Study On Differential Scattering Characteristic Extraction And Classification Method For Agricultural Crops Based On H/? Decomposition For PolSAR Images

Crop type classification is one of the most significant applications in PolSAR imagery.As a remote sensing technique,PolSAR has been proved to have the ability to provide high-resolution information of illustrated objects.However,single-temporal PolSAR data are restricted to provide sufficient information for crop identification due to the complicated condition of varying morphology within various growing stages,which limits the improvement of crop classification accuracy.With an increasing number of spaceborne PolSAR systems launched,a large amount of real PolSAR data from repeated observations of targets are being generated and used to provide great opportunities for multi-temporal analysis.Multi-temporal PolSAR data used in this paper based on classical H/? parameters to extract the differential scattering characteristics to improve crop classification accuracy.Firstly,this paper based on the H/? decomposition to analysis the scattering characteristic distribution area of H/? decomposition changes regularly on the H/? classification plane using multi-temporal quad-and dual PolSAR data,and proposed a new parameter ? for the first time to measure the changes of differential characteristics.Then,a classification method based on the new parameter ? was proposed,which is adopted to both of quad-and dualPolSAR data sets.Different classification methods(i.e.,complex Wishart,Freeman-Wishart,the classification algorithm proposed in this study and Support Vector Machine(SVM)with different inputs)based on simulated quad-and dual-polarimetric Sentinel-1 and real Sentinel-1 data sets were used to validate the advantages of the new parameter ? and the proposed classification method based on the ?.In addition,we used Convolutional Neural Network(CNN)to recognize and segment PolSAR image,and validated the proposed polarization scattering parameters ? can also optimize the classification performance of CNN.The main contents and conclusions are as follows:(1)The definition of ?,which is the new polarimetric scattering characteristicsDifferent kinds of crops change within their own growth cycles,and therefore there exist significant differences among the moving amounts of distribution.For this reason,this paper defines a new parameter(i.e.,?)to measure the change of differential characteristics of H/? distribution.A new polarization scattering parameter is proposed based on the sensitivity of the phase to the regional variation,which improved the degree of freedom of the H/? parameter.The variation of the polarization scattering characteristic on the H/? classification plane can be well described.Therefore,the variation of the polarization scattering characteristic of crops with the growth period can be well described.(2)Compared with the traditional polarization parameters,the proposed polarimetric parameters(i.e.,?)are more suitable to describe the changing law of the scattering characteristics of crops in their life cycle.In order to validate the advantages and robustness of the proposed parameter ?,the standard deviations of various scattering parameters for different sensors between different crops and within the same category are calculated out and compared.For the different sensors,the results show that the standard deviation of the proposed parameter In order to validate the advantages and robustness of the proposed parameter ? is the lowest for the same crop.Meanwhile,for different crops,standard deviation of the proposed parameter In order to validate the advantages and robustness of the proposed parameter ? is the highest.Therefore,the proposed parameter In order to validate the advantages and robustness of the proposed parameter ? provides the best homogeny for the same class and the best contrast between different crop types.(3)For the simulated quad-and dual-polarimetric Sentinel-1 data,the proposed classification algorithm and the SVM classifier with feature vector which introduced the proposed polarization scattering parameter(i.e.,?)have a better classification performance.The overall accuracy and Kappa coefficient of the classification algorithm based on the new parameter have reached 78.09% and 69.45% when the crops covered by the study area are divided into 6 classes,and there is a significant advantage compared with complex Wishart,Freeman-Wishart and SVM classifiers which introduced the traditional polarization parameters into feature vector,when the proposed parameter ? is introduced into SVM classifier,the classification can be improved effectively,and the overall accuracy and Kappa coefficient can reach 79.99% and 71.70%,respectively.Furthermore,if the crops with similar morphology are merged,the advantages will be more obvious.This advantage also exists in the simulated dual polarimetric Sentinel-1 data.(4)For the real polarimetric Sentinel-1 data,the proposed classification algorithm and the SVM classifier with feature vector which introduced the proposed polarization scattering parameter ? have better classification effective.Compared with the complex Wishart classifier,the overall accuracy and Kappa coefficient of the proposed method were increased by 6% and 9%,respectively.For the SVM classifier,the result of introducing the parameters ? into the eigenvector is the highest,and compared with the eigenvector composed of the traditional polarization parameters,the overall accuracy and Kappa coefficient were increased by about 5% and 10%.(5)The crop classification of PolSAR images is realized by using CNN network,and the proposed polarization scattering parameters ? are introduced into the feature channel,which can effectively improve the classification results.For the simulated Sentinel-1 and real Sentinel-1 data,when the porposed parameter ? are introduced into the feature channels,the overall accuracy and Kappa coefficient are highest,and reached 90.13%,86.12% and 93.30%,85.51%,respectively.