Research On Cloud Detection Algorithm Based On Satellite Remote Sensing Images

Satellite remote sensing technology has an extremely wide range of application scenarios,such as change detection,object tracking,etc.Because remote sensing images are often polluted by widely distributed clouds,and this cloud occlusion phenomenon brings difficulties and challenges to remote sensing applications.Therefore,Cloud detection is a prerequisite for many remote sensing applications and an important research field in satellite remote sensing image analysis.The existing cloud detection methods have shortcomings in the detection of thin clouds and the discrimination of bright surfaces.This dissertation mainly studies cloud detection algorithms which are based on semantic segmentation with deep learning to improve the accuracy of cloud detection.In the detection of thin clouds with low recall rate,largely because of their small sizes,sparse distributions,as well as high transparency and similarity to the non-cloud background regions and for the bright surface,due to the reflection characteristics similar to clouds,the bright surface and cloud are easy to be confused and difficult to identify.In response to the above problems,this dissertation proposes a global context dense block cloud detection network based on the U-Net architecture(GCDB-UNet),and embeds the global context dense block(GCDB)in the U-Net architecture.GCDB is mainly composed of two feature extraction units: non-local self-attention block(NSB)and channel-attention block(Ca B).NSB calculates the correlation between pixels and uses weighted summation to gather discretely distributed thin cloud pixels to improve the feature expressio n ability of thin cloud pixels and Ca B extracts the correlation features of channels by calculating the weights of the contributions of different channels to cloud detection and make full use of the multi-channel information of satellite remote sensing data to improve the ability to recognize clouds and bright ground.In addition,GCDB uses a feature-fusion block(FFB)designed based on dense connections to extract multi-level and fine-grained features.Finally,the recurrent gated fine-tuning block is introduced to fine-tune the cloud detection results.Experiments indicate that the GCDB-UNet proposed in this dissertation effectively improves the accuracy of cloud detection.In order to reduce the false recognition rate of bright surface and further improve the accuracy of cloud detection,a multi-scales cloud detection algorithm based on chain residual pooling block(CRP)is proposed in this dissertation.The CRP is used to extract the multi-scale features.Then,GCDB-CRP-UNet cloud detection network is proposed by embedding the CRP block in GCDB-UNet,which can effectively extract the rich background context feature information of remote sensing images for improving the sensitivity of the model to the background information and enhancing the discrimination ability of cloud and non-cloud areas.Experiments show that the CRP can effectively improve the detection accuracy of cloud detection.Finally,based on the research results,a visualized cloud detection system was developed to support cloud detection of FY-3D satellite and MODIS satellite images.