基于植被特征库的高光谱植被精细分类
单词
CPS：class‐pair separability
FBS：Feature Band Set
FISS：Field Imaging Spectrometer System
CPS：class-pair separability
GG：goose grass
GLCM：Grey Level Co‐occurrence Matrix
LISA：Local Indicators Spatial Association
NPSAD：neighborhood pixels’ spectral angle distances
OIF：Optimum Index Factor
PU：purslane
SVC：Sophisticated Vegetation Classification
SAD：spectral angle distance
WS：watermelon seedlings
0、Introduction
①
This paper aims to take fully advantage of the effective features of hyperspectral data which are beneficial to SVC ，and to initially construct the vegetation feature band set for SVC using hyperspectral data ．
本文目的：①充分利用高光谱数据中有益于SVC的有效特征；②用高光谱数据构造用于SVC的植被分类数据集。
②
In order to improve the accuracy of SVC ， we propose a spectral‐dimension optimization and a spatial‐dimension optimization based on current band quality optimization and feature selection algorithm of hyperspectral data ．
从两个维度（光谱维、空间维）对当前的波段质量优化和特征选择算法进行优化，以提高SVC的准确度。
1、Study Area and Data Set
①
②
③
Our study area contains four vegetation classes of interest ， including watermelon seedlings （WS ） ， goose grass （GG） ，wild amaranth （WA） and purslane （PU）
研究对象，四种植物：WS、GG、WA、PU
④
The original FISS image and the ground truth image are shown in Fig １ ．The average spectral curves of four vegetation classes are shown together in Fig ２ ．From Fig２
Fig1：原始图像
Fig2：光谱曲线
⑤
we see that C１ and C２ have similar spectral shape in the original spectral space ， while reflectance of C１ is higher than that of C２ ，especially in the near infrared bands ．However ，C３ and C４ are very similar in both spectral shape and reflectance ，making it difficult to distinguish from each other by simply utilizing original spec‐ tral information ．
C2：GG：goose grass
C4：PU：purslane
C1：WS：watermelon seedlings
C1曲线与C2类似，但是反射率比C2高。
C3和C4则有非常相似光谱曲线和反射率——这就使得通过原始光谱信息区分它们变得很困难。
2、Method
The classification strategy based on construction and op‐ timization of vegetation FBS includes the following stages ：……At last ，classify the whole image with appropri‐ ate classifier and calculate classification accuracy ．
基于FBS优化的分类策略，步骤：
①选择确定数量的训练样本，基于每种分类的样本和窗口，计算vegetation FBS中的所有特征；
②对训练样本使用基于CPS的光谱维优化算法；
③对整幅图像使用光谱维优化算法；
④对当前的FBS应用基于临近像素光谱角距离（NPSAD）的空间维优化算法；
⑤用适当的分类器对整幅图像进行分类，计算分类准确性。
2.1、Construction of vegetation feature band set（FBS）
①
The first is spectral feature obtained directly from the original bands ．
The second is vegetation spatial texture feature derived from hyperspectral images ．
The third is spectral indices by calculation between different spectral bands which are sensitive to biochemical parameters of vegetation ．
用于SVC的FBS主要包含了三种特征：
a、直接源自初始波段的光谱特征；
b、植被纹理特征；
c、通过不同波段（对植被相关的参数比较敏感的波段）的混合计算得到的光谱指标。
2.1.1、Spatial texture features
①
Grey Level Co‐occurrence Matrix （GLCM ） is one of the most common approaches to spatial and texture feature analy ‐ sis for images ，by investigating the joint distribution of two pixels and then get the spatial correlation of grey levels of the image［６］ ．
GLCM是最常用的空间、纹理特征分析方法。
②
We mainly consider seven types of texture features when constructing vegetation FBS ， including mean value ， homogeneity ，contrast ， heterogeneity ，entropy ，second moment and relevance ．
主要考虑7种纹理特征……
③
In addition to GLCM ，Local Indicators Spatial Association （LISA ） also plays an important role at texture analysis of high spatial resolution image［７］ ，including Moran index I ，Geary coefficient C and G statistic ［８ ，９］ ．
除了GLCM之外，LISA也是对高分影像进行分析的重要工具，包括……
④
In this paper ，we mainly use the localized indices when performing spatial autocorrelation texture feature extraction ．
本文中，在提取空间自相关纹理特征提取时，主要用到了局部性的指标。
⑤
Principal Component Analysis （ PCA ） is firstly performed and the first n components with accumulated contribution reaching a certain threshold （e ．g ．，９０％ ，９５％ or ９８％ ） are retained ，followed by the texture band calculation using the n components respectively based on GLCM and LISA ，to extract the texture features ．
使用PCA，通过设定阈值筛选前n个主成分，再分别用GLCM和LISA对这些主成分进行计算，提取纹理特征。
2.1.2、Spectral indices features
①
So we didn’t select biological parameters of vegetation directly ，instead the spectral indices sensitive to biological parameters of vegetation were selected when constructing FBS for SVC using hyperspectral data ．
在用高光谱数据构造FBS时，我们并不直接选取植被的生物学参数，而是选对这些参数敏感的光谱指标。
②
According to previous studies ，we selected 3０ kinds of spectral index features when constructing vegetation FBS ， as shown in Table １ ．
Tab 1中展示了本文用于构造FBS的光谱指标，有30个。
2.2、Spectral‐dimension optimization algorithm CPS
①
The common feature selection algorithm for hyperspectral data are based on information content criteria or class‐pair separability criteria
常用的高光谱数据特征提取算法有两种：a、基于信息内容；b、基于孤立的类别对（？）
②
The former selects bands with rich information and small correlation ，represented by Optimum Index Factor （OIF）［３０］ ．
前者选择那些有着大量信息、更小相关性的波段，代表是OIF。
③
The latter selects the optimal band combination according to class‐pair statistics ，the bigger the statistical distance （e ．g ．Bhattachryya Distance ［３１］ and Jeffries‐Ma‐ tusita Distance［３２］ ） ， the better the class‐pair separability is ． Bhattachryya distance （B distance） usually indicates good separability between bands without a fixed threshold ，while Jeffries‐Matusita distance （J‐M distance） distributes between ０ and ２ ，good separability between samples can be expressed by a distance value larger than １.９ ．
后者基于类别对的统计，选择最优的波段组合——两个类别的统计性距离越大，那这个类别对的分离性就越好。
给出了了两种统计性的距离B distance和J‐M distance。
④
The feature bands selected based on information content have rich information ，but the distinction between classes may not be specific ，while those selected based on class‐pair separability may have large correlation ．Therefore ，in order to improve classification accuracy ，a spectral dimension optimization algorithm is proposed by combining two types of algorithms ， which focus on the separability between different classes ， namely Class Pair Separability （CPS ） ．
②中提到的方法包含了大量的信息，但是类别之间的区分度可能不大；③中提到的方法则包含了较高的相关性。于是就提出CPS方法，用于结合方法②和③，以提高分类准确度。
这是一个光谱维优化算法。
⑤
The basic principle of the algorithm is as follows ：
First ， utilize B distance to select the feature bands with highest separability for each class pair ，get the initial FBS ．
Second ，test the separability for each class pair of the selected feature set with J‐M distance ．
Next ，update initial FBS through an iteration until the FBS have good separability for all classes ．
At last ，optimize the FBS using OIF to reduce the bands correlation and data redundancy ，then gain the FBS with spatial‐dimension optimization ．
算法原理：
a、用B距离为每个类别对选择具有最高分离度的特征波段，得到原始FBS；
b、用J-M距离测试这些类别对的分离度；
c、通过迭代，更新原始FBS，直到FBS对所有类别对都有更好的分离度；
d、用OIF优化FBS，得到空间维优化的FBS
⑥
OIF、B距离、J-M距离的算法
2.3、Spatial‐dimension optimization algorithm NPSAD
①
we propose a spatial‐dimension optimization algorithm of FBS based on neighborhood pixels’ spectral angle distance （NPSAD） so that the spectral‐dimension information of hyperspectral data could be fully used while making spatial‐dimension optimization ．
本文提出NPSAD，这样在进行空间维优化时，高光谱数据的光谱维信息也会得到充分利用。
②
The essence of this algorithm is to filter the neighboring pixels by setting window size and a threshold of spectral angle distance （SAD） based on the traditional mean filtering ．
算法的核心在于，通过传统的均值滤波（？）设置窗口尺寸和SAD阈值来对临近像素进行滤波。
③
The neighboring pixels and the central pixel are considered to be the same class for subsequent calculations only when the spatial distance and spectral angular distance meet the requirements simultaneously ．
只有当空间距离和光谱角距离同时满足要求，中心像元和临近像元才会被视为一个分类。
④
Unlike conventional window filter ，the proposed method is able to smooth the image while keeping detailed information and avoiding the edge effect effectively ． The specific algorithm process is shown in Figure ３ ．
相比于传统的窗口滤波，该方法在平滑图像的同时可以保留细节信息，避免边际效应。
Fig3展示了该方法的流程。
⑤
If there are training samples of each class ，calculate the mean SAD between training samples of each class and choose the minimum value as MinSAD ；If there are not training samples ，the MinSAD should be set manually ．
如果每个类别都有训练样本，那么计算每个类别的训练样本的平均SAD，选取其中最小值作为MinSAD；如果没有训练样本，那么MinSAD就要手动设置。
⑥
nb：波段数量
CP：中心像元的光谱
NP：临近像元的光谱
3、Experimental Results
①
The SVM classifier from ENVI is used with default parameter setting ．
The specific classification methods are shown in Table ２
使用ENVI的SVM分类器作为对照。
Tab 2给出了全部分类方法。
②
In the vegetation FBS based on FISS data ， the original bands are reflectance ；the spatial texture features are calculated using the first three principal components after PCA with the window size of ３ × ３ and the direction of ４５° ；the spectral index features are all the indices in Table １ ．
空间纹理特征通过PCA的前三个主成分进行计算（窗口3*3，方向45°）；光谱特征则是Tab 1中的全部指标。
③
According to the ２５ randomly selected training samples in each class ， utilize spectral‐dimension optimization algorithm of FBS based on class‐pair separability （CPS ） and obtain the FBS after spectral‐dimension optimizing ． We ultimately retained ２８ vegetation features ，including １０ original bands ，８ texture features and １０ spectral index features ．
The specific information of reserved vegetation feature bands after spectral dimension optimizing are as follows ：……which are indicators mainly representing chloro‐phyll ，nitrogen and carotenoids ．
通过对每个类别随机选择的25个训练样本使用基于CPS的光谱维优化算法，最终得到了28个植被特征——10个原始波段，8个纹理特征和10个光谱特征。
这些特征分别是：……
④
When making spatial‐dimension optimization of FBS ，automatic computation of MinSAD using the training sample is implemented ，with the window size of ９ × ９ and ２ times iteration ．
在进行空间维FBS优化时，对训练样本进行了自动化计算MinSAD，窗口9*9，两次迭代。
问题
1、原始数据获取与导入到程序中
2、各个类别的训练集的获取，以及MinSAD的计算
3、如果没有训练集，如何手动设置MinSAD
4、Fig 3给出的算法，倒数第二步
这里的整副图像的平滑度达到要求，其中a、平滑度如何计算；b、最终的要求如何设置。
5、在第3部分的第④个语句中，两（倍/次?）迭代的意思
6、本文给出的30个光谱维植被指标，如何具体筛选为最终的10个的？
7、最终筛选的8个空间维纹理特征，计算公式是什么？（其中6和7的这些指标，可以在第3部分第③个语句中找到）
