| Synthetic Aperture Radar(SAR)is an active remote sensing device,which makes it possible to observe and collect ground information without being restricted by climate and time.With the gradual development of SAR technology,the information in SAR images have become richer and more complex,and traditional SAR image terrain classification methods can no longer meet the needs of increasing applications.In this paper,the SAR image classification method is deeply researched from the three aspects of polarization characteristics,statistical characteristics and spatial correlation.The main research work of this paper is as follows:(1)To extract distinguishing and robust features,a SAR image classification method combining polarimetric features and convolutional neural network(CNN)is proposed.First,combine Cloude decomposition theory,span parameters and polarization coherence matrix to construct a SAR image feature representation that contains rich feature information.Then,the high-dimensional semantic features are obtained by CNN.Finally,The classification result is obtained through the Softmax classifier.This method has improved the overall classification accuracy,and has good distinguishability for different types of ground objects with similar scattering characteristics.(2)To improve the regional consistency of SAR image classification,a Markov Random Field Model(CC-WMRF)combining convolutional neural network classification confidence and Wishart distance is proposed.First,the likelihood probability is constructed by the Wishart distance between the pixel and the center of category.Secomd,sets the weight of the neighboring pixels through the classification confidence degree of each classification iteration and constructed the prior probability based on second-order neighborhood system.Then the pixel label is updated based on the Bayesian theory.This method fully combines the statistical characteristics of SAR images and the spatial correlation of neighborhoods,which can further enhance the consistency of the SAR image classification results.(3)Considering that the single-look polarimetric SAR data obey the Gamma distribution,an improved Markov Random Field Model(CC-GMRF)is proposed for water extraction based on single-look polarimetric SAR images.This method realizes the effective extraction of four different types of water bodies in plain lakes,mountain lakes,narrow rivers and wide rivers,and the extracted water bodies have smooth edges and good regional consistency.Through the application of monitoring examples of water area changes in the lower reaches of the Yangtze River,the adaptability and effectiveness of the proposed method in practical applications are verified. |
