Joint Classification Of Multispectral/hyperspectral Remote Sensing Images

The advancement of spectral imaging technology and the development of remote sensing technology have made the acquisition of high-qualit y multispectral/hyperspectral data more and more convenient and economical.People have obtained more and more hyperspectral and multispectral remote sensing data in the same scene.The collaborative processing of optical remote sensing data on the same platform has become an emerging proposition.The efficiency and cost of manually labeling data make how to efficiently and co nveniently process this information and how to use multispectral hyperspectral data for collaborative classification have become emerging propositions in remote sensing technology.Under the same scene and the same platform,multispectral data often has high spatial resolution,which is convenient for manual labeling.However,insufficient spectral resolution makes it difficult to obtain high classification accuracy for multispectral data,and hyperspectral data often has high spectral resolution.The rate makes it have a higher classification accuracy,but the lower spatial resolution increases the cost of labeling data,and multispectral images under the same platform often have a larger width.For this reason,this article will explore how to efficiently process the same platform The hyperspectral/multispectral data obtained by the same sensor is processed together to make the excellent spectral resolution of the hyperspectral image transfer to the multispectral,so that it can obtain good classification accuracy.In order to improve computational efficiency,this paper first studies the dimensionality disaster of hyperspectral images,that is,the problem o f dimensionality reduction.In order to reduce redundant information,this paper uses a local retention projection dimensionality reduction algorithm on the basis of map fusion to predict hyperspectral images.Processing reduces the dimensionality of hyperspectral images.On this basis,the feature extraction based on public subspace learning is studied.Starting from the theoretical basis of subspace learning,the theoretical basis of transfer learning,which is subspace learning,is first introduced,and public subspace learning is constructed on this theoretical basis.The model of the algorithm is improved,and the mainstream algorithms in the current research status of feature learning only consider the shortcomings of the single-modal difference between the source domain and the target domain at the original feature level,and the model is solved,optimized and analyzed for convergence.On this basis,two methods are used to further optimize the common subspace learning algorithm,and then the three classifiers used in this paper to verify the effectiveness of the feature extraction algorithm and the evaluation indicators of the experimental results are briefly introduced.Finally,the experimental results and analysis are given,and the effectiveness of the common subspace learning algorithm is summarized.Then,starting from the manifold alignment algorithm of the theoretical basis of the learnable manifold alignment algorithm,a joint classification model of the learnable manifold alignment is constructed on the basis of the manifold alignment algorithm,which improves the unsupervised method and usua lly cannot align the multi-mode well.The model is optimized and the convergence analysis is performed.Finally,the semi-supervised manifold alignment algorithm used for comparison experiments is briefly introduced,and the cross-evaluation results of all algorithms are given.Pass and semi-supervised manifold alignment algorithm The comparison verifies the effectiveness of the two main cross-modal feature extraction algorithms in this paper.