Research On Remote Sensing Image Fusion Algorithm Based On Regression Kriging And Convolution Neural Network

Remote sensing observation is an important means to obtain surface information.As an important carrier of obtained information,remote sensing images are widely used in the fields of geographic mapping,environmental monitoring,agricultural production,forestry planning,disaster prevention,infrastructure construction,national defense and security.However,due to the limitation of satellite sensor hardware design and other factors,remote sensing images have the problem of mutual restriction in spatial resolution,spectral resolution and temporal resolution,which greatly limits its application potential.Image fusion technology can make full use of the respective advantages of high-resolution panchromatic(Pan)image and low-resolution multispectral(MS)image to obtain remote sensing images with both high spectral and high spatial resolution.Therefore,this paper takes multispectral image and panchromatic image as the research object,starts from the fusion of multispectral image and panchromatic image,and discusses the purpose of improving the fusion quality of remote sensing image.The main contents of this paper are as follows:(1)The current image fusion algorithms are introduced and summarized,and the Component Substitution(CS),Multi Resolution Analysis(MRA),geostatistical-based analysis methods and network models methods are expounded elaborated.The basic norm and implementation methods of the method are discussed,and the typical algorithms are studied in detail.(2)An adaptive block-based surface-to-point collaborative kriging image fusion method based on quadtree is proposed.The algorithm is based on Area to Point Regression Kriging(ATPRK).For each individual region,regression kriging was used to process residuals using regression modeling of spatial information from highresolution PAN images.The fusion experiment is proceed using this algorithm and other traditional methods,whereafter the results are jointly evaluated from the two aspects of objective index and subjective vision.Experiments show that the fusion image fused by this method is of good quality and maintains the spectral characteristics of multispectral images.(3)A remote sensing image fusion method based on U-shaped network residual module is proposed.The algorithm uses Convolutional Neural Network as the backbone,adopts U-Net network structure,and introduces Residual Block structure to better extract image features.The multispectral image is upsampled to the same scale as the panchromatic image,the two kinds of data are superimposed as input for model training,and the model is used for fusion.Experiments were carried out using this algorithm with 5 other fusion algorithms,and the results were evaluated from both objective indicators and human vision.The results indicate that the algorithm has a good performance in preserving texture details while fusing image spectral features.