Research On Hyperspectral Remote Sensing Image Classification Based On Three-dimensional Convolutional Neural Network

Hyperspectral remote sensing refers to the process of using hyperspectral sensors to obtain the corresponding target data and apply it under certain environmental conditions in space.Hyperspectral remote sensing images refer to the images obtained by hyperspectral remote sensing technology.Hyperspectral remote sensing images consist of hundreds of continuous spectral bands in the range of ultraviolet,visible and short-wave infrared bands,and also contain rich spatial features of ground objects.Compared with the panchromatic image and the multispectral image,the hyperspectral remote sensing image has higher spectral resolution and can provide more detailed data for the classification of ground objects.Hyperspectral remote sensing image classification is widely used in environmental change monitoring,agricultural planting evaluation,meteorological forecast,modern military,natural resources exploration and so on.Traditional methods of hyperspectral image classification are gradually becoming mature.However,these methods all adopted the way of manual feature extraction,therefore the feature expression ability of the traditional methods is limited and the generalization ability is weak,which cannot meet the high requirements of classification task.In recent years,the methods of deep learning have achieved remarkable results in the field of computer vision,and the hyperspectral classification methods based on deep learning have become a research hotspot.Compared with the traditional hyperspectral classification methods,the deep learning model is similar to the human visual system,which has a hierarchical structure and can automatically extract high-level semantic information of images,and the realization process is more intelligent and flexible.Although great progress has been made in the field of hyperspectral remote sensing image classification,it is still faced with severe challenges,such as the spectral-spatial feature of hyperspectral remote sensing images can not be effectively extracted,the labelled samples are insufficient,the redundancy of too many bands,and the phenomenon of the same object with different spectrum and different objects with the same spectral characteristics.The existence of these problems also pushes researchers to further explore the field.Convolutional neural network(CNN)is a very important network structure in the field of deep learning,and it also shows excellent performance in the field of hyperspectral remote sensing image classification.At present,although there are a lot of classification methods based on convolutional neural network for hyperspectral remote sensing images,the detailed comparison and analysis of the feature extraction ability for these methods has not been reported.Therefore,three groups of convolutional neural networks are constructed to compare and analyze the feature extraction ability of different types of convolutional neural networks for hyperspectral remote sensing images.In order to solve the problem that spectral-spatial features cannot be extracted effectively and there is a lot of redundant information in the process of feature extraction for hyperspectral image classification,the 3D CNN based on spectral-spatial dense connectivity strategy and spectral-spatial attention mechanism is proposed.In order to solve the problems in the classification of hyperspectral images such as the loss of detailed information in the process of model deepening,the number of labeled samples is insufficient,and the data imbalance between different sample categories,the 3D CNN based on the data balance augmentation and spectral-spatial fractal residual structure is proposed.The experimental results prove that the proposed three-dimensional convolutional neural network has obvious advantages in the classification of hyperspectral remote sensing images.The main research contents and innovation points of this paper are introduced as follows:(1)Three groups of convolutional neural networks are constructed to compare and analyze the feature extraction ability of different types of convolutional neural networks for hyperspectral remote sensing images.Different model groups are composed of convolutional neural networks based on spectral features(1D CNN),spatial features(2D CNN)and spectral-spatial features(1D+2D CNN,3D CNN).Each model group has different characteristics: the first model group is a fully convolutional neural network;the second model group uses the convolution layer,the pooling layer and the fully connected layer;the third model group uses convolutional kernels of different sizes.The experimental results show that the classification evaluation indexes—overall accuracy,average accuracy and Kappa coefficient of the three-dimensional convolutional neural networks on three hyperspectral remote sensing image benchmark datasets—Indian Pines,Pavia University and Kennedy Space Center are far better than other classification methods based on convolutional neural networks.The above operations pave the way for the subsequent model construction of hyperspectral remote sensing image classification based on 3D CNN.(2)The 3D CNN based on spectral-spatial dense connectivity strategy and spectral-spatial attention mechanism is proposed.In the network,the spectral-spatial dense connectivity strategy which can effectively extract hyperspectral features is designed;the spectral-spatial attention mechanism which can activate the useful spectral-spatial information and surpass the useless spectral-spatial features is proposed;in addition,a series of optimization methods such as the data augmentation,batch normalization,Dropout,exponential decay learning rate,and L2 regularization are adopted to realize the high precision of hyperspectral remote sensing image classification.The overall accuracy,average accuracy and Kappa coefficient of the proposed method on two benchmark datasets of hyperspectral remote sensing images—Indian Pines and Pavia University are more than 99%.Compared with the traditional methods and the deep learning methods,the proposed method shows obvious advantages.(3)The 3D CNN based on the data balance augmentation and spectral-spatial fractal residual structure is proposed.In the network,the data balance augmentation method is proposed to increase the amount of labelled samples and balance the sample number of different categories;the spectral-spatial residual strategy is designed to extract the spectral-spatial features effectively with lightweight parameters;the spectral-spatial fractal structure is designed to increase width of the model and prevent the loss of details;and the spectral-spatial dimension transformation module is designed to effectively reduce dimension of the hyperspectral features.The overall accuracy,average accuracy and Kappa coefficient of the proposed method on three benchmark hyperspectral remote sensing image datasets—Indian Pines,Pavia University and Kennedy Space Center all exceed 99%,reaching the state-of-the-art level.