Research On Hyperspectral Image Classification Models Based On Hybrid Convolution And Attention Mechanism

With the fast progress of remote sensing technology,hyperspectral image has become one of the most popular research objects in the field o f remote sensing.It is obtained by using spectral imager,which can simultaneously image the target region in dozens to hundreds of continuous and subdivided spectral bands.Its imaging mechanism determines that it is a three-dimensional data structure.Since hyperspectral images add spectral dimension on the basis of spatial dimension,therefore,hyperspectral images not only contain a large amount of surface spatial information,but also contain extremel y rich spectral information.At present,hyperspectral data based on remote sensing has been widely used in many fields,such as ecological monitoring,municipal design,biomedicine,geological exploration and agricultural planting.However,hyperspectral images also have problems such as high data redunda ncy,lack of marked samples,mixed pixels and insufficient extraction of spatial-spectral information,which brings great challenges to the classification of hyperspectral images.This paper explores and studies hyperspectral image classification methods b y using various convolution forms and improving two classical residual structures and attention mechanisms.The main research contents are as fo llows:(1)In order to alleviate the problem that the number of model parameters used for hyperspectral image cl assification is large,the training time is long,and the number of training samples is dependent,this paper proposes a classification model based on improved SE-Net and depth-separable residual.The model first compresses the spectral dimensions of the o riginal hyperspectral image using PCA.Then,the multi-feature residual structure was connected through the three-dimensional convolutional neural networks.At the same time,the improved SE-Net module was embedded to ext ract the spatial and spectral featu res of the hyperspectral image.Finally,the extracted feature data is input into Softmax to activate the classification.By using DSC in residual structure and intr oducing GAP,the complexity of the network is reduced ef fectively and the reliability of th e model is improved.The experimental results show that the model achieves an OA of over 99% in all three publicly available datasets using a limited number of training samples.(2)In order to extract spatial and spectral features of deep hyperspectral images more effectively,a hyperspectral image classification model based on Res2 Net and the null spectral attention mechanism is proposed.The model still uses PCA to downscale the original image and enhances texture features by adding a spatial attention module to the 3D dilated convolution layer.The resulting feature data are then fed into two sets of SSDS-Res2 Net that combine the channel attention modules.Finally,the output features are converted into one-dimensional vectors by GAP and then passed through a Softmax classifier to obtain classificatio n labels.The experimental results show that the model achieves an OA of 98.95% and99.46% in two publicly available datasets,Indian Pines a nd Pavia University,using a much smaller number of traini ng samples,respectively.