1、25002500 單詞，單詞，39003900 漢字漢字 出處：出處：Du PDu P, Tao F, Hong T. Spectral Features Extraction in Hyperspectral RS Data and Its Application to, Tao F, Hong T. Spectral Features Extraction in Hyperspectral RS Data and Its Application to -298.-298. 本科畢業(yè)設(shè)計（論文）本科畢業(yè)設(shè)計（論文） 中英文對照翻譯中英文對照翻譯 院（系部） 測繪與國土信息工程學院 專業(yè)名稱測

2、繪工程 年級班級 學生姓名 指導老師 2012 年 6 月 3 日 Spectral Features Extraction in Hyperspectral RS Data andSpectral Features Extraction in Hyperspectral RS Data and Its Application to Information ProcessingIts Application to Information Processing Oriented to the demands of hyperspectral RS information processing a

3、nd applications, spectral features in hyperspectral RS image can be categorized into three scales: point scale, block scale and volume scale. Based on the properties and algorithms of different features, it is proposed that point scale features can be divided into three levels: spectral curve featur

4、es, spectral transformation features and spectral similarity measure features. Spectral curve features include direct spectra encoding, reflection and absorption features. Spectral transformation features include Normalized Difference of Vegetation Index (NDV I) , derivate spectra and other spectral

5、 computation features. Spectral similarity measure features include spectral angle ( SA ) , Spectral Information Divergence ( SID ) , spectral distance, correlation coefficient and so on. Based on analysis to those algorithms, several problems about feature extraction, matching and application are d

6、iscussed further, and it p roved that quaternary encoding, spectral angle and SID can be used to information processing effectively. 1 Introduction1 Introduction Hyperspectral Remote Sensing was one of the most important breakthroughs of Earth Observation System ( EOS) in 1990 s. It overcomes the li

7、mitations of conventional aerial and multispectral RS such as less band amount, wide band scope and rough spectral information expression, and can provide RS information with narrow band width, more band amount and fine spectral information, also it can distinguish and identify ground objects from s

8、pectral space, so hyperspectral RS has got wide applications in resources, environment, city and ecological fields. Because hyperspectral RS is different from conventional RS information obviously in both information acquisition and information processing, there are many problems should be solved in

9、 practice. One of the most important problems is about spectral features extraction and application in hyperspectral RS data including hyperspectral RS image and standard spectral database. Nowadays, studies on hyperspectral are mainly focused on band selection and dimensionality reduction, image cl

10、assification, mixed pixel decomposition and others, and studies on spectral features are few. In this paper, spectral features extraction and application will be taken as our central topic in order to provide some useful advices to hyperspectral RS applications. 2 Framework of spectral features in h

11、yperspectral RS data2 Framework of spectral features in hyperspectral RS data In general, hyperspectral RS image can be expressed by a spatial-spectral data cube ( Fig. 1). In this data cube, every coverage expressed the image of one band, and each pixel forms a spectral vector composed of albedo of

12、 ground object on every band in spectral dimension, and that vector can be visualized by spectral curve ( Fig. 2 ). Many features can be extracted from spectral vector or curve, and spectral features are the key and basis of hyperspectral RS applications. Also each spectral curve in spectral 1 datab

13、ase can be analyzed with same method. Although there are some algorithms to compute spectral features, the framework and system is still not obvious, so we would like to propose a framework for spectral features in hyperspectral RS data including hyperspectral RS image and standard spectral database

14、. Fig. 1Hyperspectral image data cubeFig. 2Reflectance spectral curve of a pixel 2. 12. 1Three scales of spectral featuresThree scales of spectral features According to the operational objects of extraction algorithms, spectral features can be categorized into three scales: point-scale, block-scale

15、and volume- Scale. Point scale takes pixel and its spectral curve as operational object and some useful features can be extracted from this spectral vector (or spectral curve).In general, hyperspectral RS image takes spectral vector of each pixel as processing object. Block scale is oriented image b

16、lock or region. Block is the set of some pixels, and it can be homogeneous or heterogeneous. Homogeneous regions are got by image segmentation and pixels in this region are similar in some given features; heterogeneous region are those image blocks with regular or irregular size, and they are cut fr

17、om original image directly, for example, an image can be segmented according to quadtree method. In hyperspectral RS image, block scale features can be computed from two aspects. One is to compute texture feature of a block on some characterized bands, and the other is to compute spectral feature of

18、 a block. If the block is homogeneous its mean vector can be computed firstly and then spectral of this mean vector can be extracted to describe the block. If the block is heterogeneous, it can be segmented to some homogeneous blocks. Volume scale combines spatial and spectral features in a whole an

19、d extracts features in 3D ( row, column and spectra ) space. Here, some 3D operational algorithms are needed, for example, 3D wavelet transformation and high order Artificial Neural Network (ANN ). Because this type of features is difficult to compute and analyze, we dont research it in current stud

20、ies. In this paper, we would like to focus on point scale feature, or those features extracted from spectral 2 vector that may be spectral vector of a pixel or mean vector of a block. 2. 22. 2Three levels of point scale featuresThree levels of point scale features From operation object, algorithm pr

21、inciples, feature properties, application modes and other aspects, we think it is feasible to categorize spectral features into three levels: spectral curve features, spectral transformation features and spectral similarity measure features. They are corresponding to analysis on spectral curve with

22、all bands, data transformation and combination with part of all bands and similarity measure of spectral vectors. In our study, data from OM IS and PHI hyperspectral image, USGS spectral database and typical spectra data in China is experimented and two examples are given in this paper. One is to se

23、lect three regions from PH I image (Region I is vegetation, Region II is built-up land, and Region III is mixed region of some land covers) , and the other is spectral curve of three ground objects from USGS spectral database, among them S1 is Actinolite_HS22. 3B, S2 is Actinolite_HS116. 3B and S3 i

24、s Albite_HS66. 3B, so S1 and S2 are similar and they are different from S3. 3 3Spectra l curve featuresSpectra l curve features Spectral curve features are computed by some algorithms based on the spectral curve of certain pixel or ground object, and it can describe shape and properties of the curve

25、. The main methods include direct encoding and feature band analysis. 3. 13. 1Direct encodingDirect encoding The important idea of spectral curve feature is to emphasize spectral curve shape, so direct encoding is a very convenient method, and binary encoding is used more widely. Its principle is to

26、 compare the attribute value at each band of a pixel with a threshold and assign the code of“0”or“1”according to its value. That can be expressed by 1ifX i T si o else Here, si is code of the ith band, X i is the original attribute value of this band, and T is the threshold. Generally, threshold is

27、the mean of spectral vector, and it can also be selected by manual method according to curve shape, sometimes median of spectral vector is probably used. Only one threshold is used in binary encoding, so the divided internal is large and precision is low. In order to improve the appoximaty and preci

28、sion, the quaternary encoding strategy is proposed in this paper. Its primary idea is as follows: ( 1 ) the mean of the total pixel spectral vector is computed and denoted by T0 , and the attribute is divided into two internal including X min , T 0 and T 0 , X min ; (2) the pixels located in the two

29、 internals are determined and the mean of each internal is got and donated by T and TR , so four internals are formed including X min , TL , T 0 , TR and TR , X min ; ( 3) each band is assigned one of the code sets 0, 1, 2, 3 according to the internal it is located; (4) to compute the ratio of match

30、ed bands number to the total band number as final matching ratio. It p roved that quaternary encoding could describe the curve shape more precisely. 3 If quarternary encoding is used, the ratio of the same region is smaller than binary encoding, but the ratio between different regions decreased dram

31、atically. So quarternary encoding is more effective in measuring the similarity between different pixels. Because direct encoding will disperse the continuous albedo into discrete code, the encoding result is affected by threshold obviously and will lead to information loss. Although its operation i

32、s very simple, it is only used to some applications requiring low precision, and the threshold should be selected according to different conditions. 3. 23. 2Spectral absorption or reflection featureSpectral absorption or reflection feature Differing from direct encoding in which all bands are used,

33、spectral absorption or reflection feature only emphasizes those bands where valleys or apexes are located. That means those bands with local maximum or minimum in spectral curve should be determined at first and then further analysis can be done. In general, albedo is used to describe the attribute

34、of a pixel, so those bands with local maximum are reflection apex and those with local minimum are absorption valley. After the location and related parameters are got, the detail analysis can be done. In general two methods are used, one is to give direct encoding and analysis to feature bands, and

35、 the other is to compute some quantitative index using feature bands and their parameters. 3.33.3Encoding of spectra l absorption or reflection featuresEncoding of spectra l absorption or reflection features The locations of feature bands are directly used in spectral feature encoding. The following

36、 will take absorption feature as an example. If one band is the location of absorp tion valley, its code will be “1 ”, otherwise its code is “0 ”. After the encoding is completed further matching and comparison can be done. Because of those uncertainties and errors in hyper spectral imaging process,

37、 the locations of feature bands perhaps move in near bands, and that will lead to low match ratio. In order to reduce the impact of band displacement, the extended encoding method is proposed and used in this paper. Its idea is that if the code of a certain band is“1”then the bands prior to and behi

38、nd it will be assigned the same code“1”, and then matching and analysis will be done. The similarity measure to code vector is matching by bit. The matching ratio is got by the ratio of matched bands to total band count. In this study, two match schemes are used. One is matching the code of all band

39、s and the other is only matching those feature bands. Based on above analysis, four schemes are used and compared. These are: ( 1) direct encoding to all bands and matching by all bands, and ( 2 ) direct encoding to all bands and matching only by feature bands, and ( 3) extended encoding and matchin

40、g by all bands, and ( 4 ) extended encoding and matching only by feature bands. From above analysis and comparison to spectral absorption and reflection feature encoding and matching, it can be found that although absorption and reflection band can describe the spectral properties of ground object,

41、effective matching operation should be used in order to overcome the impacts of noise, 4 band displacement and other factors. In practical applications, absorption and reflection can be used to extract thematic information and retrieve a certain type of object effectively. Based on spectral absorpti

42、on and reflection features, the spectral absorption index ( SA I) or spectral reflection index ( SR I) can be computed by wavelength,albedo of feature band and its left and right shoulders, and those indexes can describe spectral feature more precisely on some occasions. 4 Spectra l computation and

43、transformation features4 Spectra l computation and transformation features Both correlativity and mutual compensation exist in different bands of hyper spectral RS information, so many new features can be got by certain computation and combination to some bands and used to classification, informatio

44、n extraction and other tasks. 4. 14. 1Normalized difference of vegetation index (NDVI)Normalized difference of vegetation index (NDVI) NDVI plays very important roles in hyper spectral application. It can describe some fine information about vegetation such as Leaf Area Index (LA I) , ratio of veget

45、ation and soil, component of vegetation and so on. In some classifiers ( for example, ANN classifier) NDVI usually is used as an independent feature in classification. 4. 24. 2Derivative spectrumDerivative spectrum Derivative spectrum is also called as spectral derivative technique. One rank and two

46、 rank derivative spectrum can be computed by Equation. Each rank derivative spectrum can be computed using algorithms similar to above. After derivative computation is end, we can find that each type of ground object may have some features distinguished from other entities in a certain rank derivati

47、ve spectrum and that can be used to identify information. Sometimes derivative spectrum image can be used as the input of classifier directly. Although spectral derivative can provide new features in addition to original information, some new images will be formed after derivative operation and that

48、 will increase data volume dramatically. Formrank derivative spectrum, N - 2M bands will be formed, so how to process relationship between data volume and efficiency becomes a new question. 5 5Conclusions and discussionsConclusions and discussions In this paper, oriented to the demands of hyper spec

49、tral RS information processing to spectral features, the framework of spectral features is proposed and some major feature extraction algorithms and their applications are discussed, and some improvement, experiments and analysis are finished. From the studies in this paper, the following conclusion

50、s can be drawn: 1 ) Based on the extraction principle and algorithm, spectral features in hyper spectral RS information can be categorized into three levels: spectral curve features, spectral transformation and computation features and spectral similarity measure features. This framework is useful f

51、or further analysis and applications. 2) As the common style of pixel spectral vector, some features can be extracted and used. The 5 algorithm and computation of binary encoding is simple and easy but it will lead to loss of some detail information. Quaternary encoding can describe curve features w

52、ith highrescission and be used to matching, retrieval and other work. The reflection and absorption features based on spectral curve have wide applications in retrieval, thematic information extraction and other tasks, but effective matching strategy must be adopted in order to control errors. In th

53、is paper two new app roaches including extended encoding and matching and combined matching of reflectance and absorption features are proposed and it p roved that they can get better results than traditional methods in feature measure. 3) As the main computation and transformation features, NDV I a

54、nd derivative spectrum can provide new features participating in classification, extraction and other processing and extract those useful patterns and information hidden behind original data, so they are very useful in hyper spectral RS information processing. 4) For those spectra similarity measure

55、 indexes, Spectral Angle and SID are more effective than traditional indexes because they can measure the similarity more precisely, so they are usually used to classification, clustering and retrieval. Some topics about the feature extraction and application of spectral feature are discussed in thi

56、s paper. Our further studies will be focused on classification, object identification and thematic information extraction in hyper spectral RS information and the specific application modes of different spectral features in order to promote the development of hyper spectral RS application. 6 高光譜遙感信息

57、中的特征提取與應用研究高光譜遙感信息中的特征提取與應用研究 面向高光譜遙感信息處理和應用的需求，在高光譜遙感圖像的光譜特征可分為三個尺度：點規(guī) 模，塊規(guī)模和數(shù)量規(guī)模。根據(jù)不同的功能屬性和算法，它提出了點規(guī)模的特點，可以分為三個層次： 光譜曲線特征，光譜變換特征和光譜相似性度量功能。光譜曲線特征包括直接光譜編碼，反射和吸 收功能。光譜變換特征包括歸一化植被指數(shù)（紐卡斯爾），導數(shù)光譜和其他光譜的計算功能。光譜 相似性度量的功能包括光譜角（SA），光譜信息散度（SID）的光譜距離，相關(guān)系數(shù)等。分析這些算 法的幾個問題，關(guān)于特征提取，匹配和應用的基礎(chǔ)上進一步討論，它分析得到的第四紀編碼，光譜 角和 S

58、ID 可有效用于信息處理。 1 1 介紹介紹 高光譜遙感是在 1990 年的地球觀測系統(tǒng)（EOS），最重要的突破之一。它克服了傳統(tǒng)的天線和 多光譜 RS，如少帶量，寬波段范圍和粗糙的光譜信息表達的限制，可以提供寬窄帶，更帶量和良好 的光譜信息的遙感信息，還可以區(qū)分和識別地面光譜空間中的對象，因此高光譜遙感在資源，環(huán)境， 城市和生態(tài)領(lǐng)域得到廣泛應用。由于高光譜遙感是從傳統(tǒng)的遙感信息的信息采集和信息處理明顯不 同，有許多問題要在實踐中加以解決。最重要的問題之一是高光譜遙感圖像與標準光譜數(shù)據(jù)庫的光 譜遙感數(shù)據(jù)光譜特征提取和應用。如今，在光譜的研究主要集中在波段選擇和組合，圖像分類，混 合像元分解和他

59、人，和光譜特征的研究很少。在本文中，光譜特征的提取和應用將作為我們的中心 議題，以高光譜遙感應用提供一些有益的建議。 2 2 高光譜遙感數(shù)據(jù)光譜特征的框架高光譜遙感數(shù)據(jù)光譜特征的框架 一般情況下，高光譜遙感圖像可以表示空間光譜數(shù)據(jù)立方體（圖1）。覆蓋在這個多維數(shù)據(jù)集， 每一個面的形象，每個像素形成一個光譜，每個波段的光譜維矢量組成的地面物體的反照率，可以 通過光譜曲線（圖 2）可視化，矢量化。許多功能可以提取光譜向量或曲線，光譜特征的高光譜遙 感應用的關(guān)鍵和基礎(chǔ)。每個光譜數(shù)據(jù)庫的光譜曲線也可以用同樣的方法分析。雖然有一些算法來計 算的光譜特征，框架和體系還不夠明顯，所以我們想提出一個包括高光譜遙感圖像與標準光譜數(shù)據(jù) 庫的光譜遙感數(shù)據(jù)光譜特征的框架。 圖 1 高光譜圖像數(shù)據(jù)立方體圖 2 像素的反射光譜曲線 7 2.12.1 三個尺度的光譜特征三個尺度的光譜特征 據(jù)光譜特征提取算法的運作對象，可以分為三個尺度：點規(guī)模，塊規(guī)模和數(shù)量規(guī)模。點規(guī)模的 像素和其光譜曲線的經(jīng)營對象及一些有用的功能，可以從這個光譜的矢量（或光譜曲線）中提取。 一般情況下，高光譜遙感圖像每個像素的光譜向量作為處理對象。 塊規(guī)模為導向的圖像塊或地區(qū)。塊是一些像素的集合，它可以是同質(zhì)或異質(zhì)。同質(zhì)區(qū)域的圖像 分割，并在本地