Font Size: a A A

Improved Empirical Mode Decomposition Algorithm And Its Application To Hyperspectral Image Classification

Posted on:2015-08-16Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z HeFull Text:PDF
GTID:1108330479478773Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Hyperspectral imaging, which first emerged in the 1980 s, is a huge leap of remotesensing technology. Compared to the traditional visible or multi-spectral image, hyper-spectral image can provide more extensive information on land-cover and surface featurespectrum. Since hyperspectral image exhibits promising ability in aerospace and mili-tary fields, it has attracted greatest attention at home and abroad and become one of thehottest research topics nowadays. Intrinsically, hyperspectral data is “non-linear” and“non-stationary” signal. Therefore, how to make full use of the collected image infor-mation and improve the classification performance are di?cult problems. The empiricalmode decomposition method(EMD), which emerged in recent years, has inherent ad-vantages in dealing with the complicated “non-linear” and “non-stationary” hyperspectraldata. Unfortunately, there exists no complete and recognized theoretical foundation up tillnow. Therefore, how to improve EMD from theoretical aspect, especially end effects mit-igation and “overshoot” and “undershoot” phenomenon elimination, are major challengesfor researchers. Generally, feature extraction and classification of hyperspectral imageare the key points in this dissertation. On one hand, the traditional EMD is improvedby gray model and optimization theory in aspect of end effects mitigation and envelopecalculation. The improved EMD can then be applied to extracting spectral and spatialfeature of hyperspectral data. On the other hand, spectral-spatial hyperspectral imageclassification methods are proposed based on sparse representation classifier(SRC) andsuperpixel image segmentation. Major contributions and novelties of our dissertation canbe summarized as following:of one-dimensional EMD can be mitigated by extending 2 points of the signal generatedin the sifting process.End effects mitigation methods of one-dimensional EMD based on single variablegray model(GM-EMD) is studied. The end effect is one of the open problems of EMD.The cause of end effects lies in that the extreme points near two ends of the signal cannotbe accurately determined. First of all, the importance of extremum and envelope to EMDis proven. Secondly, the existing single variable gray model can be improved in the dis-cretization process of differential equation, and then the GM-EMD are proposed, whichtakes full advantage of the single variable gray model, i.e., requiring few original data,providing convenient calculation and achieving accurate short-term prediction accuracyin condition of not changing the properties of the original signal. As such, the end effectsEnvelopes construction methods of one-dimensional EMD based on alternating di-rection method of multipliers(ADMM) is studied. The phenomenon of “overshoot” and“undershoot” will appear in case the spline interpolation is applied to construct the en-velopes, and the cause of “overshoot” and “undershoot” lies in that there exists no con-straints on non-extreme points in process of interpolation. Therefore, strict mathemati-cal inequalities which the upper-/lower envelopes should satisfy are constructed and theoptimization problem can be solved by ADMM. Compared to spline interpolation, theADMM-based method can effectively eliminate the annoying “overshoot” and “under-shoot” phenomenon.End effects mitigation methods of bi-dimensional EMD based on multi-variable graymodel(GM-BEMD) is studied. The end effects of bi-dimensional EMD(BEMD) arecaused by inaccurate extremum near the image boundary. Main processes of the proposedGM-BEMD are as follows: the image is extended by complex-Simpson-rule-based multi-variable gray model, which can then be decomposed by traditional BEMD. As such, theextended bi-dimensional intrinsic mode function(BIMF) and residue can be obtained andthe final decomposition results can be determined by cutting out the corresponding partsof extended results. It is notable that the end effects can be mitigated effectively by GM-BEMD.Spectral-spatial hyperspectral classification methods based on kernel-based sparsemultitask learning classifier(KSMTLC) is studied. The proposed KSMTLC can dealwith the spectral and spatial features simultaneously in framework of SRC, in which theoptimization problem is solved by accelerated proximal gradient(APG). The classifica-tion results are much better than single-feature-based methods.Spectral-spatial hyperspectral classification methods based on support vector ma-chine(SVM) and superpixel image segmentation is studied. the hyperspectral image isfirst classified by spectral-feature-based methods, and then the final spectral-spatial clas-sification results can be obtained by spatial post-processing methods. In general, SVMis adopted to spectral classification, whereas superpixel image segmentation is utilized tospatial post-processing. The proposed method in this study is simple but proposed forthe first time. It is notable that the spatial post-processing benefits the improvement ofhyperspectral classification results.
Keywords/Search Tags:empirical mode decomposition, hyperspectral image, classification, support vector machine, sparse representation classifier
PDF Full Text Request
Related items