Research On Hyperspectral Remote Sensing Image Classification And Interpretability Based On 3D CNN

Hyperspectral remote sensing images have the characteristics of high spectral resolution,high spatial resolution and "map unity",which are widely used in environmental protection,precision agriculture,marine monitoring and other fields.Hyperspectral image classification is one of the key technologies and research hotspots in the above application fields.Hyperspectral remote sensing images have the characteristics of multiple bands,serious redundancy of information between related or adjacent bands,and few effective labeling samples,which makes the task of hyperspectral image classification face severe challenges.Traditional classification methods cannot mine the deep and hidden features of hyperspectral remote sensing images,resulting in poor performance of classification methods and difficulty in solving the problem of small sample classification.3D Convolution(3D CNN)has excellent deep and hidden feature extraction capabilities,which can deeply mine the spatial features of hyperspectral remote sensing images,which effectively promotes the development of hyperspectral remote sensing image classification methods.Aiming at many problems in hyperspectral remote sensing image classification,this paper proposes two classification network models based on 3D CNNs,which can achieve accurate identification of ground objects by extracting more discriminative spatial spectral features.In addition,aiming at the "black box" problem of convolutional neural network model,the interpretability of hyperspectral remote sensing image classification network model is explored.The main research contents are summarized below:Aiming at the problem of serious information redundancy between bands and insufficient number of sample labels in hyperspectral remote sensing images,this paper proposes a hyperspectral remote sensing image classification method based on multi-scale 3D convolutional neural network,aiming to deeply mine the multi-scale spatial spectral feature information of hyperspectral remote sensing images,and then improve the classification accuracy of the method.Firstly,the dataset is preprocessed,and the original hyperspectral remote sensing image is reduced by principal component analysis method,aiming to achieve dimensionality reduction and redundant feature elimination.Secondly,a multi-scale 3D CNN model is constructed.The model consists of three layers of multi-scale network:joint spatial spectral feature extraction layer,spatial feature extraction layer,and spectral feature extraction layer.Finally,the network model training is implemented and the performance of the learned network model is tested.In this experiment,four commonly used hyperspectral remote sensing image datasets were selected from PU,KSC,IP and SA,and the performance of the four classification methods was compared from the perspectives of subjective analysis and objective evaluation.Experimental results show that the proposed classification method has good performance in the hyperspectral classification of small samples.Considering the depth global semantic features and local detail features of hyperspectral remote sensing images,combined with the advantages of residual network and Inception network,this paper proposes a hyperspectral remote sensing image classification model based on multilevel and multi-branch 3D CNN.Firstly,PCA is used to reduce the dimensionality of hyperspectral data.Secondly,a multi-level and multi-branch 3D CNN classification model is constructed,which mainly includes the residual network module,the improved Inception module and the nonlinear module.The model enhances the feature extraction ability of the network by improving the multi-scale convolution in Inception to multiple asymmetric convolutions.The addition of the residual network in the model causes the input identity to be passed to the middle layer of the network to avoid the disappearance of gradients.The use of nonlinear networks enhances the nonlinear expression ability of the model.Finally,the PU,KSC,IP and SA hyperspectral remote sensing image data are used to learn the network model and test the model performance.Experimental results show that compared with the four classification algorithms,this method has better classification performance from the perspectives of subjective analysis and objective evaluation.Aiming at the problem of poor interpretability of deep neural networks,this paper analyzes the interpretability of the proposed classification model from the perspective of band selection and feature visualization.First,the hyperspectral remote sensing image bands are obscured in groups of 5 and 10 adjacent bands.Through this strategy,the influence of different adjacent band groups on the classification accuracy of various feature categories is verified.Secondly,PCA and TSNE algorithms are used to visualize the dimensionality reduction of the original data and the spatial spectral features learned after the model,which visually shows that the features learned by the network model have better separability,and then proves the effectiveness of the proposed method.