基于集成学习约束能量最小化的高光谱目标检测算法研究
|
摘要
提出了一种基于集成学习约束能量最小化(E-CEM)的高光谱图像目标检测算法。传统的高光谱检测算法通常是基于约束最小二乘法或基于高斯先验下的假设检验算法获得,然而真实环境中捕获的高光谱数据通常具有很强的非线性及非高斯特性,此时传统算法通常难以获得满意的检测效果。虽然核方法一定程度上能使传统算法获得较强的非线性表达能力,但核方法本身极易受到核函数参数的选择而表现出性能不稳定的现象。E-CEM在传统的约束能量最小化算法的基础上结合集成学习思想,使其在提升非线性光谱表达能力的同时提升检测的稳定性和稳健性。仿真高光谱图像和真实高光谱图像的实验结果都表明所提方法提升了CEM算法及其他经典算法的检测性能。
This paper proposes a new method named ensemble-learning constrained energy minimization(ECEM)for hyperspectral target detection.Traditional hyperspectral target detection methods are usually designed based on constrained least square regression method or hypothesis testing method with Gaussian distribution assumption.However,hyperspectral data captured in real environment often show strong non-linearity and nonGaussianity.Under such condition,it is hard for those classical methods to obtain a satisfied performance.Although kernel trick is able to extend some traditional methods to their nonlinear form,the kernel based methods are extremely unstable due to the sensitivity of their parameter settings.The proposed E-CEM is designed with the idea of ensemble learning based on the CEM algorithm,which gives an improvement on both of its spectral representation ability and robustness.Experiments on synthetic hyperspectral image and real hyperspectral image demonstrate that the proposed method can enhance the detection performance of CEM algorithm and other classical detection algorithms.
This paper proposes a new method named ensemble-learning constrained energy minimization(ECEM)for hyperspectral target detection.Traditional hyperspectral target detection methods are usually designed based on constrained least square regression method or hypothesis testing method with Gaussian distribution assumption.However,hyperspectral data captured in real environment often show strong non-linearity and nonGaussianity.Under such condition,it is hard for those classical methods to obtain a satisfied performance.Although kernel trick is able to extend some traditional methods to their nonlinear form,the kernel based methods are extremely unstable due to the sensitivity of their parameter settings.The proposed E-CEM is designed with the idea of ensemble learning based on the CEM algorithm,which gives an improvement on both of its spectral representation ability and robustness.Experiments on synthetic hyperspectral image and real hyperspectral image demonstrate that the proposed method can enhance the detection performance of CEM algorithm and other classical detection algorithms.
引文
[1]MANOLAKIS D,MARDEN D,SHAW G A.Hyperspectral image processing for automatic target detection applications[J].Lincoln Laboratory Journal,2003,14(1):79-116.
[2]SHAW G A,BURKE H K.Spectral imaging for remote sensing[J].Lincoln Laboratory Journal,2003,14(1):3-28.
[3]SHI Z W,QIN Z,YANG S,et al.Spatial multiple materials detection in hyperspectral imagery[C]∥3rd International Conference on Awareness Science and Technology.Dalian:[s.n.],2011:76-80.
[4]WINTER E M,MILLER M A,SIMI C G,et al.Mine detection experiments using hyperspectral sensors[J].Proc SPIE,2004,5415:1035-1041.
[5]DIPIETRO R S,TRUSLOW E,MANOLAKIS D G,et al.False-alarm characterization in hyperspectral gas-detection applications[J].Proc SPIE,2012,8515:85150I.
[6]KRUSE F A,LEFKOFF A B,BOARDMAN J W,et al.The spectral image processing system(SIPS)—interactive visualization and analysis of imaging spectrometer data[J].Remote sensing of environment,1993,44(2-3):145-163.
[7]CHANG C I.An information-theoretic approach to spectral variability,similarity,and discrimination for hyperspectral image analysis[J].IEEE Transactions on Information Theory,2000,46(5):1927-1932.
[8]FARRAND W H,HARSANYI J C.Mapping the distribution of mine tailings in the Coeur d'Alene River Valley,Idaho,through the use of a constrained energy minimization technique[J].Remote Sensing of Environment,1997,59(1):64-76.
[9]HARSANYI J C.Detection and classification of subpixel spectral signatures in hyperspectral image sequences[D].Baltimore:University of Maryland Baltimore County,1993.
[10]KAY S M.Fundamentals of statistical signal processing[M].New Jersey:Prentice Hall,1998.
[11]MANOLAKIS D G,INGLE V K,KOGON S M.Statistical and adaptive signal processing:spectral estimation,signal modeling,adaptive filtering and array processing[M].Boston:McGraw-Hill,2000.
[12]KRAUT S,SCHARF L L.The CFAR adaptive subspace detector is a scale-invariant GLRT[J].IEEE Transactions on Signal Processing,1999,47(9):2538-2541.
[13]KRAUT S,SCHARF L L,MCWHORTER L T.Adaptive subspace detectors[J].IEEE Transactions on Signal Processing,2001,49(1):1-16.
[14]KWON H,NASRABADI N M.Hyperspectral target recognition using kernel spectral matched filter[C]∥Conference on Computer Vision and Pattern Recognition Workshops.Washington:[s.n.],2004:425-430.
[15]JIAO X L,CHANG C I.Kernel-based constrained energy minimization(K-CEM)[J].Proc SPIE,2008,6966:69661S.
[16]ZOU Z X,SHI Z W,WU J,et al.Quadratic constrained energy minimization for hyperspectral target detec-tion[C]∥Geoscience and Remote Sensing Symposium(IGARSS).Milan:[s.n.],2015:4979-4982.
[2]SHAW G A,BURKE H K.Spectral imaging for remote sensing[J].Lincoln Laboratory Journal,2003,14(1):3-28.
[3]SHI Z W,QIN Z,YANG S,et al.Spatial multiple materials detection in hyperspectral imagery[C]∥3rd International Conference on Awareness Science and Technology.Dalian:[s.n.],2011:76-80.
[4]WINTER E M,MILLER M A,SIMI C G,et al.Mine detection experiments using hyperspectral sensors[J].Proc SPIE,2004,5415:1035-1041.
[5]DIPIETRO R S,TRUSLOW E,MANOLAKIS D G,et al.False-alarm characterization in hyperspectral gas-detection applications[J].Proc SPIE,2012,8515:85150I.
[6]KRUSE F A,LEFKOFF A B,BOARDMAN J W,et al.The spectral image processing system(SIPS)—interactive visualization and analysis of imaging spectrometer data[J].Remote sensing of environment,1993,44(2-3):145-163.
[7]CHANG C I.An information-theoretic approach to spectral variability,similarity,and discrimination for hyperspectral image analysis[J].IEEE Transactions on Information Theory,2000,46(5):1927-1932.
[8]FARRAND W H,HARSANYI J C.Mapping the distribution of mine tailings in the Coeur d'Alene River Valley,Idaho,through the use of a constrained energy minimization technique[J].Remote Sensing of Environment,1997,59(1):64-76.
[9]HARSANYI J C.Detection and classification of subpixel spectral signatures in hyperspectral image sequences[D].Baltimore:University of Maryland Baltimore County,1993.
[10]KAY S M.Fundamentals of statistical signal processing[M].New Jersey:Prentice Hall,1998.
[11]MANOLAKIS D G,INGLE V K,KOGON S M.Statistical and adaptive signal processing:spectral estimation,signal modeling,adaptive filtering and array processing[M].Boston:McGraw-Hill,2000.
[12]KRAUT S,SCHARF L L.The CFAR adaptive subspace detector is a scale-invariant GLRT[J].IEEE Transactions on Signal Processing,1999,47(9):2538-2541.
[13]KRAUT S,SCHARF L L,MCWHORTER L T.Adaptive subspace detectors[J].IEEE Transactions on Signal Processing,2001,49(1):1-16.
[14]KWON H,NASRABADI N M.Hyperspectral target recognition using kernel spectral matched filter[C]∥Conference on Computer Vision and Pattern Recognition Workshops.Washington:[s.n.],2004:425-430.
[15]JIAO X L,CHANG C I.Kernel-based constrained energy minimization(K-CEM)[J].Proc SPIE,2008,6966:69661S.
[16]ZOU Z X,SHI Z W,WU J,et al.Quadratic constrained energy minimization for hyperspectral target detec-tion[C]∥Geoscience and Remote Sensing Symposium(IGARSS).Milan:[s.n.],2015:4979-4982.