Research On The Classification Method For Hyperspectral Images

The enhancement of spectral resolution is the trend in development of the optical remote sensing.High spectral resolution(hyperspectral for short) remote sensing is one of the significant technological breakthroughs for the observation to ground in the past 20 years,and is the advancing front technology of the current remote sensing.The hyperspectral's latent application receives widespread attention because of its high spectral resolution.Classification is an important means of information mining for hyperspectral imagery.At present, there are lots of hard classification methods,but their classification performances are not very perfect,or some methods themselves are to be improved.Traditional spectral unmixing methods are inefficient for the participation of unrelated classes and for the deficiency of spectral unmixing model.ISODATA is a non-supervised dynamic clustering algorithm which bases on statistical pattern recognition,it has a strong usability.Support vector machine (SVM),which is based on statistical learning theory,is a good promotion of high-dimensional nonlinear data processing tools.Its core is that the nonlinearity mapping competes for high dimension space with the sample data in the principle of the minimum structure risk.In this case,the techniques of hard classification and soft classification are researched mainly based on ISODATA algorithm and SVM theory in the paper.Firstly,the concepts of hyper spectral remote sensing and the pattern of characters of hyperspectral images are introduced.Hard classification,soft classification,supervised classification and unsupervised classification are introduced in terms of the state of the art,evaluation principle,and the technique problems.The basic theory of support vector machine is also introduced.All the works are helpful of developing the research of the paper.Secondly,the principle and achieving steps of fuzzy ISODATA algorithm are described.It combines the characteristics of hyper-spectral remote sensing image to carry on a hyper-spectral remote sensing image classification,and got the ideal classification results,moreover,the effect of some certain parameters on the output of classification have been analyzed as well.Also hard ISODATA classification algorithm and Fuzzy C-means are adopted in hyper-spectral image classification,ultimately three experimental results are compared and the results show that Fuzzy ISODATA algorithm's accuracy of the classification has an improved performance comparing to the above two types of algorithms.Thirdly,a kind of multi-classes SVM with character of twice classification are proposed.It works out the hard problem about the selection of optimal punishment factor and optimal weighting vector coefficient,and overcomes the shortcoming about the need of countless experiments to determinate punishment factor.Experiments show that the proposed method,compared with 1-a-1 SVM, improves the classification accuracy of corn and soybean greatly,with very high speed for selecting punishment factor.Finally,the unsupervised soft classification method based on LSMM, supervised soft classification method based on SVM,and soft classification method based on nonlinear SVM are researched.Theory analysis and comparison experiments show that LSMM based classification method is of simple algorithm and so easy to operate,while SVM based classification method can flexibly tackle inseparable problem with high classification accuracy.Experiments show that SVM has good performance both in hard classification and in soft classification.