GF-2 Remote Sensing Image Fusion Based On Convolutional Neural Network

GF-2 satellite is a high-level civil remote sensing satellite independently developed by China.By the fusion of its high spatial resolution panchromatic image and high spectral resolution multispectral image,it can generate multispectral image with high spatial resolution.For the remote sensing image fusion,due to the inadequacy of traditional algorithms in modeling,difficulty in achieving a good balance between spectral information retention and spatial detail enhancement,and performance limitations due to the introduction of a priori assumptions,research based on convolutional neural networks start to rise.Compared with traditional algorithms,these studies have achieved a better compromise between spectral information retention and spatial detail enhancement.However,the deficiencies of these studies in the upsampling of multispectral images,the use of multi-level features,and the capture of contextual information limit the performance of the model.In view of these shortcomings,this thesis studies of GF-2 remote sensing image fusion based on convolutional neural network.The main work and innovations of this thesis are as follows:(1)In fusion models based on convolutional neural network,in order to solve the problem that the multispectral image and the panchromatic image could not be fused due to their different sizes,the multispectral image is upsampled to extract features,inaccurate information is introduced,and the calculation is increased overhead;when reconstructing the fused image,only the features extracted by the last layer are used,the multi-level features are not fully utilized,which result in loss of information in the fusion result.In response to these two shortcomings,this thesis designs an encoder-decoder fusion model based on the Laplacian pyramid,extracts features on the scale of multispectral image to solve the problem caused by different sizes when fusing;two level features are used to predict two level residuals.The fused image is reconstructed step by step on two scales;the two level fused images are supervised at the same time to ensure the stable prediction of the second level residual.Comparative experiments show that the model achieve good performance on various indicators,while enhancing the spatial details,it better maintains the original spectral information.(2)In fusion models based on convolutional neural network,most of them extract features through simple cascades of nonlinear transformation layers in shallow networks,which result in small receptive field,the capture of contextual information is limited.In response to this deficiency,in the fusion model based on the Laplacian pyramid,the feature extraction structure of the encoder is improved using dense connections to obtain more context information through a larger receptive field.Combining the features of different layers of panchromatic images to inject richer details into multispectral images.Comparative experiments show that the model after optimizing the encoder,while maintaining the original spectral information,further enhances the effect of spatial detail enhancement at the original scale.