Research On Cloud Removal Method Of Remote Sensing Image Based On Deep Learning

With the development of space remote sensing technology and electronic information technology,more abundant and accurate spatial data can be obtained by remote sensing.It has attracted people’s attention due to its advantages of being fast and having a large amount of information.However,the structural shortcomings of the existing satellite image imaging systems can easily lead to the generated images containing cloud noise.Remote sensing image is the most important data in earth observation task,in which cloud noise seriously interferes with ground object target imaging,which intuitively shows the problems of low image quality,fuzzy image information and partial information loss.These conditions greatly reduce the availability of remote sensing images.If these images cannot be recovered,a large number of remote sensing images will be wasted or their subsequent application will be wrong.In addition,due to the continuous development of electronic hardware equipment and the continuous enrichment of deep learning theory,deep learning technology plays an increasingly important role in remote sensing image cloud removal.Compared with traditional remote sensing image restoration techniques,deep neural network has the advantage of image restoration without manual measurement of other parameters.However,the existing thin cloud removal algorithms still have the problems of poor recovery effect and the system is too large and the processing time is too long.The existing thick cloud patching algorithms also have the problems of color deviation and obvious repair traces.So this paper uses the deep neural network technology and adopts two different ideas to remove the thin cloud and thick cloud respectively.The research contents are as follows:(1)In this paper,a lightweight multi-scale residual neural network model MSARDefog Net(Muti-Scale attention residual network using for cloud remove)is proposed for the task of removing thin clouds from remote sensing images.This algorithm creatively uses the multi-scale feature extraction module with a larger range of convolution kernel size variation,making full use of context information to estimate the current pixel point cloud thickness,and achieves the goal of removing thin cloud with less computing power.Moreover,the fine convolution module with channel attention mechanism provides the ability to extract fine texture features for the model.Experiments show that the images restored by the network proposed in this paper have more realistic colors,finer textures and better performance.In addition,a more comprehensive and rich data set PRSC(Pair Remote Sensing Images with Cloud)is proposed for the first time.This dataset contains many types of thin clouds,and the network trained with it can recover a clear image closer to the ground truth.(2)The information obscured by thick clouds needs to be repaired with the idea of image painting.This paper innovatively proposes a generative adversarial network TAGAN(Transformer Attention GAN using for Remote Sensing Image Inpainting)combining Transformer principle and convolution technology to improve the repainting effect.The model uses two Encoder-Decoder structures in series.The Encoder learns the required deep image features from the background information,and then uses the Decoder to reproduce the learned high-dimensional features into the required image features.The feature reproducing process is carried out twice.The first time uses the Transformer mechanism to learn the correlation between global pixels from the global background information.The location information added by this mechanism can make the training more stable and can make the network has stronger feature learning ability.The local similarity module is used for the second time to learn the similarity between background pixel blocks and foreground pixel blocks to ensure that the known information in the image is used to complete the filling of unknown information to the maximum extent.Experiments show that the proposed model can learn more complex textures from background information and repair more realistic remote sensing images.