| The coastal zone is the center of the development of human society and economy.It has dynamic changes,and it is a hot area of remote sensing monitoring.However,due to the influence of cloudy,rainy and fog in the coastal zone,it is difficult to get high resolution optical image.The high resolution SAR has the characteristics of all weather and all day.The medium optical image is rich in spectral information,and the replay cycle is short.The two kinds of data are easy to obtain.Therefore,combining the two kinds of data to carry out the classification of coastal land features,we can achieve complementary advantages and provide a solution for the difficulty of obtaining the coastal monitoring data.This paper studies from two levels: pixel-level and feature-level.We study the joint method of optical and full polarimetric SAR images in depth.And researching the applicable classification method for the joint images,then comparing the classification results,and getting the best classification scheme.The main research contents and conclusions are as follows:(1)The research on pixel-level fusion method of the medium optical image and the polarimetric SAR.We propose two pixel-level fusion methods.The first is the fusion method of polarimetric SAR and median optical images combined with PCA and HSV.The two is the fusion method considering the characteristics of polarization of polarimetric SAR and median optical image.Comparing the two proposed fusion method with traditional fusion methods,we find that the two proposed methods are superior to traditional fusion methods both qualitatively and quantitatively.(2)The research on feature extraction and combination of the medium optical image and the polarimetric SAR.We extract the spectral characteristics of NDVI,greenness and humidity.based on the median optical image,We extract and select the textural features of the median optical image.Due to SAR has a serious speckle noise,we only extract the polarization characteristics of the full polarimetric SAR image.Finally,aiming at the extracted multiple features,8 feature combination strategies are put forward for feature combination.(3)The research on classification method of the joint images.We propose object-oriented classification method for pixel-level fusion images,and comparing with high resolution optical image,we find that the fusion image obtained by combining PCA and HSV has the lowest classification accuracy,the classification accuracy of fusion images considering polarization characteristics is the highest.The addition of polarization features can obviously improve the classification accuracy.Aiming at multi-feature combination,we propose nonlinear multicore learning classification.Through analysis,we find that the classification accuracy with all the combined features is the highest.Comparing with the linear kernel SVM classification results,we find that the accuracy of nonlinear multi-core learning is better than that of linear kernel SVM classification.Nonlinear multi-core learning is more suitable for classification of multisource and heterogeneous features.Determining the optimal classification scheme.Comparing the object-oriented classification results based on the fusion image considering polarization characteristics with the nonlinear multi-kernel learning classification results based on the optimal feature combination,we find that the classification result of feature-level is better than that of pixel-level.The nonlinear multi-kernel learning classification based on the optimal feature combination is the optimal classification scheme for the coastal land surface classification with the combination of active and passive remote sensing data. |
