Band Selection And Semi-supervised Learning-based Hyperspectral Classification

Recently,remote sensing images,especially hyperspectral images,are playing a more and more important role in the scope of agriculture,environment monitoring,military as well as outer space explorations.Thus,analyzing or processing this kind of data attracts more attention as time goes by.However,due to the high dimensionality and the noise as well as the curve of dimension phenomenon of them,we are still in great need of faster and more accurate algorithms.Lots of scholars proposed different algorithms to analyze hyperspectral data from different aspects.However,most of the algorithms have drawbacks to some extent,e.g.,some of them judged the importance of the spectral bands with short information,while others ignored the spatial information of the data,so that they did not achieve a better performance.In view of the above two main problems,this paper studies some classification algorithms for hyperspectral images from various aspects.The contributions mainly include:1)A method called dynamic local cluster ratio based band selection algorithm for hyperspectral images(DLCR)is proposed.First,we need to define to what extent one spectral band is different from others,which is accomplished by a threshold and the hyperspectral data.Second,in the procedure of clustering spectral bands,the global information is prevented from affecting too much,so that the number of the isolated bands is low.Third,in the final ranking step,bands are dynamically added to the result instead of ranking them once and for all.The experiments suggest the accuracy gain especially with few expected bands.2)A method called novel location-based DNA matching algorithm for Hyperspectral Image(LDNAMA)is proposed.First,this algorithm is based on evolutional algorithms,and the coding for spatial information is added first in this work,what’s more,the length of the total coding is cut down here.Then the elite-preserving strategy is presented along with a specifically designed method for mutation and crossover operations.In the last step,after matching and classifying every pixel,the final result should be modified with the help of the locations of labeled data,which will achieve a better performance.3)A method called fusion area based semi-supervised learning for hyperspectral images(FASSL)is proposed.First,the several spatial maps are extracted and fuse into a single map using the seed-growing algorithm.Different spatial maps aim to present different ranges of brightness and darkness,so that the single fusion map can better represent the data.Second,the classification is conducted merely on spectral information,but the fusion map is used to enlarge the training set in a semi-supervised view to deal with the lack of labeled samples.In the end,the results can be further smoothed similar to the former work.