Research On Feature Mining And Classification For Hyperspectral Remote Sensing Image

Abundant spectral characteristic information can be provided by hyperspectral remote sensing image.These information can represent the specific properties of different objects,and it is one of the important ways for human society to understand the earth.At present,the information of hyperspectral data sets is widely used in the fields of modern military,mineral exploration,precision agriculture and environmental disaster monitoring.The study of hyperspectral image classification is one of the most active subjects in the field of target recognition and remote sensing.Advanced machine learning algorithm is the main research direction to solve classification problems.However,there are still great difficulties and challenges in achieving high precision classification.The main problems are described below.First,the spectral dimension of hyperspectral data sets is too high,and the data redundancy is large.Especially for the spectral features with high similarity,how to carry out effective feature mining is the key issue.Second,computational complexity is high and marker samples are lacking in the classification process,how to solve the contradiction between small sample and high dimension and develop efficient classification model is the key issue.Third,the statistical properties of the intraclass and interclass are more complex in hyperspectral data sets,how to enhance the image features by increasing the class gap and build relevant classifiers to improve the classification effect is the key issue.For the above issues,the characteristics of hyperspectral image data sets are analyzed in this dissertation.The feature mining method and efficient classification models of hyperspectral image are studied,and the proposed models are analyzed and evaluated in the real hyperspectral image data sets.The main results of this dissertation are as follows.(1)Due to the high dimensionality of hyperspectral data sets and the strong correlation between adjacent bands,there are a lot of redundant information in the hyperspectral data sets,The spectral feature mining method for hyperspectral image based on cumulative variation quotient(CVQ)is proposed in this dissertation.The method is improved according to the characteristics of statistical coefficient of variation,by the quotiens of the cumulative variation values between the intraclass and the interclass,the influence of each band on the classification effect is analyzed,the feature mining of hyperspectral images is realized by selecting the efficient band according to its contribution degree,and it is verified that the proposed method is not affected by the order of the samples.Through the experiment comparison analysis,the spectral feature mining method of hyperspectral image based on cumulative variation quotient can extract spectral features more effectively,and reduce the redundancy of information.It provides necessary technical support for the subsequent processing of hyperspectral image.(2)Due to the high complexity of hyperspectral image classification model and the small number of labeled samples,the ensemble extreme learning machine with cumulative variation quotient for hyperspectral image classification is proposed in this dissertation.Weighted random grouping is used to further mine the characteristics of hyperspectral data sets,and then the comprehensive evaluation is proposed on the classification conclusions of ensemble extreme learning machine.Through the experiment comparison analysis,the proposed method optimizes band selection and sample selection,and it also reduces the computational complexity of the model.The classification accuracy is improved,and the problem of sample deficiency is solved well.(3)Due to the complex structural features of the intraclass and interclass and the low recognition rate of small samples,the bidirectional depth recursive classification model of hyperspectral image based on feature enhancement is proposed.The model is made up of the feature enhancement part and the bidirectional deep recursive neural network part.In combination with the above research results,we continue to study from the perspective of feature vector sequence continuity,and the remote sensing image information enhancement technology is used to improve the characteristic gap between sample classes.This proposed method can better mine the characteristics of hyperspectral data sets.According to the characteristics of feature vector sequence,a matching classification model is constructed.Through the experiment comparison analysis,the proposed method fully considers the characteristic gap between sample classes,and then each output is fused into the hyperspectral input's feature vector to enhance the classification properties of hyperspectral image.The poor classification effect of small samples in hyperspectral data sets is improved,and the classification accuracy is further improved.There are 58 Figures,16 Tables and 164 References.