Research On Band Selection For Hyperspectral Imagery

With the development of remote sensing technology and imaging spectrometer,hyperspectral imagery (HSI) applications are more and more widely. However, HSI has anumber of bands and a huge amount of data so that it has brought great difficulties for theclassification and recognition of HSI, such as high redundancy information, large data storagespace, long processing time and proning to the curse of dimensionality phenomenon becauseof a large number of bands of HSI. Therefore, it is necessary to reduce the dimension of HSIto reduce the amount of data and save resources in ensuring terrain classification recognitionrate. Feature extraction and band selection are the two main dimensionality reduction of HSI.Feature extraction is a complexity algorithm with the large amount of calculation, and makessome kind of transformation to achieve the purpose of dimensionality reduction. It changesthe physical meaning of the original data and is not conducive to data interpretation. Incontrast, band selection is to select a subset of bands from all of bands of HSI which plays amajor role in HSI. Band selection not only greatly reduces the dimension of HSI, but alsorelatively intacts usefull information. It is of more special meaning.Band selection of HSI is a very complex band combination optimization problem. Theselected band combination requires better performances which are a large amount ofinformation, less correlation and better class separability. Band selection techniques includetwo aspects: one is the criterion function, and the other is the search method. Most of theexisting band selection methods can not take into account both time efficiency and refinedsolution efficiency, so it is urgent to study a band selection method which has time efficiencyand refined solution efficiency. Based on this, we do in-depth research on the basis of theresults of previous studies about band selection of HSI. The main contents are as follows:1. Band selection for HSI based on the combination of genetic algorithm and ant colonyalgorithm. In-depth study of the genetic algorithm and ant colony algorithm, we know thatgenetic algorithm has rapidity, randomness and global convergence and ant colony algorithmhas positive feedbackness and high efficiency of solving. In the algorithm, the geneticalgorithm is used to search for some better solutions quickly which are initialized theinformation list of ant colony algorithm, and then ant colony algorithm can effectively searchfor the best solution. In the part of the genetic algorithm, quaternary encoding is employed, that makes encoding/decoding and genetic operation simple and uses less space of computer.In the part of ant colony algorithm, subspace division is used to deal with hyperspectralimagery, which reduces the range of ants search, improves the search efficiency, and reducesthe correlation and redundancy of the output band of hyperspectral imagery. The algorithmthat makes good use of the advantages of both genetic algorithm and ant colony algorithm andovercomes their defects is a fewer time-consuming and better performace-restraining methodfor band selection. AVIRIS imagery was used for experiment with the proposed algorithm,which proved that this algorithm of hyperspectal dimension reduction is effective in terms ofband selection performance and execution time consumption.2. Artificial physics optimization (APO) algorithm combined band selection for HSI.APO algorithm which is a new group random search algorithm has better global searchcapability, not easy to fall into local optimum, and has a stable and fast convergence andbetter robustness. The algorithm has been in single/multi-objective optimization problemsuccessfully, so APO algorithm is used to select band subsets for HSI in this paper. Due toAPO algorithm is a relatively new algorithm, and the mass function and gravitationalparameters have a critical impact on the performance of the algorithm, therefore, this papermade a thorough study of these two important factors, respectively. In this paper, thecombination between-class separability and information of band groups is used as fitnessfunction, and subspace division is used to deal with hyperspectral image before band selectionin order to reduce the correlation and redundancy of the output band of hyperspectral image.AVIRIS imagery was used for experiment with the proposed algorithm and other classicalalgorithms which are ant colony algorithm, genetic algorithm and particle swarm optimization,and it turned out that this algorithm is more effective in both band chosen performance andcomputing time consumption than others.3. Fast SGA is used for the process of HSI. The original SGA which is a spectralendmember selection method mainly calculates the single body combined by choosedendmembers to select endmembers. This algorithm can get stable endmember combinationand solve the N-FINDR algorithm unstable, but its calculation is very explex. The fast SGAthat uses distance calculation to replace volume calculation improves the efficiency of SGA.In addition, due to the endmember selection can transplant band selection, the articleproposed an idea of the use of the fast SGA for band selection. In this paper, three algorithms respectively have their advantages. In the case of the morefeatures, Band selection for hyperspectral imagery based on combination of genetic algorithmand ant colony algorithm can get band-subset with large amount of information. Bandselection for hyperspectral imagery which based on artificial physics optimization algorithmnot only has a distinct advantage in the band selection efficiency, but also gets a band-subsetconducive to the classification of the image. Fast simplex growing algorithm (SGA) algorithmhas a prominent advantage of the efficiency of the search time.