Study On Polarimetric SAR Image Classification Method Based On Polarization Scattering Characteristics And SVM

Polarimetric SAR Image Classification is a very important research area in radar remote sensing applications.Result of the recent research of polarimetric SAR target decomposition theory and machine learning algorithm,the classification method combined polarization scattering and nonparametric model has been developed.Due to the high dimensionality of polarimetric SAR image features the information used in the classification method has become redundancy.Facing the problem we study the classification method based on polarization scattering characteristics and SVM both on the full polarimetric SAR data and dual polarimetric SAR data.In this study,the Yigen agroforestry ecotone of Hulunbeier city was used as the experimental area of full polarimetric SAR image classification research.We obtain a GF-3 full polarimetric SAR image as research data.The ground truth data used to verify the classification result were abstained by synchronous field investigation.For the purpose of researching Dual-pol SAR classification method.We chose Hulunbeier city as our experimental area and collected 63 scenes of GF-3 dual polarimetric images in this area.In the same year of imaging time,a field survey of the city's land cover and forest distribution has been token.The acquired land label data are used as ground truth to verify our mapping result.Firstly,based on the target decomposition theory and Stokes vector characteristics of polarimetric SAR data,we analyzed polarization scattering features of typical surface object of the research area using both full polarimetric data and Dual-pol data of GF-3.The research validate the effectiveness of Stokes vector features using as alternative classification features.Secondly,based on the polarization scattering characteristics and texture parameter of polarimetric SAR data,we compare the results and validation accuracy of the SVM model classification based the optimal features combination selected by genetic algorithm,ReliefF algorithm and SVM_RFE algorithm.The experimental results show that the feature extraction method combined with genetic algorithm and SVM verification accuracy can effectively avoid getting into local optimal feature combination and obtain the feature set with the best verification precision.Compared with genetic algorithm,the SVM_RFE algorithm is easy to fall into the local optimal solution,but it has a significant advantage in the program running time and the verification accuracy is still in the acceptable range.Finally,based on the conclusion of the first two parts of this paper.Using the dual-pol mode data of GF-3,we extract the elements of the scattering matrix elements,stokes vectors features and the texture features of the span image to form the classification features,adopt theSVM-RFE algorithm to select final classification features set and classify it by the SVM model one monorail image at a time.The classification results were spliced and the forest-non forest classification map of the experimental area was obtained,and the accuracy of the results was verified according to the field survey data.Total accuracy reached 87.9%,and the user precision of forest land was 92.9% and the producer precision was 88.1%.