Study Of High Resolution Remote Sensing Images Feature Extraction And Classification Based On Data Analysis

Hyperspectral remote sensing images(HRSIs)are three-dimensional data cubes formed by stacking two-dimensional images in different wavelength bands in increasing order,which have rich spectral and spatial information;radar high resolution range profile(HRRP)is a vector of the energy and composition of radar echoes at the scattering point of the target within each distance unit,which contains the surface characteristics and structural features of the target.Both HRSIs and HRRP are high-resolution remote sensing images.In order to improve the terrain classification accuracy for HRSIs,fractional differentiation is used to extract the spectral fractional differentiation feature(Spe FD)of HRSIs pixel,after which the validity of the Spe FD feature is verified by classical classifiers,and full connected network(FCN)and 1-dimensional convolutional neural network(1DCNN)are used to extract the deep Spe FD feature respectively,3-dimensional convolutional neural network(3DCNN)and hybrid spectral net(Hybrid SN)are used to extract the deep Spe FD-Spa feature containing spatial information respectively;In addition,a two-dimensional fractional differential mask is designed to extract the spatial fractional differentiation feature(Spa FD)of HRSIs pixel,and the Spa FD feature is fused with the original feature to form the hybrid feature Spa FD-Spe-Spa,then the deep Spa FD feature and the deep Spa FD-Spe-Spa feature are further extracted by 3DCNN and Hybrid SN respectively.For HRRP,the fractional differentiation is used to extract the amplitude spectrum fractional differentiation(ASFD)feature,and in order to solve the problem of low recognition rate of propeller aircraft,a two-stage classifier fusion strategy is designed based on the ASFD feature,which can improve the overall recognition rate on the basis of significantly increasing the recognition rate of propeller aircraft;1DCNN is designed to extract the deep ASFD feature and classify.The details of the research are as follows:(1)In order to improve the accuracy of HRSIs terrain classification,the Spe FD feature of HRSIs pixel is extracted by fractional differentiation,then the deep Spe FD feature is extracted by FCN and 1DCNN respectively,and the deep Spe FD-Spa feature is extracted by3 DCNN and Hybrid SN respectively.Firstly,an order selection criterion based on spectral separability criterion is proposed to select the appropriate differential order to extract the Spe FD feature for each HRSIs pixel;then the obtained Spe FD feature and and the Spe FD feature after dimensionality reduction by linear discriminant analysis are verified by the minimum distance classifier,support vector machine and K-nearest neighbor classifier;finally,the Spe FD feature is sent to FCN and 1DCNN to extract the deep Spe FD feature respectively,and the Spe FD-Spa feature cube is sent to 3DCNN and Hybrid SN to extract the deep Spe FD-Spa feature respectively.The classification results on four real HRSIs demonstrate that introducing the Spe FD feature into the deep learning environment can further improve the terrain classification accuracy for HRSIs,and the improvement is better in the case of small samples.(2)In order to effectively extract the spatial feature of HRSIs,a two-dimensional fractional differential mask is designed to extract the Spa FD feature for HRSIs pixel,then the Spa FD feature are fused with the original feature to obtain the hybrid feature Spa FD-Spe-Spa,and the 3DCNN and Hybrid SN are used to extract the deep Spa FD feature and the deep Spa FD-Spe-Spa feature.Firstly,a spectral-spatial joint criterion is proposed to select the appropriate order of differential mask,the two-dimensional image of each band is processed with fractional differential mask,and the results are stacked to form the Spa FD feature;then fuse the Spa FD feature with the original feature to obtain the hybrid feature Spa FD-Spe-Spa;finally,the Spa FD feature and the Spa FD-Spe-Spa feature are sent to the3 DCNN and Hybrid SN for deep feature extraction respectively.The classification results on four real HRSIs demonstrate that the Spa FD feature and the Spa FD-Spe-Spa feature can effectively improve the terrain classification accuracy compared with the original feature Spe-Spa.(3)In order to improve the target recognition rate for HRRP,fractional differentiation is used to extract the ASFD feature of HRRP,and a two-stage classification fusion strategy is designed based on the extracted ASFD feature to solve the problem of low recognition rate of propeller aircraft,and finally 1DCNN is used to extract the deep ASFD feature.Firstly,the amplitude spectrum feature of HRRP is obtained by fast Fourier transform,and then the fractional differential feature ASFD of the amplitude spectrum is extracted,and a two-stage classification fusion strategy is proposed: in the first stage,for the problem of low recognition rate of propeller aircraft,the propeller aircraft is selected from the angle of the energy difference of adjacent amplitude spectrum,and then the remaining samples are extracted from the high-order ASFD feature to further select the propeller aircraft;in the second stage,the low-order ASFD feature is extracted from the remaining amplitude spectrum feature after the propeller aircraft is selected in the first stage,and the final recognition results are obtained after classification.1DCNN is designed to extract and classify the deep feature of the extracted ASFD feature.The experimental results based on the real HRRP show that the ASFD feature combined with the two-stage classification fusion strategy can greatly improve the recognition rate of propeller aircraft,thus improve the overall recognition;compared with the deep amplitude spectrum feature,the deep ASFD feature is more effective and can effectively improve the target recognition rate.