Research On Dimension Reduction Method Of Hyperspectral Image For Classification Of Vegetation And Similar Color Features

Unlike other remote sensing images,hyperspectral images contain rich spectral information of ground objects,and their ability to finely analyze ground objects has been significantly improved.Vegetation is one of the ground objects that play an important role in the background of the feature environment.Some special man-made objects are hidden by means of imitating the colors of the vegetation background.Both spectra for some bands are similar,and the phenomenon of "Same-Spectrum Foreign Matter" is formed in these similar bands.This phenomenon greatly increases the difficulty of hyperspectral image classification,resulting in a large amount of "salt and pepper noise" in the classification results.On the other hand,hyperspectral image data has the characteristics of high dimension and high correlation,which contains a lot of redundant information,which can easily lead to the "Hughes Phenomenon".It is an open challenge in the field of hyperspectral image application that how to effectively select the features of ground objects,avoid the influence of some spectral feature similar bands that produce the phenomenon of "Foreign Objects In The Same Spectrum",and eliminate the redundant information of hyperspectral image on the basis of ensuring the classification accuracy of vegetation and ground objects with similar colors.This paper plans to solve the shortcomings of hyperspectral information redundancy and similar sample colors that reduce the accuracy of ground object classification,this paper studies the feature selection of vegetation and features with similar colors.The following research work has been carried out:1.In order to improve classification accuracy by eliminating hyperspectral image information redundancy,this paper proposes a wavelength selection algorithm,LBIBPSO(Binary Particle Swarm Optimization With Local Band Index),which combines the amount of information and the separability between classes.Based on the information calculation model of local band index in IOIF(Improved Optimal Index Factor),this method introduces the inter class distance to measure the inter class separability of various samples,and puts forward a reasonable fitness function.The algorithm effectively eliminates the redundant information between wavelengths and retains the wavelengths beneficial to distinguishing features.The selected wavelength combination contains more information,less correlation between bands,and good separability between ground objects,which improves the classification accuracy of samples with similar colors.2.In order to test the classification performance of dimensionality reduction results,hyperspectral simulation images of vegetation and its similar color features are constructed in this paper.Firstly,the hyperspectral reflectance of six ground object samples is measured,and the hyperspectral image is simulated based on these measured data and hyperspectral mixed pixel theory.The simulation data set can flexibly adjust the pixel abundance ratio and SNR of the sample through the setting of parameters,so as to flexibly verify the classification performance of the dimensionality reduction results obtained by the wavelength selection algorithm from different test conditions.Then,LBI-BPSO,GA-BPSO(Particle Swarm Optimization With A Genetic Algorithm)and IOIF are compared to classify the hyperspectral simulation images of similar color features.The performance of the algorithm is analyzed from three aspects:SNR = 320,different endmember abundance and different SNRs.When the condition of SNR=320 is established,the overall classification accuracy OA of LBI-BPSO is improved by 2.43%,the average classification accuracy AA is improved by 3.07%,and the kappa coefficient is improved by3.06% compared with the IOIF with the second effect;For mixed pixels with different abundance ratios,LBI-BPSO can achieve the highest overall average;When the SNR changes,LBI-BPSO shows stronger tolerance to noise.Calculate the relative value difference of classification results under the conditions of the highest SNR(SNR = 320)and the lowest SNR(SNR = 10).Better than IOIF,the values of OA,AA and kappa coefficients of LBI-BPSO are reduced by33.31%,37.26% and 58.59% respectively;Compared with GA-BPSO,the values of OA,AA and kappa coefficients of LBI-BPSO decreased by 22.47%,26.63% and 41.33%respectively.In this paper,an improved algorithm LBI-BPSO is proposed for the wavelength selection direction of hyperspectral images.It is verified by the simulation data set established by vegetation and similar color features,and compared with the relevant algorithms.The research conclusions and achievements of ground objects with similar vegetation colors have important practical significance for the development of hyperspectral image application technology to satisfy the demand of camouflage target detection in military field and agricultural and forestry ecological monitoring in civil field.