Research On Remote Sensing Image Fusion Method Based On Deep Learnin

Hyperspectral images have been widely used in the fields of mineral exploration,environmental monitoring,and ground classification.Hyperspectral images usually have high spectral resolution,while the spatial resolution is low due to the hardware limitations of the imaging sensor.Therefore,the fusion of hyperspectral images and auxiliary images(such as multispectral images)has received greate attention.How to explore the complementary information of multispectral images and hyperspectral images,and realize the enhancement of the spatial resolution of hyperspectral images through spatial-spectral information fusion,is an important issue in the field of remote sensing images fusion at present.The current methods include model-driven and data-driven methods.Among them,multispectral-hyperspectral image fusion based on deep learning is a current research hotspot.This paper introduces representative fusion methods based on deep learning,and deeply studies the mechanism of multi-source remote sensing images fusion.On this basis,considering the complementary characteristics of multi-source remote sensing images,we propose new hyperspectral-multispectral image fusion methods,and develop a multi-source spatial-spectral remote sensing image fusion system.The main research results are as follows:(1)We propose a new hyperspectral-multispectral image fusion method based on a multisupervised dense recursive network.First,a residual unit is used to construct a recursive block,which increases the depth and expands the receptive field of the network,while greatly reducing the scale of training parameters.Secondly,the upsampling recursive sub-network is used to replace the traditional interpolation method to avoid the problem of possible spatial-spectral information loss.Finally,in the stage of image fusion,multi-supervised dense connections are used to make full use of the low,medium,and high-level features of the network,and better guide the network training at the same time.Simulation experiments show that compared with other CNN-based methods,this method can effectively reduces spectral distortion,and obtains better image fusion quality.(2)We propose a new hyperspectral-multispectral image fusion method based on a forward-backward channel attention network.First,we designs a forward-backward channel attention residual block(FBCARB).FBCARB redistributes weights in the extracted feature channels,so that channels with high-frequency information in the network get more "attention".At the same time,the forward feature channel is enhanced,and the enhanced forward feature and the backward feature after redistributing the weight are added as the final output.In addition,a non-local operation module(NLM)is introduced at the end of the network to capture the longdistance dependence of different pixels on the feature map.Simulation experiments show that the proposed FBCAN and NLM can effectively improve the quality of fusion image,and have a better effect on the reconstruction of image details and textures.(3)We designe and implement a multi-source spatial-spectral remote sensing image fusion system,which supports multispectral-panchromatic image fusion and hyperspectralmultispectral image fusion.The former can generate high-resolution multispectral images by fusing high-resolution panchromatic images and low-resolution multispectral images obtained by Gaofen-1,Gaofen-2 and Superview-1 satellites under real scenes.The latter implements the two algorithms proposed in this article.Finally,we compare the performance of our system and ENVI software on multispectral-panchromatic image fusion function,and the results show that our system performs better in fusion quality and processing time consumption.