Classification Of Satellite Remote Sensing Images Based On Multi Feature Fusion

With the rapid development of remote sensing,research on satellite remote sensing images have received more and more attention.Hyperspectral image(HSI),as an important branch of satellite remote sensing image with high dimensional data,contains a wealth of information for us to learn.In the field of remote sensing image research,HSI classification is a hot issue,which involves computer graphics,mathematical statistics,matrix theory and other disciplines.For HSI classification,we prefer to use machine learning classification algorithm(including supervised learning and unsupervised learning)to predict the category for each pixel in the image.Owing to the high dimension of the feature vector(spectral feature)corresponding to each pixel and less available training samples,when we use the supervised classification method,we will meet the "Hughes" phenomenon,which is why the hyperspectral image is sensitive to the choice of classification method.After a lot of experiments,we learned that support vector machine(SVM)is suitable for HSI classification problem,and SVM achieved good results in the field of HSI classification,which can overcome the problem of less available training samples and high dimensionality.Considering the problem of HSI classification,the existing methods generally depend on the spectral features,but the spatial features were ignored which reflects the spatial and geographic information.Therefore,it is very important to solve that how to effectively fuse the information provided by these two kinds of features in the classification process for improving the accuracy of HSI classification.In this paper,we use attribute profiles(APs)method to extract spatial features,which spans each channel of HSI by attribute filters(AFs)and calculates the attribute values covered by the AFs.After a series of comparison,we can obtain the spatial feature of every pixel.By adding spectral feature and spatial feature with weights,we can realize feature fusion and use the new integrated features to build the classification model.Before integrating the features,it is necessary to conduct denoising processing.Recently,sparse representation based on dictionary learning has been paid more and more attention in data processing.In the field of image processing,low-rank matrix decomposition is a representative method in sparse representation,which is called low-rank representation(LRR)combining with dictionary learning.In this paper,by considering the assumption that adjacent elements have the same category in HSI,we can divide the whole image into blocks and do LRR on each block respectively.We propose the method: using LRR based on region division on both spectral and spatial feature to denoise,then integrating the new features and build the SVM classification model.According to the experimental results,our method gets more accurate classification results for HSI classification.Compared with the classical methods of integrating features based on kernels,our model has a great advantage.