Hyperspectral Image Classification Based On Multi-scale Spatial-spectral Feature Fusion And Deep Learning

With the continuous development of remote sensing technology,hyperspectral images have attracted more and more attention in agricultural planning,disaster prevention,resource exploration,environmental monitoring and other fields.Different from general natural images,hyperspectral images have more spectral bands and contain a large number of spectral features.These features not only provide information for hyperspectral image classification tasks,but also cause many problems,such as insufficient feature extraction,too many redundant parameters,too complex model,noise pollution,"dimension disaster",limited number of samples and so on.In recent years,hyperspectral image classification method based on deep learning has become a research hotspot in this field.Deep convolutional neural network not only has strong feature extraction ability,but also has universal network model.The same model can be applied to multiple hyperspectral image data sets,but it needs sufficient training samples in the training process.In order to solve the problem of insufficient samples,generative adversarial network can generate high-quality samples and expand the training sample data set by using the game and confrontation between generative network and discriminant network.Therefore,based on the above two network frameworks,this paper proposes a new network model for hyperspectral image classification,aiming at the problems of insufficient feature extraction,too many redundant parameters,too complex model,noise pollution,"dimension disaster",and limited number of samples.The main work of this paper is as follows:(1)A hyperspectral image classification method based on multi-scale and multi-level spatialspectral feature fusion network is proposed,aiming to solve the problem that the feature extraction in traditional hyperspectral image classification model is not enough,which affects the classification performance.The method takes neighborhood blocks of different scales as input,and introduces multi-scale features for the model.At the same time,a new3D-2D alternating residual block is designed.The fusion of spectral and spatial features is realized effectively,and the fusion of high-level and low-level features is realized.The threedimensional convolution neural network and two-dimensional convolution neural network in this model are connected in a cascaded way to ensure the continuity of feature extraction.The original hyperspectral image is used as the input of the network,and the feature engineering is not used in the method.The spatial information and spectral information in the original hyperspectral image are greatly preserved,thus improving the classification accuracy of the model.The experimental results on four real hyperspectral image data sets fully validate the effectiveness of the method.(2)A hyperspectral image classification method based on attention mechanism and weight sharing fusion network is proposed to solve the problems of large number of parameters,complex model and being difficult to extract effective information from original samples in multi-branch depth neural network.Using the functions and characteristics of siamese network and human visual attention mechanism,the weight sharing module and attention mechanism module are designed,which can greatly reduce the amount of model parameters and complexity.At the same time,the network model can improve the transmission weight of information that is beneficial to classification and reduce the transmission of noise information in the model.Compared with other existing methods,this method is competitive in classification performance.(3)A hyperspectral image classification method based on multi-features fitting and residual learning generative adversarial network model is proposed to solve the over-fitting problem of hyperspectral image classification model based on deep learning due to limited samples.In this method,a new MRGAN sample expansion model is designed to expand the training sample set of hyperspectral images.In this model,3D-2D alternating residual blocks are used as the basic module of the generator to better fit the spatial and spectral information of hyperspectral images,so as to generate false samples that are similar to the real samples.At the same time,the classical residual network is used as the basic module of the discriminator to improve the discrimination ability of the model.This method uses the idea of game and confrontation between generative network and discriminant network to generate highquality samples and improve the richness of training sample set.Finally,the advantages of this method are validated by experiments.