Research On Multitemporal Alignment And Classification Of High Resolution Remote Sensing Images

Multi-temporal images classification mainly uses a labeled image to classify unlabeled image which is collected at other time.Multitemporal classification includes multitemporal alignment and aligned data classification.Among them,multitemporal alignment is relative to the spectral drift problem caused by the different imaging environment between different remote sensing images.By changing the statistical distribution,the distribution difference between the same objects in the different temporal is eliminated,and the multitemporal image can be used together.Multitemporal classification effectively solves the problem of image interpretation without labels of unclassified image.In recent years,with the development of remote sensing technology,the spatial resolution of remote sensing images has been increasing.For classifying multitemporal remote sensing images under high spatial resolution,besides the existed spectral drift,there are many other problems: data massing caused by the imaging conditions,large internal divergence,low separability of statistical distribution and multi-model in the high resolution images,and the expression of objects under high resolution from traditional pixels to patch or objects.In order to solve the problem of high resolution multi-temporal image classification,this paper studies three aspects,namely,high resolution multi-temporal missing data recovery,drift alignment between different multi-temporal high resolution images,and multiple kernel classification and sparse multiple kernel learning for the aligned data classification.The main research results of this paper are as following:(1)In order to solve the missing data problem of multitemporal images due to imaging conditions,a multitemporal missing data recovery method without reference is studied by using the correlation characteristic of temporal-spectral.First,the concept of temporal spectral angle for multi-temporal remote sensing images and the corresponding calculation method are constructed.On this basis,a full scene oriented fast similarity search and information recovery algorithm is designed,and a high precision large area cloud-cover data recovery algorithm is realized,which can effectively solve the missing data problem existed in multi-temporal image classification.(2)Aiming at the problem of low spectral bands,complicated statistical distribution and serious spectral drift,a multitemporal manifold alignment method with label-based topology optimization is studied.First,on the basis of an unsupervised alignment method with explicit projection,based on the label information of the source phase,a majority voting method combining the similar weight is constructed to count the largest connection classes in the topology structure and delete the error class connection to achieve the optimization of the manifold topology in the alignment framework.Beside,an improved similarity measurement is proposed to optimize the alignment performance of same scene multitemporal images(3)For the multidimensional changes of multitemporal HR remote sensing images under multimode observation(multi angle,multi-resolution,multi source and so on),with the spatial-spectral structure of the high resolution remote sensing scene objects themselves,a multi-dimensional cooperative alignment for local space spectrum data is studied to align high resolution multitemporal multimode images.First,based on the tensor Patch data with spatial-spectral structure,multitemporal data is mapped to the corresponding tensor subspace.By constructing a multi-dimensional mapping association matrix in the tensor subspace,multi-dimensional alignment of multitemporal Patch data is realized.Secondly,a multichannel maximum likelihood method for estimating the intrinsic dimensions of the tensor data is proposed.The estimation algorithm is used to estimate the optimal parameters in tensor alignment and reduce the computational complexity of the tensor alignment model.(4)In view of the transformation from pixel to object in the classification of high resolution multitemporal remote sensing image,an object-based multitemporal alignment method is studied.First,multitemporal image is segmented by SLIC algorithm.Based on the feature of the center point spectrum of the superpixel,the object-based multitemporal alignment is realized by label-based topology optimization manifold alignment.Further,for same scene multitemporal image,by using supervoxel segmentation which keeps the unchanged objects consistency.A temporal-spatial supervoxel segmentation algorithm is proposed to improve the accuracy of alignment of objects under the same scene.(5)Aiming at the complex distribution of data after alignment,the nonlinearity of remote sensing images still exists.In order to further improve classificationa ccuracy of aligned HR images,a sparse multiple kernel learning(Sparse MKL)is studied.Through mappings of the aligned data with multiple kernels and the sparse PCA of the weighted fusion kernel matrix,the sparse maximum variance projection in the fusion kernel matrix is obtained.Thus the sparse scales(best sparse kernels)are obtained,and the selection of the effective kernels and the elimination of the redundant kernels are complete.