Research On Information Fusion Of Remote Sensing Images

With the rapid development of remote sensing technology in recent years,more and more satellites can provide rich remote sensing data for human beings.These data can be applied to many fields related to national economy and people’s livelihood,such as military,agriculture,and industry,and play an important role.On the other hand,due to the increasing variety of sensors,the number of multi-source remote sensing data is increasing.How to effectively fuse the complementary information and remove redundant information of multi-source remote sensing data has gradually become a hot research topic in the field of remote sensing.Information fusion of remote sensing images can enrich image information,help image understanding and interpretation tasks,and improve the confidence of decision-making,which has important research significance.Due to the limitations of physical technology,a single satellite sensor cannot obtain high-resolution multispectral images directly.A large part of satellites only provides highresolution panchromatic images and low-resolution multispectral images of the same scene.In order to obtain high-quality images with both high spectral resolution and high spatial resolution,researchers proposed to fuse panchromatic images with multispectral images,a technique also known as pansharpening.Regarding to the problem that existing pansharpening methods cannot preserve spatial information and spectral information at the same time,this paper proposes two novel and effective pansharpening methods by using Bayesian theory and deep learning tools,respectively.In addition,in view of the case that some panchromatic images have no corresponding multispectral images,inspired by the natural image colorization research,a panchromatic image spectrum restoration algorithm based on reference multispectral image is proposed.The basic work and innovations of this article are summarized as follows:1.We proposed an effective pansharpening model based on Bayesian theory.Based on three reasonable assumptions,three novel probability distributions are designed to characterize the relationship between the latent high-resolution multispectral image and the available low-resolution multispectral image/high-resolution panchromatic image,and combined into the Bayesian framework.In addition,a multi-level gradient operator is introduced into the proposed model,which can capture finer spatial detail information of the image to significantly improve the pansharpening results.Experiments demonstrate that,compared with other traditional pansharpening methods,the proposed model can obtain fusion results with sharper edge and less spectral distortion.2.We proposed a multiscale pansharpening network based on the Framelet framework,which is the first attempt to convert the pansharpening task to the Framelet coefficient prediction task.First,multiscale features of multispectral images and panchromatic images are extracted and aggregated,then they are fused to make full use of effective information.Additionally,we introduce a hybrid residual connection to ensure promising spectral and spatial information preservation.Specifically,we regard the combination of the appropriate coefficients of multi-spectral images and detail coefficients of panchromatic images as the initialized Framelet coefficients of fused images and apply a residual network architecture to learn the difference between the initialized and the latent Framelet coefficients.Experiments show that the pansharpening network proposed in this paper outperforms other pan-sharpening methods on real and simulated data sets.3.We proposed a panchromatic image spectrum restoration method based on reference multispectral images.This task is regarded as a problem of selecting color candidates at super-pixel level.First,we tackle the problem by the use of superpixel representation rather than pixel level.This improvement speeds up the processing,and enables the features extracted from superpixels more discriminable so that the selected color candidates are more reliable.Besides,we propose a variational framework to model the color selection by exploiting superpixel-based nonlocal self-similarity and local spatial consistency simultaneously.Experiments show that the spectrum recovery method proposed in this paper not only achieves better results than other algorithms in the field of natural image colorization,but also can achieve promising spectrum recovery of panchromatic images.