Resolution Reconstruction Based On Deep Learning Using Jilin-1 Spectrum Satellite Images

With the application of remote sensing technology is becoming wider and wider,the demand for high-resolution remote sensing images is increasing.Therefore,how to obtain the high quality of high resolution remote sensing image is becoming a hot spot of research problem,traditional methods is high frequency information of the panchromatic image with a specific algorithm to extract,in the form of superposition or replace a component to make up for the problem of insufficient resolution multispectral data,the traditional methods of feature extraction algorithm is relatively fixed,however,is easy to appear spectral or space distortion.In recent years,with the gradual rise of machine learning methods,represented by convolutional neural network,it shows powerful complex data processing ability and potential of feature extraction,which provides a new method for the field of remote sensing image resolution reconstruction.Based on the idea of traditional image fusion,this paper proposes a two-channel network from the perspective of spectral and spatial components.A residual structure is added to the spectral channel to maintain the spectral characteristics of the original image,and a multi-size convolution kernel is added to the spatial channel.The multilevel features in the spatial channel are fused to extract the high-frequency spatial information in the panchromatic network,and at the end,the channel attention module is used to adaptively filter the features,the characteristic of the results with the traditional bicubic interpolation and GS method and compared the Pan Net classic panharpening network.At the same time,in order to quantitatively analyze the role of panchromatic images in resolution reconstruction,this paper uses Jilin-1 Spectral Satellite imagery to create two data sets of multispectral number and multispectral plus panchromatic,and train the network separately.Quantitative analysis is carried out from the two perspectives of image quality and band independence of the result image to evaluate the influence of panchromatic data on the reconstructed image.The results show that CNN has obvious advantages compared with the two traditional resolution reconstruction methods.The traditional methods often only focus on one of the spectrum or space,while the advantages of the CNN method are mainly reflected in the ability to achieve better results in both spatial and spectral components,to get a better overall quality.Compared with the Pan Net network,the proposed network in this paper has an increase of about 5% in the spectral component and approximately 6% in the spatial component.The comparison of the results of the two different data sets shows that the panchromatic image not only provides high-frequency information for the spatial component during the resolution reconstruction process,but also improves the spectral information to a certain extent.The increase in the spatial component is about 20%,the increase in spectral components is about 6%.In addition,the results of the 4x reconstruction of the two data sets are exactly the same.The presence or absence of panchromatic has little effect on the results,indicating that when the spectral information and panchromatic information in the data set are too different,it is difficult for the network to extract the corresponding high-frequency information to the multi-spectral images.In subsequent research,it is necessary to explore other network structures or methods to extract high-frequency information from panchromatic data more efficiently.