Research On Limited Sample Classification Method Of Hyperspectral Image Based On Convolutional Neural Network

Hyperspectral remote sensing images have nanoscale spectral resolution and can reflect the subtle features of ground objects.Therefore,hyperspectral remote sensing images have great advantages in ground object identification tasks and are usually used in agricultural production,military target identification,mineral mapping,environmental monitoring and other fields.Hyperspectral image classification refers to the classification of all pixels in the image into specific categories,which is the primary task in many applications,and the accuracy of classification will directly affect the quality of production results.Therefore,accurate classification of hyperspectral images is of great significance and application value.Although the existing algorithms have achieved high accuracy on hyperspectral data sets with small image size,the classification effect of hyperspectral images with large size is not ideal in the case of limited samples.In this paper,from the perspective of making full use of the effective information in the limited samples,a two-channel classification model based on convolutional neural network is constructed to extract the joint multi-scale spatial and spectral features,so as to improve the classification accuracy of large-size hyperspectral images under the condition of limited samples.Specific research contents are as follows:Firstly,a high-resolution feature map mechanism was proposed to reduce the influence of noise introduced during the sampling of pixel blocks on the feature learning process,and a HRVDRN model was built by combining residual learning and variable dimensional convolution structure.Experimental results show that the overall classification accuracy of the proposed model on Houston 2018 and Ma Ti Wan datasets can reach 90.69% and 98.24%,respectively,when only 1% of the samples are used for training,and both of which are better than the other five algorithms.Secondly,a multi-scale feature weighted fusion module is designed,which adaptively adjusts the weights of different scale features by learning.By introducing this module into HRVDRN model,MS-HR-VDN model is constructed,which can extract and fuse multi-scale features in images and mine diversified features from limited information.The control experiment results show that the overall accuracy of the improved model is improved effectively on the two kinds of data sets.Finally,a 1D-ACNN model which can dynamically adjust the size of the convolution kernel is designed,and the complete spectral information in the original hyperspectral image is used as input to avoid the problem of the degradation of the ground feature discrimination ability caused by the destruction of the high-dimensional data structure of the spectral information in the dimension-reduction process.A two-channel network model was built by combining MS-HR-VDN and 1D-ACNN to realize the joint extraction of spatial and spectral features with more discriminability.The experimental results show that when 200 samples are selected from each category as training data,the overall accuracy of the two-channel network model on the two types of data sets reaches 84.19% and 86.81%,respectively.Compared with the five comparison algorithms,the two-channel network model can significantly improve the classification accuracy under the condition of extremely limited samples and has strong robustness.