Research On Land Cover Classification Using High-Resolution SAR Image

Synthetic Aperture Radar(SAR)is a coherent microwave remote sensing imaging system,which can acquire SAR images of the region of interest in all-day and all-weather.Nowadays,SAR has entered into a high-resolution age.Compared to low-and medium-resolution SAR images,high-resolution SAR images exhibit rich texture details and clear geometry,giving people the opportunity to get more detailed information about the region of interest.However,opportunities and challenges coexist,and automatic interpretation of high-resolution SAR images requires the introduction of new knowledge and the development of new technologies.In this context,this paper aims at the problem of land use and land cover(LULC)classification in the automatic interpretation of high-resolution SAR images.Using knowledge in the fields of machine learning and pattern recognition,a variety of different strategies are proposed to improve the LULC classification performance of high-resolution SAR images.The main work and contributions of this paper are as follows:1.The commonly used speckle filtering algorithms may blur and destroy the edge and texture details in high-resolution SAR images.To solve this problem,a high-resolution SAR image filtering method based on nonlocal low-rank minimization model is proposed.Under the framework of Alternating Direction Method of Multipliers(ADMM),the low-rank minimization problem involved in the method can be effectively solved and guaranteed to converge to a local minimum.Extensive experiments on simulated SAR image and high-resolution real SAR images show that the proposed filtering method can not only sufficiently filter out the speckle noise but also preserve the edge and texture details in the SAR images.2.Classification methods based on texture features are easy to produce misclassification similar to "salt and salt noise".To solve this problem,from the perspective of classifier design,the composite kernels are designed by using the properties of the kernel function.The composite kernels consist of the original texture feature kernel and a newly designed context kernel.The context kernel explores adaptive context information with superpixels as neighborhoods,and the calculations are simple and efficient.Classification results on high-resolution real SAR images show that compared with the traditional classification method,classification methods based on composite kerhels can proauce smoother and more accurate classification results without significantly increasing computational complexity.3.Traditional texture features have insufficient ability to distinguish different texture in high-resolution SAR images.To solve the problem,a new texture covariance feature is proposed.Considering that the texture covariance feature belongs to the symmetric positive definite matrix and forms the Riemannian manifold,classification methods developed in the Euclidean space ignore the geometry of the Riemannian manifold and cannot obtain the optimal classification result.We further propose to classify the texture covariance features by using the Gaussian kernel based on the Log-Euclidean distance.Since the Log-Euclidean distance is the geodesic distance on the Riemannian manifold,the geometry of the texture covariance feature can be preserved.Classification results on high-resolution real SAR images demonstrate the effectiveness of the proposed classification method.