Research On The Classification Methods Of Coastal Zone Land Cover Types Based On Fully Polarimetric SAR Multi-features

Coastal ecosystems are fragile and sensitive,Climate change and human activities have brought tremendous pressure on the coastal environment,causes the deterioration of ecological environment.The climate in this area is cloudy and rainy,traditional optical remote sensing data has high failure rate,therefore,it is of great significance to use the full-polarization SAR data to develop coastal classification.Full-polarization SAR data contains rich target information,which can extract a large number of features.This article carried out classification experiments based on Radarsat-2 full-polarization SAR data for mining the complementarity of multiple features and removing redundant information.The main research results include the following aspects:(1)According to the feature significance analysis,it was found that in the statistical features channel,in addition to saline soda and seawater,the other 9 land-use features showed complementary separability.In the physical scattering channel,the large class separability is significant,and the small class confusion serious.Artificial plants,Reeds,Spartina alterniflora,Saussurea have significantly different response characteristics in different physical scattering channels.Therefore,the linear combination of scattering features can be used as an important new feature to improve the accuracy of these types of features.Texture features can be used as excellent auxiliary features to gain classification effects.(2)Combining scattering model-Wishart unsupervised classification and Multi-kernel Learning(MKL)supervising classification,a classification framework of tree structure was esTablelished.Firstly,the Freeman decomposition scattering features are used as the input training set for Wishart unsupervised classification to obtain low-level classification results.Then for the problem that the heterogeneous features have different response to different kernel functions,the MKL method with prior label is used to classify the low-level categories accurately.The experimental results show that the overall accuracy of the proposed method and the classification accuracy of the single ground object are superior to the traditional SVM and MKL.This method can use the complementary characteristics of multiple features of polarimetric SAR to effectively improve the classification accuracy.(3)Analysis of Phase Sensitivity of Five Planting Vegetations under Different Phenological Conditions Based on Multi-temporal Data to set the reconstruction threshold for PCA dimension reduction for multi-temporal features.Through the six cross-contrast tests of different phase data and classifiers,it was found that the vegetation withered period data had better classification accuracy than the prosperity period.The classification accuracy of multitemporal data in vegetation is higher than that of two types of single-phase data.Multi-kernel learning classifiers have advantages in dealing with the classification of high-dimensional features,however,when the feature amount is less,the classification efficiency is lower than SVM.