| The booming development of hyperspectral remote sensing technology has made up for the shortcomings of the previous multi-spectral remote sensing technology which defects about less bands and wide gap,and can detect more detailed features of ground objects.So it can be said to be an important milestone in the development of remote sensing technology.Hyperspectral remote sensing images have the characteristics of“integration of maps and spectra” and higher and higher spectral resolution.They show great advantages in feature recognition and classification,and the classification of hyperspectral remote sensing images has become the main method for obtaining their feature information.Therefore,on the basis of existing research,finding better classification methods and effectively improving classification accuracy have become the goals that pursued by researchers.In view of the characteristics of hyperspectral remote sensing image data which has large amount and small number of labeled samples,also take the inadequacies of complicated algorithms of traditional image dimensionality reduction and classification into account,which utilize the spatial information deficiently.So,based on the end-to-end working mechanism of deep learning,this paper discusses the classification algorithm of hyperspectral remote sensing images based on the convolutional neural network model.The main work contents are as follows:In the classification of hyperspectral remote sensing images,aiming at the shortcomings that the loss of spectral information as well as the spectral and spatial features cannot be extracted simultaneously when using two-dimensional convolutional neural network(2D-CNN);Considering that the three-dimensional convolutional neural network(3D-CNN)does not need to perform additional band dimensionality reduction processing,the 3D convolution kernel is directly used to synchronously extract the spectral characteristics and spatial characteristics of the hyperspectral remote sensing image,and it can make full use of the advantage of hyperspectral remote sensing image which has the three-dimensional characteristics of “integration of maps and spectra”;In this paper,3D-CNN is used to carry out research on the classification of objects in hyperspectral remote sensing images,and the classification effect is better than that of2D-CNN.The classification accuracy has an improvement of one-to-two percentage points.In addition,compared with traditional image classification methods,the training process does not require any pre-processing and post-processing work,and directly processes the original hyperspectral remote sensing image 3D cube,which truly realizes the automatic classification of images and improves the operation efficiency.Furthermore,the influence of different network depths on the performance of network classification is discussed,and the problem of "network degradation" is verified,which lays the foundation for the further introduction of residual learning.In the process of network training,aiming at the network degradation problem that may occur with the increase of network depth,based on the principle of residual learning,this paper constructs a deep residual network based on hyperspectral remote sensing image classification.On the basis of 3D-CNN,this paper introduces a residual structure and increases the network depth appropriately,uses two forms of 3D convolution kernels to extract spectral features and spatial features,which effectively improves the accuracy of feature recognition.The residual structure includes the spectral residual module and the spatial residual module,which can learn spectral and spatial features separately and continuously.And the information transmitted along the "jump connection" structure in residual modules can protect the integrity of information to some extent.Compared with 3D-CNN,the problem of network degradation is alleviated while increasing the depth of the network,thereby further improving the classification accuracy.The designed network model was verified on three publicly labeled data sets.The results show that 3D-CNN can effectively use the semantic information provided by the joint spectral-spatial features,and it is the most effective to process the 3D data of hyperspectral remote sensing images.The deep residual network based on the principle of residual learning optimizes the network performance while increasing the depth of the network,and realizes the classification accuracy improvement again on the basis of3D-CNN. |
