Research On Efficient Hyperspectral Image Classification Method Based On Spatial-spectral Features

Hyperspectral image classification task,as a basic research project in the field of remote sensing images,has received extensive attention from scholars at home and abroad.Using a spectral imager to image the terrestrial scene,hyperspectral images with hundreds of spectral narrow bands can be obtained.Hyperspectral images contain fine spectral information and rich spatial information,and are therefore widely used in many fields,such as geology and minerals exploration,anomaly detection,weather recognition,atmospheric science,agriculture and so on.However,on the other hand,it is precisely because of its rich information characteristics that hyperspectral images have problems such as too high spectral dimensions and less labeled data.So how to improve the performance of hyperspectral classification models is still a huge challenge.In recent years,with the development of deep learning,the convolutional neural network has been able to solve the above problems to a certain extent,but there are still some problems,such as the large amount of model parameters and calculation,and the bad performance in the classification of some categories.Therefore,how to fully mine the information of hyperspectral image space and spectral information with less computational consumption,and then improve the classification effect of the model is a key issue in the task of hyperspectral image classification.This paper conducts in-depth research on the above problems,mainly including the following aspects:(1)In view of how to fully exploit the information of hyperspectral image’s space and spectral information with less calculation amount,a lightweight operator,Involution is introduced from the perspective of spectrum for the first time,which is different from the traditional convolution kernel.Furthermore,in order to exploit the performance of the Involution operator,this paper innovatively proposes Involution-2D and Involution-1D based on pooling layers and 1D convolution.The former can more effectively combine the spatial information of hyperspectral images without increasing the amount of parameters,and the latter can greatly reduce the amount of parameters and computation without losing performance,making it applicable to more devices.(2)In order to give full play to the performance of these two new operators,an Efficient Spatial-Spectral Interaction Network(ESSINet)is proposed for hyperspectral image classification.The skip-connection and the improved Involution can make the spatial spectral information in the hyperspectral image more fully interacted.Finally,the experimental results on four public datasets demonstrate the superior performance of the ESSINet proposed in this paper.