Research On Hyperspectral Object Classification Algorithm Based On Convolutional Neural Network

Hyperspectral images have more abundant characteristic information,and have been widely used in national defense construction,resource survey,urban development,precision agriculture and forestry and other fields in recent years.Hyperspectral image classification is an important research topic of remote sensing image classification.At present,it is the mainstream research direction to use convolution neural network to solve hyperspectral image classification problem.However,how to get effective classification results accurately and quickly is still the difficulty in the classification of ground objects.In this paper,the characteristics of hyperspectral image data are deeply discussed,the feature extraction method of hyperspectral remote sensing image and the classification algorithm based on convolution neural network are studied,and the model is analyzed and evaluated in the open hyperspectral remote sensing data set.The main achievements of this paper are as follows:(1)Aiming at the problems of high dimension of hyperspectral data structure.abundant sample information,resulting in redundant information and insufficient accuracy of ground feature classification,this paper proposes a hyperspectral ground feature classification algorithm,SENet-residual,which combines multi-dimensional convolutional neural networks.The algorithm is divided into two parts:multi-feature weight excitation module and multi-scale wide residual network module.The multi-feature weight excitation module collects rich feature information through feature selection,excitation and weight distribution,and performs feature fusion.Multi-scale wide residual blocks are obtained by merging convolution operations of different scales into the residunt network,and the multi-scale network structure is formed by combining a plurality of serially connected multi-scale residual blocks.The final results are weighted and fused,and the average value obtained is used to distinguish hyperspectral ground objects.The experimental results show that using SENet-residual to fuse features of different scales can effectively improve the average classification accuracy,and the weight incentive can also improve the quality of feature extraction to some extent.(2)To solve the problems of unreasonable feature extraction,unstable classification accuracy and long training time caused by the difference between spatial features and spectral features of hyperspectral remote sensing images,a hyperspectral image classification algorithm based on 3D dense total convolution(3D-DSFCN)is proposed.The algorithm extracts spectral features and spatial features by 3D convolution kernel in dense module,replaces the pool layer and full connection layer in traditional network by feature mapping module,and finally classifies them by softmax classifier.The experimental results show that the HSI classification method based on 3D-DSFCN improves the accuracy of ground object classification and enhances the classification stability of low-frequency tags.