Research On Remote Sensing Image Fusion Algorithm Based On Deep Learning

Geographic information management is based on satellite remote sensing technology in order to obtain remote sensing data for multidimensional monitoring of resource utilisation.In order to correctly reflect the use of resources in all regions of the world and to implement more efficient resource management of geographic information resources,it is necessary to have multispectral images with high resolution as the basis for data processing.However,due to the current level of sensor technology,remote sensing satellites are not yet able to collect such images directly.For this reason,satellites generally carry a variety of different sensors.Therefore,different image data can be provided,such as multispectral images with high spectral resolution and panchromatic images with high spatial resolution.Remote sensing image fusion technology,based on the similarity and information between two images,can overcome traditional hardware limitations and generate multispectral images with high resolution,thus achieving an accurate description of a specific scene.Remote sensing image fusion,as a pre-process in image processing,generates high-resolution fusion results with both sufficient spatial and spectral information to provide a key guarantee for the use and management of resources,which is also known as pan-sharpening of remote sensing images.The main work of this thesis is as follows.（1）A deep learning-based progressive general sharpening framework is proposed for coarse-to-fine reconstruction of multispectral images with high spatial resolution,combined with a channel attention module and a supervised attention module to enable the network to learn important spatial and spectral information and improve the information retention capability of the network.Finally a high spatial resolution multispectral image with complete spatial detail and uniform spectral distribution is generated.The first two stages of the network use an encoder-decoder design,echoing the channel attention mechanism,to complete the stitching and transmission of features in the channel domain.In the encoder,an inverse residual block is used instead of traditional downsampling to improve the expressiveness of the network,reduce the computational effort and improve the robustness of the network.In addition to designing skip connection between codecs,cross-stage fusion and supervised attention modules are added between each two stages,allowing the network to extract a more comprehensive range of features.In addition,a resolution enhancement network is added in the third stage to further enable the fusion results to cover rich spatial and spectral information.（2）A U-Net based method for deep information fusion of remote sensing images is proposed,using an end-to-end network structure to generate high quality pan-sharpened images.In order to perform deep extraction of spatial spectral information and reconstruct the fused images,the remote sensing image fusion in this work is represented as a task similar to an encoder-decoder framework,which can be divided into the following three tasks: 1)feature extraction;2)feature fusion;and 3)image reconstruction.This encoderdecoder-like network structure allows the differences and complementarities between multilevel features to be fully exploited for remote sensing image pansharpening.The input of the network is a panchromatic image with high spatial resolution and a multispectral image with high spectral resolution,and the output is a multispectral image with high resolution.Two modules,feature extraction and feature fusion,act as encoders to extract and encode the multilayer features of the multispectral and panchromatic images.The feature fusion module fuses the multi-layer features extracted by the network to reduce detail loss during downsampling.A channel attention module is introduced between the feature fusion and image reconstruction stages to focus the network on key information.Finally,the image reconstruction module decodes the fused feature map to produce a high spatial resolution multispectral image.