Deep Learning-based Multispectral Image Super-resolution Research On Sentinel-2

Super resolution reconstruction(SR)refers to the technology of recovering high-resolution images from single or multiple consecutive low-resolution images.With the rapid development of remote sensing technology in agricultural detection and military strategy,sentinel-2 images have attracted more and more attention.The Sentinel-2 multispectral image is an important tool for surface exploration research.Its super-resolution problem has also become a hot spot.It researches the surface by collecting images with three resolutions of10 m,20m,and 60 m.However,the low-resolution images collected can not meet the current application requirements,so the Sentinel-2 image super-resolution based on remote sensing image fusion appears.The super-resolved image not only supplements the spatial information of the low-resolution image,but also maintains the spectral structure.It also enhances the intelligibility of the image and has important application value.However,most traditional super-resolution methods of Sentinel-2 images can only extract the shallow features and the quality of super-resolved images is not high.While the deep learning-based Sentinel-2 image super-resolution methods can automatically learn the deep feature information of the input image through the deep neural network and improve the quality of the super-resolved images.Also,the depth features of different sources can be effectively extracted by inputting high-and low-resolution images into the depth network together.Further,the network which extracts different image features are connected to form an end-to-end network structure,which can effectively fuse the extracted deep features and enhance the network’s representation capabilities.Therefore,this paper draws on the idea of fusion and uses deep neural networks to propose super-resolution methods of Sentinel 2 multispectral image based on deep learning.The main research contents of this article include:1)This paper proposes and implements the super-resolution algorithm of Sentinel-2multispectral image based on deep channel attention network.Traditional convolutional neural network-based methods treat the functions of each channel equally when dealing with super-resolution problems and these functions lack the flexibility to handle different types of information.Sentinel-2 image super-resolution draws on the idea of fusion and injects high-frequency information in high-resolution bands into low-resolution bands.Traditional convolutional neural networks lack discriminative learning capabilities across feature channels,which ultimately hinders the representation capabilities of deep network.At the same time,high-frequency features are more important than precision low-frequency features in accurately recovering Sentinel-2 images.Therefore,a channel attention mechanism is proposed to adaptively readjust channel features by considering the interdependence between channels,so that the network pays more attention to the fine details of high-resolution input.Then fused features are input into the residual network for training and experiments show that this method can obtain better performance in objective evaluation.2)This thesis proposes and implements the super-resolution algorithm of Sentinel-2multispectral image based on deep feature extraction network.The high-and low-resolution bands of Sentinel-2 images have different spatial resolutions and the spatial information scales are different.Therefore,these two kinds of images need to use different networks to extract features and fuse at the feature layer.In order to be able to fully extract the spectral and spatial features from the source image,the network has multiple branches and a main thread is proposed.First,the convolutional layers with different convolution kernel sizes are used to extract the spatial and spectral features of multiple input images.The features extracted from the main thread network fusion branch further obtain high-resolution images.Compared with other super-resolution methods,the proposed framework can simultaneously extract and fuse features,and obtain better performance in both visual effects and objective evaluation.3)This paper proposes and implements the super-resolution algorithm of Sentinel-2multispectral image based on deep non-local neural network.The Sentinel-2 image shows a wealth of surface information,so the relevant information between distant pixels in the image is also valuable.However,the traditional convolution operation mainly considers the local area of the input feature map and does not consider the correlation information between the pixels at a long distance.Therefore,a non-local operation module is introduced.First,shallow feature extraction is performed on the high-and low-resolution input of Sentinel-2multispectral image,and it is pre-processed to obtain rich feature information of multiple inputs.Then non-local operation is used on the fused features to capture long-distance spatial content information,and directly calculate the similarity of the two positional relationships to quickly capture long-range dependencies,getting more informative features and improving the performance of the network.Experimental results show that the proposed method achieves the best performance in terms of visual effects and various indicators.