM和几何特征而言,特征内重要性差异较大。例如,GLCM Angular second moment与GLCM Entropy是排名前两位的GLCM特征,二者反映了图像中纹理的均匀或复杂程度,即图像灰度分布的均匀性,若灰度共生矩阵内所有值越相近,则越均匀,而淘汰的GLCM Dissimilarity与GLCM Correlation则表示纹理的差异性或一致性,当灰度共生矩阵行或列元素值之间越相近,则一致性越高。这表明对于城市绿化乔木树种分类的纹理特征选取,应选用与均匀性相关的特征。被淘汰的几何特征为Border index,体现了分割对象的不规则程度,该值越大,对象多边形越扭曲,即与真实树冠的形状相差越大。

4.3 分类的结果

首先利用NIR1波段对水体和阴影进行掩膜,其次通过SAVI对植被进行提取,由于草地和灌木在影像分割中的面积都相对较大,因此内部像元数量较多,利用对象内像元数量（Number of pixels）对其进行掩膜,最终得到绿化乔木占总面积的32.23%,其中各树种占其面积比例见表3。
Tab. 3 The statistics of areas and proportion of tree species

表3 树种分类的面积和比例统计

树种 面积/m2 百分比/%
香樟 57 047.10 30.00
木棉 6693.53 3.52
重阳木 3232.67 1.70
垂叶榕 15 649.92 8.23
杧果 59 462.09 31.27
榕树 33 410.58 17.57
羊蹄甲 14 661.10 7.71
分类结果如图6所示,使用约70%的样本点数据（511个样点）进行精度验证（表4）。由图4可知,当使用34个特征时分类精度最高,总精度为74.95%,kappa系数为0.67。
Fig. 6 The false color composite image and mapping results of the study area

图6 遥感影像与分类结果

Tab. 4 Accuracy assessment of results

表4 精度验证

验证数据 行合计 使用者精度/%
木棉 榕树 羊蹄甲 香樟 杧果 重阳木 垂叶榕
木棉 16 0 0 1 0 0 0 17 94.11
榕树 0 53 0 4 2 2 0 61 86.88
羊蹄甲 2 0 38 4 7 0 1 52 73.07
香樟 0 3 5 49 19 3 5 84 58.33
杧果 0 6 6 17 176 3 18 226 77.87
重阳木 0 1 0 0 5 18 0 24 75.00
垂叶榕 0 1 1 4 8 0 33 47 70.21
列合计 18 64 50 79 217 26 57
生产者精度/% 88.89 82.81 76.00 62.02 81.11 69.23 57.89
总精度/% 74.95
Kappa系数 0.67
木棉的分类精度最高,这是由于研究区内木棉长势良好,与其他树种相比树冠明显并且旁边基本没有其他乔木生长。垂叶榕的生产者精度较低,有一部分被误分为杧果,原因是由于研究区内存在大量杧果,占研究区乔木的31.27%,RF会对样本多的类别有一定的倾向性,从而导致分类结果的增多,尽管香樟所占比例也有30%,但由于之前提到的其叶色黄绿,因此,光谱可分性较高,只有少量的垂叶榕被误分为香樟。香樟的生产者精度和使用者精度都比较低,这是因为树木的光谱特征由大量的生化和结构参数所决定,相似的特征可以通过不同的参数组合所产生[13],且研究区内香樟树龄差异大,由此产生了混淆,对分类结果造成了影响。

5 结论

本研究通过面向对象的方法,对光谱信息丰富的WorldView-2卫星影像进行特征提取,对福州大学旗山校区北部片区内种植的榕树、杧果、香樟、重阳木、羊蹄甲、垂叶榕以及木棉7种主要绿化乔木树种进行分类。通过RF的特征重要性分析对42种特征进行特征选择,对大量特征的重要性进行量化并排序,去除不重要的特征,分析分类精度随不重要特征减少时的变化情况,并研究不同的特征对于分类的重要程度。结果表明,通过特征选择可以简化特征空间,从而明显提高同类型遥感数据树种分类的精度,获取的城市绿化树种信息,可为城市规划和绿化管理提供新的方法和科学依据。
分类结果表明：① 利用RF算法进行特征选择明显有助于提高分类精度,当使用34个特征对城市绿化乔木树种进行分类时,精度达到最高的74.95%,Kappa系数为0.67;② 特征重要性分析表明,黄光波段的光谱平均值、GLCM Angular second moment、Number of edges为重要性排名前三的特征,各波段的标准差的光谱特征排名均在后20%,GLDV提取的纹理特征重要性最低,全部被淘汰;③ WorldView-2影像的黄光、红边波段以及新增4个波段所构成的比值型指数重要性排名靠前,在植被遥感领域,特别是树种分类的研究中极具应用前景;④ 基于GLCM的纹理特征和几何特征存在较大的内部差异,反映纹理均匀或复杂程度的特征重要性较高,但反映一致或差异性的特征重要性却很低,对于城市绿化乔木树种的研究,应选用表征均匀或复杂程度的特征。然而,各种提取特征对树种分类精度的影响机理,还有待于进一步深入探索。
对于高空间分辨率卫星影像的树种分类而言,仍存在着不小的挑战。本文的分类方法对于树龄差异不大的乔木,分类精度较高,而对于树龄相差较大的树种,分类精度偏低。今后的研究中可以考虑进一步按树龄细分训练样本,如幼林和成熟林,以提高分类精度,或加入LiDAR数据以获取树木的结构参数特征,如胸径和树高等,用来训练更加精确的估算模型。后续研究中也可以结合实地测量的高光谱数据,完善分类模型,并有助于解析各类光谱特征对分类精度的影响。近年来,深度学习和主动学习的多种算法逐渐应用于遥感领域之中,今后可尝试结合多源遥感数据以及新算法,在更大区域的城市环境中进行绿化树种分类的研究,以获得更大尺度、更高精度的分类方法,作进一步的研究与探讨。

The authors have declared that no competing interests exist.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Shang X, Chisholm L A.Classification of australian native forest species using hyperspectral remote sensing and machine-learning classification algorithms[J]. IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, 2014,7(6):2481-2489.Mapping forest species is highly relevant for many ecological and forestry applications. In Australia, the classification of native forest species using remote sensing data remains a particular challenge since there are many eucalyptus species that belong to the same genus and, thus, exhibit similar biophysical characteristics. This study assessed the potential of using hyperspectral remote sensing data and state-of-the-art machine-learning classification algorithms to classify Australian forest species at the leaf, canopy and community levels in Beecroft Peninsula, NSW, Australia. Spectral reflectance was acquired from an ASD spectrometer and airborne Hymap imagery for seven native forest species over an Australian eucalyptus forest. Three machine-learning classification algorithms: Support Vector Machine (SVM), AdaBoost and Random Forest (RF) were applied to classify the species. A comparative study was carried out between machine-learning classification algorithms and Linear Discriminant Analysis (LDA). The classification results show that all machine-leaning classification algorithms significantly improve the results produced by LDA. At the leaf level, RF achieved the best classification accuracy (94.7%), and SVM outperformed the other algorithms at both the canopy (84.5%) and community levels (75.5%). This study demonstrates that hyperspectral remote sensing and machine-learning classification has substantial potential for the classification of Australian native forest species.

DOI

[2]
Jensen R R, Hardin P J, Hardin A J.Classification of urban tree species using hyperspectral imagery[J]. Geocarto International, 2012,27(5):443-458.Urban areas serve as humanity''s principal habitat. Because of this, it is important to understand the biophysical components of the urban environment including the urban forest. The goal of this study was to determine the potential to classify individual urban trees as a function of spectral features derived from airborne hyperspectral data. To determine this, 500 urban trees were identified (through fieldwork) in the built-up zone of Provo-Orem, Utah, USA. Visible and near infrared airborne hyperspectral imagery was collected over the same area. The 500 trees were identified on the images, and spectral features of each tree were extracted. Principal components, vegetation indices, band means, and band ratios were all used as features to discriminate between different tree species. The tree classification was 82% accurate when just the six principal components were used. Classification accuracy increased to 91.4% after combining vegetation indices, band mean values and band ratios.

DOI

[3]
黄慧萍,吴炳方,李苗苗,等.高分辨率影像城市绿地快速提取技术与应用[J].遥感学报,2004,8(1):68-74.高分辨率遥感影像是城市绿地信息快速提取的主要数据源 ,文中以多尺度影像分割与面向对象影像分析方法为主要技术 ,利用样本多边形对象的成员函数建立训练区 ,自动提取大庆市城市绿地覆盖信息 ,达到清查城市绿地的目的。该方法信息获取周期短、精度高、成本低 ,实现了城市绿地信息精确获取与快速更新。

DOI

[ Huang H P, Wu B F, Li M M, et al.Detecting urban vegetation efficiently with high resolution remote sensing data[J]. Journal of Remote Sensing, 2004,8(1):68-74. ]

[4]
李成范,尹京苑,赵俊娟.一种面向对象的遥感影像城市绿地提取方法[J].测绘科学,2011,36(5):112-114.针对目前面向对象方法在高分辨率遥感影像中提取绿地专题信息的特 点,以及受到城市区域范围尺度限制因素的影响,本文提出一种利用高分辨率遥感影像提取城市区域范围绿地专题信息的方法,结合影像分割、影像面向对象模板统 计和基于多阀值的模糊分类技术的优势,林地、密草地和疏草地专题信息被提取出来.研究中提出利用基于面向对象多尺度分割算法和最大似然监督分类法的提取结 果进行精度评价,实验结果证明,在城市区域范围尺度上,该方法计算简便且实现了93.72％的高精度和总Kappa系数为0 8236的评价结果.

[ Li C F, Yin J Y, Zhao J J.An extraction algorithm of urban vegetation from remote sensing image based on object-oriented approach[J]. Science of Surveying and Mapping, 2011,36(5):112-114. ]

[5]
李登秋,居为民,郑光,等.基于生态过程模型和森林清查数据的森林生长量估算对比研究[J].生态环境学报,2013,22(10):1647-1657.利用遥感驱动的生态过程模型-Boreal Ecosystem Productivity Simulator(BEPS)、2001-2006年国家森林资源连续清查数据(一类清查-样地尺度)和2003-2009年森林资源规划设计调查数据(二类调查-区域尺度),分别计算江西省吉安市的森林生态系统生长量,从不同空间尺度和森林类型对3种数据源估算的森林生长量进行了分析。结果表明,样点尺度上,BEPS模型模拟的森林生长量(4.18 Mg·hm-2·a-1)低于群落生长量(5.86 Mg·hm-2·a-1),与乔木层生长量(4.29 Mg·hm-2·a-1)基本一致,模型模拟结果与两者的拟合R2分别为0.48和0.43。区域尺度上,BEPS模型模拟、二类调查数据计算的群落及乔木层生长量分别为4.65、4.36和3.34 Mg·hm-2·a-1,BEPS模型估算的吉安市各县森林总生长量与二类调查数据计算的群落、乔木层生长总量拟合R2分别达0.84和0.83。一类清查数据计算结果高于二类清查数据计算结果,BEPS模型模拟森林生长量分别与基于一类清查数据计算的乔木层生长量及二类调查数据群落生长量较为一致。从研究区两种主要森林类型来看,常绿阔叶林年平均生长量高于常绿针叶林,常绿针叶林与模型估算结果差异小于常绿阔叶林。最后利用模型估算了研究区2001-2010年平均生长量,为认识研究区的森林生长空间分布差异及更新森林生物量提供支持。

DOI

[ Li D Q, Ju W M, Zheng G, et al.Comparison of estimated forest biomass increment rate based on a process-based ecological model and forest inventory data[J]. Ecology and Environmental Sciences, 2013,22(10):1647-1657. ]

[6]
Korpela I, Ørka H O, Maltamo M, et al.Tree species classifcation using airborne LiDAR-effects of stand and tree parameters downsizing of training set intensity normalization and Sensor Type[J]. Silva Fennica, 2010,44(2):319-339.ABSTRACT Tree species identification constitutes a bottleneck in remote sensing-based forest inventory. In passive images the differentiating features overlap and bidirectional reflectance hampers analysis. Airborne LiDAR provides radiometric and geometric information. We examined the single-trees-level response of two LiDAR sensors in over 13 000 forest trees in southern Finland. We focused on the commercially important species. Our aims were to 1) explore the relevant LiDAR features and study their dependencies on stand and tree variables, 2) examine two sensors and their fusion, 3) quantify the gain from intensity normalizations, 4) examine the importance of the size of the training set, and 5) determine the effects of stand age and site fertility. A set of 570 semiurban broad-leaved trees and exotic conifers was analyzed to 6) examine the LiDAR signal in the economically less important species. An accuracy of 88-90% was achieved in the classification of Scots pine, Norway spruce, and birch, using intensity variables. Spruce and birch showed the highest levels of confusion. Downsizing the training set from 30% to 2.5% of all trees had only a marginal effect on the performance of classifiers. The intensity features were dependent on the absolute and relative sizes of trees, especially for birch. The results suggest that leaf size, orientation, and foliage density affect the intensity, which is thus not affected by reflectance only. Some of the ecologically important species in Finland may be separable, since they gave rise to high intensity values. Comparison of the sensors implies that performance of the intensity data for species classification varies between sensors for reasons that remained uncertain. Both range and gain receiver normalization improved species classification. Weighting of the intensity values improved the fusion of two LiDAR datasets.

DOI

[7]
宫鹏,浦瑞良,郁彬.不同季相针叶树种高光谱数据识别分析[J].遥感学报,1998,2(3):211-217.利用高分辨率光谱仪在实地测得的光谱数据来识别美国加州的6种主要针叶树种。树冠阴面和阳面的高光谱数据分别在1996年夏、秋测得。首先对原始光谱数据作简单处理, 然后进行6种数据变换：对数变换、一阶微分变换、对数变换后一阶微分变换、归一化变换、归一化变换后一阶微分变换及归一化后对数变换。采用相邻窄波段逐步加宽的办法, 测试不同波段宽度对树种识别精度的影响。所有的变换方法及波段宽度试验最后均由神经元网络算法产生的树种分类精度来评价。试验结果表明对数变换后一阶微分和归一化变换后一阶微分能够获得高于94%的平均精度

DOI

[ Gong P, Pu R L, Yu B.Conifer species recognition with seasonal hyperspectral data[J]. Journal of Remote Sensing, 1998,2(3):211-217. ]

[8]
王芳,黎夏,卓莉,等.基于Hyperion高光谱数据的城市植被胁迫评价[J].应用生态学报,2007,18(6):1286-1292.快速获取城市植被的胁迫状态,不仅对城市植被健康状况的维护,而且对城市生态环境的改善具有重要意义.在对受胁迫植被的生理特征和光谱特征进行分析的基础上,利用星载高光谱Hyperion数据,计算出与胁迫相关的14种高光谱植被指数,在此基础上运用BP神经网络算法建立了城市植被胁迫强度分类器,对城市植被的胁迫强度进行了识别与分析.结果表明:城市中心商住区的植被受胁迫程度明显高于城乡结合部和郊区;植被的受胁迫现象在大块绿地外围呈环状分布;构建的植被胁迫强度分类器能够较为准确地反映植被受胁迫的强度信息,可为大面积城市植被胁迫监测提供一种较为可靠而快捷的方法.

DOI

[ Wang F, Li X, Zhuo L, et al.A stressed level of urban vegetation: Its assessment based on Hyperion hyperspectral data[J]. Chinese Journal of Applied Ecology, 2007,18(6):1286-1292. ]

[9]
Agarwal S, Vailshery L S, Jaganmohan M, et al.Mapping urban tree species using very high resolution satellite imagery:Comparing pixel-based and object-based approaches[J]. ISPRS International Journal of Geo-information, 2013,2(1):220-236.

DOI

[10]
毕恺艺,牛铮,黄妮,等.基于Sentinel-2A时序数据和面向对象决策树方法的植被识别[J].地理与地理信息科学,2017,33(5):16-27.Sentine-2A数据具有较高的空间分辨率和时间分辨率,克服了以往时序数据难以获取或空间分辨率低的问 题.该文以山西省吕梁市陈家湾流域为研究区,基于 Sentine-2A时序数据,根据归一化植被指数（ NDVI）时序曲 线特征和光谱特征,构建基于面向对象决策树方法的分层分类模型,成功提取了陈家湾流域的植被信息,分类总体 精度达到89. 7%, Kappa系数为0 87.基于面向对象决策树方法的多时相分类结果与单时相分类结果相比,可以 有效改善波谱特征相近和受地形影响较大地物的区分,减少混分现象;基于 Sentinel2 2A 时序数据和面向对象决策 树分类方法能够有效提高植被分类的精度.

DOI

[ Bi K Y, Niu Z, Huang N, et al.Identifying vegetation with decision tree model based on object-oriented method using multi-temporal Sentinel-2A images[J]. Geography and Geo-information Science, 2017,33(5):16-27. ]

[11]
关舒婧,韩鹏鹏,王月如,等.华南地区典型种植园地遥感分类研究[J].地球信息科学学报,2017,19(11):1538-1546.华南地区种植园地广泛分布,类型混杂多样,导致园地分布信息难以正确获取,为农业管理造成了较大困难。本研究基于Landsat8 OLI数据,通过数据融合、特征优化,应用随机森林算法构建面向对象的种植园地分类规则集,对华南地区典型经济作物香蕉、柑橘、葡萄、蒲葵、海枣、番木瓜和火龙果等进行类别识别,同时对比贝叶斯分类法、K最邻近分类法、支持向量机法、决策树分类法的分类效果。结果表明：数据融合会在一定程度上影响分类结果精度;植株形态、光谱特征接近,种植期交错是影响华南地区典型园地分类精度的重要原因;以中分辨率影像为数据源,面向对象的随机森林算法应用于种植园地分类研究总体精度可达88.05%,Kappa系数0.87,可以有效区分华南地区典型种植园地类别;相比于其他算法,随机森林算法在分类精度、可靠性和稳定性上具有一定优势,可为园地作物生长监测和种植管理提供科学依据。

DOI

[ Guan S J, Han P P, Wang Y R, et al.Study on the classification of typical plantations in South China[J]. Journal of Geo-information Science, 2017,19(11):1538-1546. ]

[12]
Dian Y Y, Li Z, Pang Y.Spectral and texture features combined for forest tree species classification with airborne hyperspectral imagery[J]. Journal of the Indian Society of Remote sensing, 2015,43(1):101-107.Forest precision classification products were the basic data for surveying of forest resource, updating forest subplot information, logging and management of forest. However, due to the diversity of stand structure, complexity of the forest growth environment, it is difficult to discriminate forest tree species using multi-spectral image. The airborne hyper-spectral images can obtain high spatial and spectral resolution imagery of forest canopy, so it may be useful for tree species level classification. The aim of this paper was to test the effective of combining spatial and spectral features in airborne hyper-spectral image classification. The CASI hyper spectral image data were acquired from Liangshui natural reserves area. First the MNF (minimum noise fraction) transform method for to reduce the hyperspectral image dimensionality and highlighting variation. Second, the grey level co-occurrence matrix (GLCM) is used to extract the texture features of forest tree canopy. Thirdly the texture and the spectral features of forest canopy were fused to classify the trees species using support vector machine (SVM) with different kernel functions. The results showed that when using the SVM classifier, MNF and texture-based features combined with linear kernel function can achieve the best overall accuracy which was 85.92 %. It also confirmed the belief that combined the spatial and spectral information can improve the accuracy of tree species classification.

DOI

[13]
Immitzer M, Atzberger C, Koukal T.Tree species classification with random forest using very high spatial resolution 8-band WorldView-2 satellite data[J]. Remote Sensing, 2012,4(9):2661-2693.

DOI

[14]
Pu R L, Landry S.A comparative analysis of high spatial resolution IKONOS and WorldView-2 imagery for mapping urban tree species[J]. Remote Sensing of Environment, 2012,124:516-533.78 High spatial/spectral resolution data are vital to mapping tree species. 78 WorldView-2 has a great potential to improve the vegetation classification. 78 WorldView-2 outperforms IKONOS for identifying tree species/groups. 78 A stepwise masking system is effective for mapping tree species.

DOI

[15]
Naidoo L, Cho M A, Mathieu R, et al.Classification of savanna tree species in the Greater Kruger National Park region by integrating hyperspectral and LiDAR data in a random forest datamining environment[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2012,69(3):167-179.The accurate classification and mapping of individual trees at species level in the savanna ecosystem can provide numerous benefits for the managerial authorities. Such benefits include the mapping of economically useful tree species, which are a key source of food production and fuel wood for the local communities, and of problematic alien invasive and bush encroaching species, which can threaten the integrity of the environment and livelihoods of the local communities. Species level mapping is particularly challenging in African savannas which are complex, heterogeneous, and open environments with high intra-species spectral variability due to differences in geology, topography, rainfall, herbivory and human impacts within relatively short distances. Savanna vegetation are also highly irregular in canopy and crown shape, height and other structural dimensions with a combination of open grassland patches and dense woody thicket a stark contrast to the more homogeneous forest vegetation. This study classified eight common savanna tree species in the Greater Kruger National Park region, South Africa, using a combination of hyperspectral and Light Detection and Ranging (LiDAR)-derived structural parameters, in the form of seven predictor datasets, in an automated Random Forest modelling approach. The most important predictors, which were found to play an important role in the different classification models and contributed to the success of the hybrid dataset model when combined, were species tree height; NDVI; the chlorophyll b wavelength (466nm) and a selection of raw, continuum removed and Spectral Angle Mapper (SAM) bands. It was also concluded that the hybrid predictor dataset Random Forest model yielded the highest classification accuracy and prediction success for the eight savanna tree species with an overall classification accuracy of 87.68% and KHAT value of 0.843.

DOI

[16]
李丹,柯樱海,宫辉力,等.基于高分辨率遥感影像的城市典型乔木树种分类研究[J].地理与地理信息科学,2016,32(1):84-89.为探索高分辨率遥感影像对城市复杂环境优势乔木树种分类的有效性,采用面向对象分类方法,基于WorldView-2影像对首都师范大学及周边地区（CNU）、北京师范大学及周边地区（BNU）两个研究区进行优势乔木树种（泡桐、法国梧桐、杨树、国槐、银杏）分类。首先对WorldView-2影像进行分割,获得树冠区域及其49个属性特征,包括31个光谱属性和18个纹理属性;随后利用随机森林RF与支持向量机SVM两种分类算法对树冠区域进行分类。CNU研究区SVM与RF总体分类精度分别为86.5%、75.8%,Kappa系数为0.801、0.648;BNU研究区SVM与RF总体分类精度分别为66.9%、65.3%,Kappa系数为0.541、0.520。实验表明WorldView-2影像能有效实现城市非阴影区域优势乔木树种分类,但异质性较高、树种分布分散的区域分类精度低于异质性较小、树种分布密集的区域;WorldView-2影像的4个新增波段尤其是红边波段的派生属性在分类过程中所占权重值较高。

DOI

[ Li D, Ke Y H, Gong H L, et al.Urban tree species classification with machine learning classifier using WorldView-2 imagery[J]. Geography and Geo-information Science, 2016,32(1):84-89. ]

[17]
刘怀鹏,安慧君,王冰,等.基于递归纹理特征消除的WorldView-2树种分类[J].北京林业大学学报,2015,37(8):53-59.

[ Liu H P, An H J, Wang B, et al.Tree species classification using WorldView-2 images based on recursive texture feature elimination[J]. Journal of Beijing Forestry University, 2015,37(8):53-59. ]

[18]
Pu R L, Landry S, Yu Q.Object-based urban detailed land cover classification with high spatial resolution IKONOS imagery[J]. International Journal of Remote Sensing, 2011,32(12):3285-3308.

DOI

[19]
Huete A R.A soil-adjusted vegetation index (SAVI)[J]. Remote Sensing of Environment, 1988,25(3):295-309.A transformation technique is presented to minimize soil brightness influences from spectral vegetation indices involving red and near-infrared (NIR) wavelengths. Graphically, the transformation involves a shifting of the origin of reflectance spectra plotted in NIR-red wavelength space to account for first-order soil-vegetation interactions and differential red and NIR flux extinction through vegetated canopies. For cotton ( Gossypium hirsutum L. var DPI-70) and range grass ( Eragrostics lehmanniana Nees) canopies, underlain with different soil backgrounds, the transformation nearly eliminated soil-induced variations in vegetation indices. A physical basis for the soil-adjusted vegetation index (SAVI) is subsequently presented. The SAVI was found to be an important step toward the establishment of simple lobal that can describe dynamic soil-vegetation systems from remotely sensed data.

DOI

[20]
Puissant A, Rougier S, Stumpf A.Object-oriented mapping of urban trees using random forest classifiers[J]. International Journal of Applied Earth Observation and Geoinformation, 2014,26(1):235-245.Since vegetation in urban areas delivers crucial ecological services as a support to human well-being and to the urban population in general, its monitoring is a major issue for urban planners. Mapping and monitoring the changes in urban green spaces are important tasks because of their functions such as the management of air, climate and water quality, the reduction of noise, the protection of species and the development of recreational activities. In this context, the objective of this work is to propose a methodology to inventory and map the urban tree spaces from a mono-temporal very high resolution (VHR) optical image using a Random Forest classifier in combination with object-oriented approaches. The methodology is developed and its performance is evaluated on a dataset of the city of Strasbourg (France) for different categories of built-up areas. The results indicate a good accuracy and a high robustness for the classification of the green elements in terms of user and producer accuracies.

DOI

[21]
Leo Breiman.Random forests[J]. Machine Learning, 2001,45(1):5-32.

DOI

[22]
刘凯,汤国安,黄骁力,等.面向地形特征的DEM与影像纹理差异分析[J].地球信息科学学报,2016,18(3):386-395.lt;p>纹理分析方法在宏观地形特征分析方面具有较大的优势与潜力,但当前缺少对DEM与影像数据纹理特征差异的系统分析研究.本文采用灰度共生矩阵为纹理量化模型,选取了8个不同地貌单元的样本数据,对DEM和遥感影像2类数据的纹理进行了特征值对比分析,纹理特征稳定性分析,纹理特征组间差异性分析.实验结果表明,在所测试的二阶角矩,对比度,方差,熵4个纹理指标中,DEM和影像的对比度特征值间具有显著的相关性;通过不同地貌样区纹理特征值对比分析发现,DEM数据在地形起伏较大区域纹理特征更为明显,遥感影像数据则受地表覆盖物影响较大;从地形特征的稳定性角度分析,DEM数据在丘陵和山地分析有优势,影像数据则在平原和台地分析表现更好;从地形特征差异性角度分析,DEM数据要优于影像数据.进一步采用光照模拟和坡度数据以增加DEM纹理信息,研究结果表明,DEM派生的2类数据在地形量化差异性方面改进明显,并大大优于影像数据.</p>

DOI

[ Liu K, Tang G A, Huang X L, et al.Research on the difference between textures derived from DEM and remote-sensing image for topographic analysis[J]. Journal of Geo-information Science, 2016,18(3):386-395. ]

[23]
徐涵秋. 城市遥感生态指数的创建及其应用[J].生态学报,2013,33(24):7854-7862.

[ Xu H Q.A remote sensing urban ecological index and its application[J]. Acta Ecologica Sinica, 2013,33(24):7854-7862. ]

[24]
Blackburn G A.Quantifying chlorophylls and caroteniods at leaf and canopy scales: An evaluation of some hyperspectral approaches[J]. Remote Sensing of Environment, 1998,66(3):273-285.

DOI

Options
Outlines

/

〈 〉