Hyperspectral Imagery Classification Based On Hybrid Neural Network

Hyperspectral remote sensing not only records the spatial information of ground object targets,but also collects the reflection information of high-dimensional spectra.It is one of the research directions of remote sensing and widely used in civil,military,agricultural and other fields.Hyperspectral imagery classification is one of the basic research directions in hyperspectral image processing,and the use of machine learning related algorithms for effective feature extraction and classification is the mainstream in the field.However,due to its inability to effectively combine spectral-spatial feature information,image data redundancy,less tag data and other difficult problems,accurate classification of ground objects in hyperspectral imagery is still challenging.Based on the related theory of deep learning and combining the characteristics of hyperspectral imagery,this dissertation analyzes and improves the relevant deep learning algorithm,studies the above problems in the classification of hyperspectral imagery,and designs the hybrid neural network.The main work of this dissertation is as follows:Firstly,in view of the problem that convolutional neural network cannot be applied to high-dimensional data in the hyperspectral imagery,a three-dimensional convolutional hybrid neural network is proposed.The weighted convolution of the characteristics of each spectral channels is proposed to obtain the characteristics of spectral-spatial feature.The parameters of the network are optimized by feedforward operation.Moreover,the small convolution kernel is adopted to avoid overfitting problem,speed up the network training,and enhance the extraction ability of spectral-spatial feature.In order to obtain spectral feature information effectively,the network combines with the stacked automatic encoder and improves it,and references the spatial context feature extraction layer to realize the extraction of spectral-spatial information fusion,so as to further improve the feature extraction and classification performance of the network.Secondly,the above method causes the loss of feature information in the pooling process,and cannot effectively use the shallow feature information.Aiming at this problem,a deep deconvolution hybrid neural network classification method based on skip structure is proposed for the ground object target classification.In order to overcome the over-fitting problem caused by high-dimensional hyperspectral data,the network introduces the band selection layer to remove redundant data information.The deep deconvolution network restores the lost feature information in the process of pooling by using deconvolution layers,and overcomes the problem of gradient disappearance and insufficient feature caused by a small number of training samples through using skip architecture,extracts more expressive feature information to realize the combination of deep semantic and shallow representation information,and multi-level feature fusion classification.The above designs combined with the stacked automatic encoder construct an end-to-end spectral-spatial feature classification network for hyperspectral imagery,which could achieve more accurate ground object classification for hyperspectral imagery.Experimental results show that this method has better classification performance than other deep learning based algorithms in hyperspectral imagery classification regardless of whether the training samples are sufficient or not.To sum up,in combination with the practical characteristics of hyperspectral imagery,and in view of the current major problems of the in-depth,conbined with the relevant deep learning neural network classification algorithm,this dissertation designs hybrid neural networks algorithm to extract rich and representative spectral-spatial feature information and achieve more accurate hyperspectral imagery classification.It provides an effective reference suggestion and method for the intelligent translation and practical application of hyperspectral imagery.