Building Extraction Method Of Feature Optimization Combined GA And SVM Using Polarization SAR Data

As its all-weather and all-weather characteristics,Synthetic Aperture Radar(SAR)becomes one of the most important remote sensing technology means to obtain surface information.Compared with single-channel SAR,Polarimetric SAR(PolSAR)can provide more detailed and ground scattering information,how to interpret these information quickly and accurately has become the focus of SAR research.As one of the research hotspots in the interpretation of polarimetric SAR images,building extraction has been attracted more and more scholars’ attention.The polarimetric SAR image contains more polarization scattering features and texture features.How to select the most effective and essential features is the key to optimize the feature space and improve the performance of the classifier.This paper analyzes and extracts the polarization features and texture features suitable for building extraction using the GF-3 data and uses a feature optimization algorithm.Using the proposed method,the feature space optimized,and the optimal features subset of building extraction are selected.The main work and results are shown as follows:(1)Analyze the representation of typical buildings in the PolSAR image and extract the polarization scattering and texture features which sensitive for the buildings.The feature set has 74-dimensional features which includes:target decomposition parameters,polarization correlation coefficients,polarization coherence matrix and covariance matrix correlation parameters,and statistical texture parameters extracted by the gray level co-occurrence matrix algorithm.The contribution of different types of features to building extraction is discussed and analyzed.Experiments results show that contribution ranged from large to small is texture features,target decomposition and polarization matrix parameters,and the polarization correlation coefficient.(2)This paper uses a GA and SVM coupled feature optimization algorithm.It uses the test sample accuracy and feature dimensions obtained by the SVM classifier to construct the fitness function in the genetic algorithm.After multiple iterations of optimization,the original 74-dimensional feature is reduced to 16 dimensions.The SVM algorithm is used to evaluate the building extraction results compared before and after feature selection.Experimental results show that the overall accuracy and Kappa coefficient of building extraction after feature selection are improved,which proves the importance and effectiveness of the proposed feature optimization method.(3)The study compares the building extraction results of different feature sets based on the feature optimization algorithm in this paper.Experimental results show that adding texture features to traditional polarization scattering features improves the accuracy of building extraction.It is verified that the feature set extracted in this paper has advantages in proposed feature optimization algorithm based on GA and SVM.