1、Exploiting Manifold Geometry in Hyperspectral Imagery441-454共14頁46篇參考文獻(xiàn)， google scholar引用183次目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery2目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery3目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery4關(guān)鍵詞1.Airbone Visible

2、Imaging Spectrometer(AVIRIS), 2.Bidirectional reflectance distribution function(BRDF),pression, 4.Cuprite, 5.Geodesic distance, 6.hyperspherical, 7.Isometric mapping(ISOMAP), 8.Land-cover classification9.Linear mixing10.Local linear embedding(LLE)11.Maniflod coordinates12.Nonlinearity13.PROBE214.Vir

3、ginia Coast Reserve目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery5Abstract A new algorithm for exploiting the nonlinear structure of hyperspectral imagery is developed and compared against the de facto standard of linear mixing. Idea：把scene分成小的tiles，每個(gè)tile用isomap降維，然后再對齊tiles，重新補(bǔ)全scen

4、e Experiment：classification Compare:MNF-minimum noise fraction最小噪聲分離目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery69部分1.Background and Introduction2.Manifold Coordinate Systems and Hyperspectral Data3.Results for Hyperspectral Data Compression:Manifold Coordinates Versus MNF4.Manifold

5、 Representations of Hyperspectral Scene Tiles5.Scaling to Large-Scale Remote Sensing Data Sets6.Results: Aligning Tile Manifolds7.Automating the alignment of Tile Manifolds8.Computational Scaling and Optimization9.Conclusion目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery71Background an

6、d Introduction “best band”方法依賴于所感興趣的類別會出現(xiàn)在窄波段 Linear mixing：假設(shè)：場景中每個(gè)像素可分解成有限個(gè)連續(xù)連續(xù)端元(endmembers)目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery81Background and Introduction 特別是land-cover classification中，產(chǎn)生非線性的根源 1、本質(zhì)是散射的非線性，描述為雙向反射分布函數(shù)，最小反射波長展示最大非線性化 2、特別是海岸環(huán)境下，海岸濕地，水變化的表示 其他：多次散射，次元成份的

7、多樣性(heterogeneity of subpixel consistents)目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery91Background and Introduction Isomap雖然全局最優(yōu)，但僅適用于小數(shù)據(jù)集 1時(shí)間復(fù)雜度O(N3)計(jì)算量較大 2空間復(fù)雜度O(N2)存儲需求高 LLE計(jì)算量O(N2logN) SPE:最快速估計(jì)流形坐標(biāo)的方法，計(jì)算量O(N)，但是簡化導(dǎo)致退化解。 策略：divide -conquer-merge目錄Title:Exploiting Manifold Geomet

8、ry in Hyperspectral Imagery102 Manifold Coordinate Systems and Hyperspectral Data A目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery112 Manifold Coordinate Systems and Hyperspectral Data A目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery122 Manifold Coordinate Systems and Hyp

9、erspectral Data A目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery132 Manifold Coordinate Systems and Hyperspectral Data A 局部度量 1 譜角spectral angle可以消除各種光照的影響 2歐氏距離 3Mahalanobis距離常用于hyperspectral分類和檢測，需要計(jì)算covariance matrix 局部變量的選擇并不重要目錄Title:Exploiting Manifold Geometry in Hyperspectral I

10、magery142 Manifold Coordinate Systems and Hyperspectral Data B Manifold Formal Definitions目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery152 Manifold Coordinate Systems and Hyperspectral Data C. ISOMAP Floyd算法O(N3)用于dense graph Dijkstra算法用于sparse graph(指邊數(shù)遠(yuǎn)少于點(diǎn)數(shù)平方) 孤立點(diǎn)處理： 假設(shè)：孤立點(diǎn)在主分布上不連通

11、，但在孤立點(diǎn)包內(nèi)也許是完全連通的 方法：把最近的孤立點(diǎn)連接起來，然后建立主分布與孤立點(diǎn)包的最近連接，而主分布的測地距離不變目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery162 Manifold Coordinate Systems and Hyperspectral Data D LLE 鄰域點(diǎn)線性組合 流形坐標(biāo) 重構(gòu)誤差 復(fù)雜度O(N2logN)目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery173 Results for Hyperspect

12、ral Data Compression:Manifold Coordinates Versus MNF目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery184Manifold Representations of Hyperspectral Scene Tiles目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery194Manifold Representations of Hyperspectral Scene Tiles目錄Title:Exploi

13、ting Manifold Geometry in Hyperspectral Imagery204Manifold Representations of Hyperspectral Scene Tiles目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery215 Scaling to Large-Scale Remote Sensing Data Sets 三種對齊策略: 1用tile和tile之間相同的tie-points來擴(kuò)展tile。 Tile到tile的變換公式： 2從每個(gè)tile中隨機(jī)找s個(gè)點(diǎn)構(gòu)成backbone

14、。 tile到backbone流形的變換： 3Ti中的點(diǎn)n，在原空間Tj上找k鄰域。即n用Tj的k鄰域線性表示，Ti中的fi個(gè)點(diǎn)用LLE解在Tj上的流形坐標(biāo)。目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery225 Scaling to Large-Scale Remote Sensing Data Sets目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery236Results: Aligning Tile Manifolds目錄Title:Explo

15、iting Manifold Geometry in Hyperspectral Imagery246Results: Aligning Tile Manifolds目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery257Automating the alignment of Tile Manifolds目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery268Computational Scaling and Optimization G：tile數(shù)，

16、N：像素?cái)?shù) 單個(gè)tile計(jì)算 相似圖內(nèi)存需求 拼接計(jì)算目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery279 Conclusion Tile的大小 鄰域的大小 人造痕跡的改進(jìn)目錄Title:Exploiting Manifold Geometry in Hyperspectral Imagery28后續(xù)工作1.Bachmann, CM. 2006. Improved manifold coordinate representations of large-scale hyperspectral scenes. IEEE

17、Transactions on Geoscience and Remote Sensing. 44:2786-2803.2.Thomas L. Ainsworth, Charles M. Bachmann, Robert A. Fusina: Local intrinsic dimensionality of hyperspectral imagery from non-linear manifold coordinates. IGARSS 2007: 1541-15423.Charles M. Bachmann, Thomas L. Ainsworth, Robert A. Fusina:

18、Automated Estimation of Spectral Neighborhood Size in Manifold Coordinate Representations of Hyperspectral Imagery: Implications for Anomaly Finding, Bathymetry Retrieval, and Land Applications. IGARSS (1) 2008: 56-574.Charles M. Bachmann, Thomas L. Ainsworth, Robert A. Fusina, Marcos J. Montes, Jeffrey Bowles, Daniel Korwan, David Gillis: Bathymetric Retrieval From Hyperspectral Imagery Using Manifold Coordinate Representations. IEEE T. Geoscience and Remote Sensing 47(3): 884-897 (2009)5.Charles M. Ba