Band Selection Based On Chaotic Cuckoo Search Algorithm And Semi-Supervised Classification For Hyperspectral Image

The advent of hyperspectral remote sensing promoting by the development of imaging spectrometer,provides a new research direction for remote sensing science and technology.Hypercpectral image has unity of graph and spectrum,it can get continuous and fine spectral curves combined a continuous imaging spectral range and high spectral resolution.That provides basis for fine detection of hyperspectral remote sensing.However,hypercpectral image brings difficulties and challenges to the image processing,such as more band numbers,strong correlation in bands,and hughes phenomenon,that can lead to low classification accuracy of hypercpectral remote sensing image.Therefore,how to carry on the dimension reduction of hyperspectral remote sensing images,and improve image classification accuracy,is of great significance.Aiming at the search strategy's choice of band selection for hyperspectral remote sensing image,the chaotic cuckoo search algorithm is proposed.As a newly proposed evolutionary algorithm,cuckoo search algorithm has the preferable global optimization ability,and chaotic map could improve the local optimization ability of evolutionary algorithm.In the proposed algorithm,standard cuckoo search algorithm was utilized to obtain some better solutions by optimizing the whole population,and chaotic map could quickly search for the best solution.Experimental results demonstrate that the search ability of chaotic cuckoo search algorithm is better than genetic algorithm,particle swarm optimization algorithm,and standard cuckoo search algorithm.To solve the problem of more band numbers and strong correlation in bands,the proposed chaotic cuckoo search algorithm is utilized to band selection for hyperspectral remote sensing image.The band selection of hyperspectral remote sensing image is ed into the search process of the cuckoo bird's nest location.First,band selection which numbers is not fixed is processed for decreasing band target numbers.Then determining the band target numbers according to the hughes phenomenon.Last,selecting bands for hyperspectral remote sensing image.The experimental results show that the proposed algorithm is finally selected band number more primitive band decreases the number of close to 98%of the feature numbers,the higher classification accuracy can be obtained.For low classification accuracy of image because of less labeled samples,a tri-training semi-supervised collaborative classification method combined spatial and spectral information is proposed.The process of tri-training semi-supervised collaborative method is combined spatial information and spectral angle mapper(SAM),includes the second selection of unlabeled samples by 8 neighborhood information of labeled sample and SAM between labeled sample and unlabeled sample,and integrated classification with three random forest classifier.The experimental results show that the selection method combined spatial and spectral information introduce the accuracy of unlabeled samples,and the proposed method greatly improve the generalization ability of the classifier,and the classification accuracy of the hyperspectral remote sensing image.