Research On Hyperspectral Image Classification Based On Spatial-Spectral Feature Learning And Deep Network

Hyperspectral remote sensing is a common type of remote sensing.The hyperspectral image has a very high spectral size and is finely segmented in the spectral dimension,so that each pixel of the hyperspectral image can contain a rich spectral curve of the ground object,which can reflect the subtle characteristics of the object spectrum.With the development of science and technology,hyperspectral images are more and more used in various fields,such as military mapping,agriculture,meteorological observation,land planning and earthquake monitoring.Through the classification of hyperspectral images,more valuable information is obtained,which has become one of the research hotspots of hyperspectral image processing technology.Basing on the advantages of depth network in feature learning,this dissertation proposes models based on spatial-spectral feature learning and deept network,and applies them to hyperspectral image classification tasks.The main contributions are as follows:1)A hyperspectral image classification method based on Wasserstein Generative Adversarial Network Gradient Penalty(WGAN-GP)and Progressive Growing Generative Adversarial Network(PG-GAN)is proposed.The training process of PG-GAN is used to make the training process more smooth.The loss function of WGAN-GP is used to make the training easier to reach convergence and balance.On the basis of the combination of the two methods,the network structure suitable for hyperspectral image is designed,and the classifier is added,finally the network model of hyperspectral image classification is obtained.The experimental results show that the proposed method achieves the best performance,which proves that the method can significantly improve the training process of the GAN based hyperspectral image classification method.2)A novel method is proposed based on Relational Recurrent Neural Networks(Relational-RNN)for hyperspectral image classification.,whose iedeal is inspired by that hyperspectral pixels can be considered as sequential data.Relational-RNN designs a new memory module-a Relational Memory Core(RMC),which contains a fixed set of memory slots and employs multi-head dot product attention to allow memory slots to interact for relational reasoning.Therefore,compared with recurrent neural networks,Relational-RNN for hyperspectral image classification not only considers the intrinsic sequence-based data structure of hyperspectral pixels,but also selects information between bands interactively.As a result,the variation and spectral correlation between bands of hyperspectral images were obtained meanwhile.Experiments indicate that Relational-RNN shows statistically higher accuracy than SVM-RBF and CN-N for hyperspectral image classification.In the future,further experiments will be carried out to fully confirm the characteristics of deep RRNN in hyperspectral image processing,so as to provide more accurate analysis for remote sensing,such as remote sensing big data analysis and transfer learning for change detection.3)A hyperspectral image classification algorithm based on feature fusion and Multi-Layered Gradient Boosting Decision Trees(mGBDT)is proposed.Hyperspectral image exists data redundancy and need a lot of computation because of its high spectral dimension.Considering the high dimension of data,the principal component analysis is used to reduce the dimension of hyperspectral image.Multi-feature fusion method is adopted.Firstly,EMP is used to extract geometric information of hyperspectral image,and then the texture features of the hyperspectral image are extracted by gray level co-occurrence matrix.Four main attributes are gotten,including energy,entropy,contrast and correlation.Finally,the linear multi-scale spatial features of the image are obtained by using Gaussian convolution kernel.The extracted multi-features are combined to form a one-dimensional feature vector into the mGBDT for classification.The feasibility analysis proves that the hyperspectral classification method based on feature fusion and mGBDT is efficient and has excellent performance even on small samples.4)A novel hyperspectral image classification algorithm based on double channel temporal dense network is proposed to solve the problem that most hyperspectral image classification methods cannot extract insufficient information and the classification accuracy is not high enough.First,the proposed method uses Temporal Convolutional Network(TCN)and Dense Convolutional Network(DenseNets)to process the spectral information and spatial information of hyperspectral image data separately,and then fuses the features extracted by the two networks.Finally,the merged features are sent to the softmax classifier for classification.Simulation experiment is carried out on the University of Pavia dataset.The proposed algorithm is compared with traditional hyperspectral image classification algorithm,spatial hyperspectral image classification algorithm and spectral classification algorithm.The experimental results show that,compared to other classic classification algorithms,the proposed algorithm can effectively extract the feature of the target from the spatial structure and spectral channel,and achieve 99%classification accuracy on the classical dataset,2%to 3%higher than other methods.