Deep Learning Based Domestic High-resolution Remote Sensing Images Fusion

In recent years,domestic high-resolution remote sensing satellites have developed rapidly,and the space and time resolution of images have been continuously improved.They have been widely used in land,surveying,mapping,environment,and agriculture to provide data support for planning,management,and decision-making.Due to signal transmission bandwidth and limitations of imaging sensor storage,remote sensing satellites usually provide low-resolution multi-spectral images and high-resolution panchromatic images,while in many remote sensing applications(such as vegetation cover maps and environmental monitoring),there is a large demand for satellite images with high resolution and multi-spectrum at the same time.In general,researchers use fusion technology to obtain remote sensing images with both high resolution and multi-spectrum advantages.In view of the high spatial and temporal resolution of domestic high-resolution remote sensing imagery and the strong demand for satellite images with both high-resolution and multi-spectral characteristics,this article uses the high-resolution image data as an example to analyze high-resolution domestically produced high-resolution images.Study on fusion methods of remote sensing imagery,combined with deep learning theory,one of the current academic research hotspots,studies deep-learning-assisted image fusion methods,and uses convolutional neural networks in deep learning algorithms to improve resolution of panchromatic images,image fusion based on improved resolution images.The research content mainly includes:(1)Study the blind restoration method of remote sensing image based on deep learning.A low-resolution single-image blind restoration model of deep composite convolutional neural network was researched.The characteristics of low resolution and corresponding high-resolution images were simulated by neural network learning.This model can predict the multi-scale end-to-end feature map based on the whole image,and learn model parameters through sample training.In this paper,additional maximum pooling and upsampling operations are used to implement deep composite networks by connecting lowlevel and high-level feature maps.The repair model ensures the stability of spatial local input patterns through jump-connected filters.Use the method of modifying the learning rate to improve the convergence efficiency to improve the stability of the sample training.(2)Based on the foundation of the depth composite convolutional neural network model,the image fusion method is studied.The Gram-Schmidt transformation is based on the statistical analysis theory,the best matching of all bands to be fused is histogram,the multispectral image is converted to orthogonal space to eliminate redundant information,and the transformed components are in the orthogonal space.All of them are orthogonal,which makes them able to maintain the spectral information and at the same time have the advantage of improving the clarity of the fused image.(3)Using the method of remote sensing image restoration based on deep learning combined with Gram-Schmidt transform to conduct experiments and analyze the experimental results.The experimental results show that the image restoration method based on deep learning has a significant effect on the detail restoration of low-resolution images.Combining the Gram-Schmidt transform algorithm can effectively improve the spatial resolution of the fusion image,and the spectrum information in the original image can be better preserved.