| As an important remote sensing technology,hyperspectral remote sensing plays an important role in geological monitoring,resource exploration,disaster prevention and other fields.As a key step in hyperspectral remote sensing technology,hyperspectral image classification is worthy of intensive study.Hyperspectral image can provide abundant features for the recognition and classification of ground objects because of its high resolution and combining image and spectrum.However,it also suffers many problems such as noise pollution,high cost of manual marking,difficulty in feature extraction and "dimension disaster".The low-rank based methods can effectively remove the noise in hyperspectral images and improve the consistency of samples,but these methods are faced with the problem of poor generalization performance and high computational cost.On the other hand,the deep learning model with the ability of adaptive feature extraction has good usability and generalization,but its classification performance largely depends on the quality and quantity of training samples.These two methods can complement each other well,so this thesis combines these two methods for hyperspectral image classification and completes the following works:(1)A hyperspectral image classification method based on convolutional neural network combining low-rank and sparse information is proposed to solve the problems of high calculation cost and poor generalization ability in low-rank based methods,which caused by the complex low rank subspace estimation(LRSE)operation.In this method,a convolutional network with two-branch structure is used to extract the low-rank and sparse components of hyperspectral image simultaneously,and then the fusion features are constructed for classification.On the one hand,this operation ensures the integrity of essential information in the case of inaccurate low-rank subspace estimation,so that the complex LRSE operation can be omitted.On the other hand,the fusion of multiple features improves the classification accuracy and robustness of the model.(2)A hyperspectral image classification method based on adaptive low-rank representation extraction convolutional network is proposed to solve the problems of high delay and waste of computing resources in the existing two-stage low-rank based methods.This method integrates the low-rank representation extraction of samples and the classification based on the low-rank representation into an end-to-end convolutional network,and designs a network training method based on weight migration and a low-rank representation extraction method based on spatial and category information to ensure the extraction ability of low-rank representation of the network.This operation solves the problem of high latency caused by the separation of feature extraction and classification process.On the other hand,the calculation cost of manual extraction of low-rank representation is saved by utilizing the adaptive feature extraction ability of neural network.(3)A hyperspectral image classification method based on low-rank representation and multi-network fusion is proposed to solve the problems of "dimension disaster" and over-fitting caused by insufficient training samples.In this method,the low-rank representation of hyperspectral images is divided into several low-resolution subgraphs in spectral dimension,and a number of lightweight convolutional networks are designed to extract the features of each subgraph respectively.Finally,the output of multiple subnets is fused with three different strategies to obtain the final classification results.On the one hand,on the premise of ensuring the integrity of spectral information,this operation reduces the spectral dimensions,expands the training samples,and enables the network models to get more adequate training.On the other hand,the idea of ensemble learning is introduced to enhance the classification performance of the model through multi-network fusion.In this thesis,three hyperspectral image data sets are used to verify the effectiveness of the above methods.In addition,some state-of-the-art hyperspectral image classification methods are introduced for comparison to prove the competitiveness of the proposed methods in classification performance. |
