Deep Learning Based Hyperspectral Image Super-Resolution And Classification Methods

Hyperspectral image(HSI)is data cube which combines the spatial and spectral information of land-cover objects.Due to the rich spectal resolution of HSI,a continuous and fine spectral curve can be obtained from the HSI at any point in the space.And,it will benefit accurate land-cover analysis.Therefore,HSI is widely used in agricultural remote sensing,pest and disease monitoring,vegetation survey,environmental monitoring,military detection and other fields.In recent years,the rapid development of hyperspectral remote sensing systems offers new opportunities for earth observation technology,but it also presents new challenge for hyperspectral image processing,which mainly include the following parts.(i)Affected by the restriction of HSI sensors and complex imaging environment,the spatial resolution of HSI is limitation which can not meet the applications of fine object recognition and classification.(?)Due to the diversity of land-cover object and the complex conditions of imaging system,different classes of land-cover objects may have similar spectral features,while the land-cover objects of same class may have various spectral characteristics.And,this phenomenon is very common in HSI,leading to the difficulties to high-precision object recognition and classification.This thesis firstly analyzes the existing hyperspectral image prcessing technology.We force on the aforementationed limitations of HSI,and proposed novel deep learning based spatial resolution enhancement and classification methods.The proposed methods can fully utilize the spatial-spectral characters of HSI,and provide superior image super resolution and classification performance.The main contributions are as below.(1)We force on the limitation of HSI spatial resolution,two novel single hyperspectral image super-resolution models and a hybrid loss function are proposed.The first model is separable-spectral convolution and inception block network(SSIN),which can fully utilize the spatial information of HSI image for each band.And then,a feature fution and progressive reconstruction strategy are used to restore the high-frequence characteristic of target high resolution image with step-by-step approach.The second model is dense connection residual network(DCRN),which can take advantage of the intermediate features of network.The hybrid loss function is proposed which contains spatial constraint,spectral constraint and smooth constraint.The experimental results show that the proposed methos can significantly improve the spatial and spectral quality of reconstruction image.(2)A novel hyperspectal and multi-spectral image unsupervised fusion method is proposed which can adaptive handle arbitrary point spread function and spectral response function.The multi-spectral image provides detailed spatial contextual information and the hyperspectral image preserves the spectral information.The basic idea of proposed method is based on matrix factorization.According to the shared endmember of linear unmixing and relationship between different abundance,three interrelated autoencoder network are constructed.The point spread function and spectral response function are adaptively obtained in the autoencoder training.In addition,the strategy of generative adversarial network is used to enforce the semantic similarity between the input and output of autoencoder.The experimental results show that the proposed method can effectively enhance the adaptability and robustness of the fusion model,and it has higher reconstruction accuracy than other methods.(3)A novel deep learning based spatial-spectral classification method is proposed.Firstly,an end-to-end classification framework is proposed,which changes the computational redundancy and low training efficiency caused by the sample point slice in the traditional deep learning-based classification methods.Secondly,a multi-level spatial feature pyramid is constructed,which can take advantage of the shallow,middle and deep feature of ResNet.Thirdly,a multi-receptive-field fusion structure is proposed,which contains multiple dilated convolution and can fuse multiple-receptive-field features using a top-to-bottom approach.This fusion structure can effectively enhance multi-scale spatial-spectral feature utilization.The experimental results show that the proposed classification method can avoid the "salt"classification results,and can effectively improve the classification accuracies.(4)The classification results of single hyperspectral image super-resoltuion and fusion based hyperspectral image super-resolution is fully compared.Single image super-resolution reconstruction and fusion-based super-resolution reconstruction algorithm is used to reconstruct remote sensing images of 2x,4x,8x and 16x resolution respectively.SVM and the proposed spatial-spectral classification method is used to test the classification results of different super-resolution image.The experiment results show that the spatial resolution enhanced image of single hyperspectral image super resolution algorithm can effectively improve the classification effect and accuracies.The fusion based resolution enhanced image can increase image spatial information and improve classification effect and accuracies.Compared with pixel-wised classification method,the proposed spatial-spectral classification model can significantly improve the classification effect and accuracy.