Research On Hyperspectral Image Classification Method Based On Deep Convolution Neural Network

In recent years,deep learning,especially deep convolutional neural network(CNN),has become more and more widely used in the field of hyperspectral image(HSI)classification,and has achieved better classification performance.However,with the increasing spatial resolution and spectral resolution of hyperspectral images,there are still problems such as insufficient utilization of spatial spectrum information fusion,redundancy of spectral information,high computational complexity and serious noise.And the deep learning based classification method always needs a large-scale dataset to support the training,and the insufficient number of samples will affect the classification accuracy.In response to these problems,this paper has carried out the following research:Firstly,as the basis of this paper,several classical CNN-based classification methods are analyzed and validated.They are the classification methods based on spectral information(SPE),based on spatial information(PCA1),and based on two types of fused spatial information(PCA1_SPE and PCA3).Through the analysis and verification of these methods,not only the advantages of the method of integrating the space spectrum information have been proved,but also a foundation for the subsequent research has been laid.Subsequently,a hyperspectral image classification method based on information measure(IM)is proposed.The preliminary selection based on the entropy and color matching function and the secondary selection based on the minimum mutual information are effective.Spectral image dimensionality reduction.And through the selection of the band to synthesize false color images,the visual prediction of the ground object information is realized.Compared with the SPE method,the total classification accuracy rate of the IM method has increased by at least 7%.Compared with the PCA1 and PCA1_SPE methods,the total classification accuracy rate of the IM method has increased by at least 4%.On this basis,a classification method based on information measurement dimension reduction and spectral information enhancement(IM_SPE)is proposed,and the spatial data input is further integrated into CNN.The classification accuracy is higher than that of PCA3 and IM methods.To some extent,the problem of spectral information redundancy is solved,and a better classification effect is obtained by using the optical spectrum fusion.Finally,the classification method based on deep migration learning is studied and analyzed,and a classification method based on deep migration learning and neighborhood noise reduction is proposed,which obtains an average classification accuracy rate of over 98% on small samples.Compared with the non-migration learning method and the PCA3 method,the total classification accuracy rate has increased at least 3% and 2%,respectively.This method reduces the computational complexity to a certain extent,and solves the problem that the training samples are insufficient and the noise is severe,resulting in low classification accuracy.Through the performance comparison experiments of all the above classification methods,it is verified that the proposed methods improve the classification accuracy to a certain extent,especially have the most stable and excellent performance on the classification of a high-dimensional small sample dataset.