Hyperspectral Image Classification Based On Deep Learning

Hyperspectral image is a high-dimensional image data that includes spectral information and spatial information,which plays an important role in vegetation detection,ocean monitoring,urban remote sensing,and medical research.Hyperspectral image classification is the process of identifying the ground objects.It provides the premise for the remote sensing researchers to obtain the information of the remote sensing data.Therefore,the accuracy of the classification affects the subsequent hyperspectral image processing stage.In order to improve the accuracy of hyperspectral image classification,it is important to focus on the essential features of hyperspectral image.The traditional classification method treats hyperspectral images as ordinary images.The shallow classification model is difficult to extract the essential features of the images and affect the accuracy of classification,thus affecting the classification accuracy.However,deep learning has the advantage of learning the deep features of input data and has been widely applied to related fields of images.In this paper,the spectral information and spatial information of hyperspectral images are combined in two different ways,and two different deep neural networks are used to learn the essential features to realize hyperspectral image classification.The research results obtained are as follows:(1)From the perspective of using spectral information,this paper designs a spectral reflectance curve matrix to achieve the purpose of the react the difference of shape of the spectral reflectance curve.On this basis,the spectral reflectance curve matrices of the central pixel and all the pixels in its neighboring space are accumulated to form a stacked spatial spectral reflectance curve matrix to joint spectral information and spatial information.And through the visualization,the matrix can reduce the negative impact of ‘same object with different spectral' and improve the discrimination of different objects.(2)Aiming at the stacked spatial spectral reflectance curve matrix,this paper constructs a feature fused convolutional neural network capable of combining shallow features and deep features,it can learn the essential features and get the classification.The convolutional neural network reduces training parameters by 1×1 convolution and implements feature fused based on concatenation of multiple fully connected layers.Experiments show that the feature fused convolutional neural network with stack spatial spectral reflectance curve matrix as input can effectively improve the classification effect.(3)From the perspective of utilizing spatial information,this paper proposes a method of extracting band spatial features,which realizes the efficient comprehensive utilization of spectral information and spatial information.This method convolutes each spectral band of each neighborhood pixel block with different convolution kernels and extracts multiple spatial feature maps of each band.On this basis,the spatial feature maps of each spectral band are fused into a spatial fusion feature map by using 3D average pooling.Spatial fusion feature maps of all bands of each neighborhood block form a sequence of fusion spectral spatial feature maps arranged in spectral order.(4)For the sequence of fusion spectral spatial feature maps,this paper first converts them into spatial spectral sequences based on the near-neighbor band block strategy,and then inputs the spatial spectral sequences into the long-term and short-term memory network for learning and classification.Experiments show that the classification effect can be effectively improved by using long short term memory network after feature map sequence made by near-neighbor band block strategy.