Hyperspectral Band Selection Methods Based On Artificial Bee Colony Optimization

With the development of remote sensing technology,hyperspectral images have been widely used in many fields.A large number of bands data in hyperspectral images can provide rich information for many tasks such as detecting ground objects and targets recognition.However,with the rapid increase in the number of bands,there are many problems such as information redundancy and difficulty in obtaining object lables when applying hyperspectral data.Without the guidance of label information,how to effectively select a small number of bands from a large number of hyperspectral data has become one of the research hotspots in the field of hyperspectral remote sensing.In view of this,the thesis studies the theory and method of unsupervised band selection based on artificial bee colony algorithm(ABC),by combing the characteristics of hyperspectral remote sensing images.(1)Aiming at the problem that it is difficult to obtain object labels for hyperspectral images,an artificial bee colony band selection algorithm based on pseudo-labels generation is propsoed.Firstly,using the super-pixel centers to replace traditional pixel points,a hypergraph evolutionary clustering method with low computational cost is developed,so that the high-quality pseudo-labels can be generated based on the results of clustering hyperspectral data;Then,according to the generated pseudo-labels,a supervised band selection algorithm based on artificial bee colony is proposed by defining the classification accuracy as optimization index.In addition,in order to ensure the accuracy of the generated pseudo-labels,a noise filtering index based on mesh division is designed.(2)A multi-task bee colony band selection algorithm based on variable-scale clustering is proposed to get multiple band subsets of different sizes at the same time.Firstly,a variable-scale clustering method based on gradual elimination of the worst class is introduced to model the hyperspectral band selection problem as a multi-task optimization problem.Following that,an ABC-based multi-task multi-microcluster band selection algorithm with variable code length is developed to simultaneously select multiple band subsets with different sizes from multiple clustering results.In addition,considering the unlabeled characteristic of band selection problem,an individual(or band subset)evaluation model based on manifold preservation is presented.With the help of the parallel collaborative capability of the multi-task evolutionary optimization method,the proposed algorithm can simultaneously obtain multiple high-quality band subsets of different scales.(3)A multi-objective ABC(Artificial Bee Colony)method based on multi-strategy integration is proposed by regarding the trade-off between band correlation and redundancy.Firstly,using the discrete coefficient and the cross-correlation function to describe the amount of information in a band subset and the correlation between bands respectively,a multi-objective hyperspectral band selection model integrating the information amount and correlation is established.Secondly,a multi-strategy integrated multi-objective ABC algorithm(MABC-BS)is proposed to deal with the model above.New operators,including as the multi-direction search guided by diversity and convergence,the following probability based on crowding degree and the adaptive mutation,are proposed to strengthen the search ability of MABC-BS.These band selection algorithms above are applied to many typical hyperspectral data,including Indian Pines,Pavia University and Salinas,and are compared with several existing representative band selection algorithms.Experiments verify the effectiveness of the proposed three algorithms.The research results of the thesis enrich the theory of hyperspectral band selection and broaden the application fields of artificial bee colony optimization algorithms.The thesis includes 42 figures,13 tables and 140 references...