Research On Multi-level Joint Processing Methods Of Cloud And Cloud Shadow In Optical Satellite Images

Remote sensing is considered to be the most important macro-monitoring tool for earth surface.However,as an important data source for satellite earth observation,optical satellite images are inevitably affected by cloud coverage due to the physical limitations of the imaging system.The presence of clouds and their projected shadows impedes the acquisition of useful surface information by optical satellites,resulting in the missing information in the acquired images,and reducing the availability of the images and causing biases in its subsequent processing and applications such as land change monitoring.Therefore,the detection and removal of clouds and their shadows are essential for the precise processing and analysis of optical satellite images.However,existing optical satellite images suffer from different degrees of cloud and cloud shadow detection accuracy problems,and the processing of multi-source images also put forward higher requirements for the generality of cloud and cloud shadow detection methods.In addition,the accuracy of multi-temporal cloud and cloud shadow removal methods in high spatial resolution images need to be further improved.Therefore,this thesis will focus on the specific problems of current cloud and cloud shadow processing methods,conduct research on multi-level joint processing methods of cloud and cloud shadow at the feature,method and application levels,respectively,to present available methods in practical applications,and provide theoretical and methodological support for the realization of high-resolution seamless satellite earth observation.The main contents of this thesis are as follows:1)In order to address the problem of low accuracy of cloud detection due to insufficient spectral information in images only contain visible and near-infrared bands,this thesis proposed a multi-feature combined cloud and cloud shadow detection method,which combines spectral,geometric,and texture features.Specifically,after the rough cloud mask is generated based on the spectral features,the edges of clouds in the mask are refined by guided filtering,and then the geometric and texture features are combined to remove the non-cloud bright objects,and finally the cloud and cloud shadow matching and correction are performed based on the result of shadow detection.The proposed method reduced bright surface commission and thin clouds omission in cloud masks,and achieve high-precision cloud and shadow detection in images with insufficient spectral information.2)In view of the fact that most of the current cloud detection methods are only applied to specific types of images,and it is difficult to generalize to multi-sensor images with a single method,this thesis proposed a deep learning cloud detection method based on multi-scale convolution feature fusion,which is a fully convolutional neural network with encoding-decoding architecture.Specifically,a multi-scale convolutional feature fusion module is designed and added to the decoder module to better integrate the multi-scale spatial and semantic information of clouds and cloud shadows.Besides,the combination uses of dilated convolution and residual network units make output results have better spatial details and models are easier to converge in training stage,respectively,enabling high-precision cloud and cloud shadow detection for medium and high-resolution images of multiple sensors.3)In response to the problems of spatial detail loss and color distortion in the thick cloud removal results of high-resolution images by existing methods,a multi-temporal thick cloud removal method based on joint radiometric adjustment and residual correction is proposed in this thesis.The cloud edges in the mask of target cloud-contaminated image are first optimized to ensure the spatial continuity at the edges of the area to be recovered.The local radiation of available complementary information in the auxiliary image is then stepwisely adjusted,and used to fill the area covered by cloud in the target image to achieve thick cloud removal.Finally,the filled area is corrected by global optimization based residual correction,to achieve the spatially seamless high-fidelity thick cloud removal in high-resolution images.4)With regard to the problems of cloud and cloud shadow coverage,geometric misalignment,and radiometric inconsistency faced in large-scale high-resolution urban remote sensing mapping,this thesis proposed the procedure and technical methods for high-resolution remote sensing mapping in large-scale urban areas,which cover a series of joint processing,including cloud and cloud shadow detection and removal,spatial-spectra fusion,image registration,and mosaicking,etc.Taking the mapping of the main urban area of Nanning,Guangxi with high spatial resolution domestic satellite images as an example,this thesis focuses on the processing of cloud and cloud shadow in experimental images,and the results of large-scale,high-quality and high-resolution seamless urban remote sensing mapping are finally generated.