Research And Application Of Stochastic Resonance And Deep Learning In Hyperspectral Image Shadow Information Extraction

Hyperspectral images(HSIs)contain abundant spatial information and spectral information at the same time,which has great application value and development space in military and civil fields.However,the shadow area caused by sun illumination seriously hinders the information extraction of HSIs.How to enhance the information expression ability of the target in shadow region of HSIs has become a problem to be solved.In recent years,dynamic stochastic resonance(DSR)has attracted the attention of most researchers because of its unique mechanism of enhancing weak signals with noise.Dynamic stochastic resonance theory holds that the existence of internal noise or external noise in nonlinear system can enhance the response of system output,which has been widely used in weak signal detection,digital signal enhancement,mechanical fault diagnosis,color image enhancement and so on.Based on this,this paper introduces the dynamic stochastic resonance theory for the first time to enhance the shadow areas of hyperspectral images,further mining the hidden targets and information in the shadows of hyperspectral images,and improving the information utilization of hyperspectral images.The main research contents and achievements of this paper are as follows:Firstly,this paper analyzes and studies the feasibility and specific methods of applying the dynamic stochastic resonance theory to the enhancement of shadow areas in hyperspectral images.This paper discusses the method of introducing the dynamic stochastic resonance theory into the study of shadow region enhancement from the spectral dimension perspective of hyperspectral images.Simulation results show that the target spectral lines in the shadow area of dynamic stochastic resonance spectral enhanced hyperspectral image have higher amplitude and more obvious spectral line characteristics than those in the shadow area of original hyperspectral image.At the same time,a variety of hyperspectral image classification methods are used to classify hyperspectral images before and after dynamic stochastic resonance spectral enhancement.The classification results also show that dynamic stochastic resonance spectral dimension shadow enhancement can effectively enhance the information expression ability of the shadow area of hyperspectral images.From the perspective of spatial dimension of hyperspectral image,this paper discusses the method of introducing dynamic stochastic resonance theory into the study of shadow region enhancement of hyperspectral image.Simulation results show that compared with the original hyperspectral image,the shadow area of dynamic stochastic resonance spatial dimension enhanced hyperspectral image has improved brightness and contrast.At the same time,a variety of hyperspectral image classification methods are used to classify hyperspectral images before and after spatial dimension dynamic stochastic resonance shadow enhancement.The classification results also show that dynamic stochastic resonance spatial dimension shadow enhancement can effectively improve the information expression ability of the shadow area of hyperspectral images.Secondly,the dynamic stochastic resonance spectral enhanced hyperspectral image and dynamic stochastic resonance spatial enhanced hyperspectral image could be fused to reduce the hyperspectral image spatial information loss caused by dynamic stochastic resonance spectral processing and the hyperspectral image spectral information loss caused by dynamic stochastic resonance spatial processing.The experimental results show that the spatial and spectral information continuity of hyperspectral images caused by dynamic stochastic resonance can be minimized by fusing the dynamic stochastic resonance spectral enhanced hyperspectral image and dynamic stochastic resonance spatial enhanced hyperspectral image in a ratio of about 3: 1.Finally,in order to make full use of the spatial and spectral information,and further compensate for the damage of spatial and spectral information of hyperspectral images caused by shadow enhancement in different dimensions by dynamic stochastic resonance of hyperspectral images,this paper also proposes a multi-modal convolution neural network(MM-CNN)for hyperspectral image classification,which has a doublebranch structure and can simultaneously extract 2D spatial mode and 3D tensor mode features from hyperspectral images.A series of experiments show that the multi-modal convolution neural network can make full use of the spatial information and spectral information of hyperspectral images,and can achieve higher classification accuracy and better robustness compared with many mainstream hyperspectral image classification methods based on deep learning and neural networks.