Research On Panchromatic And Multi-spectral Image Fusion Method For New Satellites

In recent years,a variety of satellites carrying panchromatic and multi-spectral image sensors have been launched,and the characteristics of remote sensing satellite images changed significantly.Mainly for two aspects,the multi-spectral images of the new satellites such as the World View series have a more detailed band division and a larger number of bands,at the same time,the spectral overlapping range of the panchromatic image and the multi-spectral image becomes narrower and the spectral matching in the overlapping range becomes worse.These changes lead to the poor adaptability of the existing fusion methods.In view of this,the thesis studies the panchromatic and multi-spectral image fusion method for new satellites.Through the study and analysis of deep learning and its related theories,this thesis constructs a multichannel deep model using convolutional neural networks.This method achieves the fusion of the panchromatic and multi-spectral images for new satellites,and effectively preserves the spatial details of the panchromatic image and maintains the spectral characteristics of the multispectral image.The main research content and innovative achievements are summarized as follows:(1)According to the characteristics of new satellite images,this thesis proposes a panchromatic and multi-spectral image fusion method based on multi-channel deep model.Firstly,a threelayer convolutional neural networks is constructed,and the input image of the network are constituted by the panchromatic and multi-spectral images together.Then the network is used to extract the image features of each band of the input image,and it effectively solves the problem that the existing methods is limited by the number of image bands,and the poor spectral matching of the panchromatic and multi-spectral images.Secondly,the network training set is constructed by capturing the image blocks of the panchromatic and multi-spectral images,and it effectively solves the problem that the remote sensing images are not easy to obtain.Furthermore,by constructing a multi-channel deep model and training images of different sizes,the corresponding different weight and bias are respectively applied to different channels in the multi-channel deep model.Finally,the result of each channel are integrated through the pooling layer to finally obtain the fusion result image.The multi-channel deep model can synthesize the advantages of different channels and improve the space detail information and spectral information in the fusion result image,and it prevents the network training over-fitting and effectively realizes the fusion of the panchromatic and multi-spectral images for the new satellite.(2)Based on the above research,this thesis expands the research on the following three aspects to further improve the accuracy and stability of the proposed method.The main idea is as follows: firstly,the multi-spectral image is sharpened before constructing the training set to emphasize the high frequency information,it can make full use of the existing image information,including the spatial detail information in the multi-spectral image which is neglected in the existing methods.Then the number of the training images are increased to four times of the original by the data augmentation,it can reduce the instability of the network training effect and effectively prevents the over-fitting in the training process.Finally,the pooling layer is added into the multi-channel deep model,and the parameter self-learning method is designed.By learning the weighting coefficients of the pooling layer,the randomness of the artificial setting parameters is reduced.The above work can effectively enhance the accuracy and robustness of the proposed method.The experimental results show that the proposed method can effectively fusion the spatial details of the panchromatic image and the spectral information of the multi-spectral image,and adapt to the image characteristics of the new satellite.The fusion effect is better than the existing methods.