Cloud Detection And Information Restoration Methods For GF-1 WFV Remote Sensing Image Based On Deep Learning

Optical remote sensing imagery has become an important means of obtaining surface information due to its abundant spatial texture information features,and with its continuous development,it has been widely used in urbanization construction,land cover detection,crop yield assessment and so on.Cloud is the main obstacle that restricts the application of optical remote sensing image.High-precision cloud detection and corresponding area information restoration are of great significance to improve image utilization and expand the scope of image application.Cloud detection and cloud removal are essentially image segmentation and repair problems.Deep learning has strong feature extraction and analysis learning capabilities.After being introduced into the field of image segmentation and repair,its performance has far surpassed traditional algorithms.In the cloud detection task,due to Cloud are very similar to highlighted features,which are prone to misjudgments and missed judgments,resulting in a decrease in detection accuracy.In the subsequent information restoration task,there are situations such as difficult restoration of thick clouds in single-view images and poor visual consistency.In view of the above problems,this paper proposes a semantic segmentation network that integrates geographic information features and multi-scale context information in the cloud detection task.In the cloud removal task,the image inpainting method is improved according to the features of remote sensing images,and a cloud area information repair dataset of GF-1 is produced.The main research contents and results of this paper are as follows:(1)In this paper,a cloud detection semantic segmentation network is constructed based on geographic information features and multi-scale context information.This method can integrate geographic information and image information,expand the perception range of context information,and improve the distinction between cloud and snow and the overall segmentation accuracy.The accuracy of cloud layer detection in non-snowy areas of Gaofen-1 remote sensing images has been greatly improved,and the accuracy of cloud detection in snowy areas has been greatly improved.The average F1-score,Precision,Io U,and Kappa coefficient were 86.27%,93.12%,77.89%,and83.93%,respectively.Compared with the cloud detection semantic segmentation network CSAMNet,the comprehensive improvements are 2.00%,8.83%,1.93% and2.17% respectively.According to the quantitative and qualitative analysis results,the accuracy of the cloud detection results of this method has been improved to a certain extent.(2)This paper integrates gated convolution and coherent semantic attention mechanism into remote sensing image cloud information restoration network,adopts large-scale convolution kernel and modifies upsampling method,and proposes GF-1remote sensing image cloud information restoration network model GC-CSAGAN.This method can effectively distinguish invalid pixels from valid pixels,and enhance the correlation between repaired pixels.The experimental results show that the method in this paper can improve the visual consistency of the information restoration results of the cloud occluded area of the GF-1 remote sensing image,and its peak signal-tonoise ratio and structural similarity are 20.5294 and 0.7043 respectively.Compared with the PIC image inpainting model,the inpainting accuracy is improved by 1.0124 and 0.1131 respectively.