Research On Feature Extraction Method Of City Feature Of High Resolution SAR Image Based On Manifold Learning

Synthetic Aperture Radar(SAR)can penetrate the cloud,has the unique advantages of its all-time,all-weather observation capabilities to compensate defects of optical sensors that unable to obtain valid data in the cloudy rain and fog area,and has become an importance technique of remote sensing information extraction.With the development of SAR technique,the constantly enriched high resolution SAR data of urban area in high resolution SAR image,the urban typical objects extraction has become an important application of SAR,providing the basic data for urban planning,land use monitoring,the survey population density and so on.Owing to the complex scatter characteristics of high-resolution SAR image features,the extraction accuracy of urban typical objects based on high-resolution SAR images is not high and the high-dimensional nonlinear characteristics of high-resolution SAR data made automatic extraction of urban typical objects more difficult.As a machine learning method,manifold learning can find the within characteristics of data and the manifold learning which is good at dealing with the nonlinear data,is applied to feature extraction of the high-dimensional nonlinear SAR image,which is beneficial to improve the accuracy of target recognition.Therefore,in order to improve the extraction accuracy of urban typical objects and study the automatic extraction technology of urban typical objects,this paper studies the extraction method of high-resolution SAR image urban typical objects based on manifold learning.The main research contents are as follows:(1)The feature of object in high-resolution SAR was analyzed and the high-dimensional data set was constructed.Firstly,detailed analysis of high-resolution SAR urban images was conducted based on the radiation characteristics and the geometry characteristics of images,and the image characteristics of various types of objects.Secondly,eight texture features were obtained by the classic second order probability statistical method(gray level Co-occurrence matrix,GLCM),then the data set were constructed with eight texture features and gray feature,which original high-dimensional data set of manifold learning.Finally,the optimal window size for calculation of GLCM texture features were determined by experimental analysis.(2)Five typical manifold learning methods,such as Laplacian Eigenmap(LE),Locally Linear Embedding(LLE),Hessian Locally Linear Embedding(HLLE),Local Tangent Space Alignment(LTSA),Locality Preserving Projections(LPP),are selected for each of the types of objects.The high dimensional feature set of the type(building area,water body,playground and square)is reduced dimension,and finally the three types of objects are extracted.The results of the three methods were evaluated and the appropriate manifold learning method was selected to be improved.(3)The local Tangent Space Alignment(LTSA)method is used to improve the distribution of SAR data.Considering the manifold structure and the Euclidean distance,the weight reduction of the original method is given,and a distance and structure weighting method based on distance and structure weighting is proposed.The algorithm is applied to the dimension reduction of the high-dimensional feature set of high-resolution SAR image.The effectiveness of the DSWLTSA algorithm is verified by comparing the DSWLTSA and LTSA algorithms with four typical features.The applicability and application value of the DSWLTSA algorithm are analyzed in detail by experiment.(4)Aiming at the problem that the sample sparseness is not good,a new method based on the homogenization of the LLE algorithm(DHLLE)is proposed.The new method is based on the localized linear embedding algorithm(LLE)The recalculation of distance makes the problem improved.And the algorithm is applied to the dimension reduction of the high-dimensional feature set of high-resolution SAR image.Taking DHLLE and LLE algorithm as an example,the validity of DHLLE algorithm is verified by experiment.The applicability and application value of DHLLE algorithm are analyzed by experiment.