Thin-cloud Removal For Remotely Sensed Optical Imagery

Remotely sensed optical images are inevitably affected by atmospheric conditions or clouds.However,when the clouds are thin,the ground objects underneath the clouds can be vaguely detected,making it possible to restore the clear image.The thesis aims at thin-cloud removal algorithms based on a single image and its application on data without a cirrus detection band.The thesis first reviews the slow feature analysis(SFA)algorithm,which can effectively remove thin clouds in the visible and near-infrared bands.Although the SFA algorithm can also be applied to data without a cirrus band,its cloud-removal performance is negatively affected.Thus,an improved SFA algorithm without including the cirrus band is proposed by introducing a cloud-free auxiliary image.Different datasets were analyzed for cloud removal,proving that the algorithm could effectively improve cloud removal accuracy without a cirrus band.The coefficients of the four visible bands increased from 0.6039,0.6635,0.7417 and 0.8750 to 0.7047,0.7430,0.8010 and 0.8706 at least.The algorithm is insensitive to the acquisition date of the cloud-free auxiliary image.Nevertheless,an auxiliary image having the closest acquisition or anniversary date with the cloud-covered data should be chosen to achieve the best cloud-removal results despite the algorithm's robustness.This thesis then studies an extension algorithm using linear discriminant analysis(LDA),which can remove clouds based on the original cloudy image itself.LDA is also based on image transformation,requires label data to know categories.This algorithm proposes to use the first band of the cloudy image itself to obtain the label data.Through a median filter,ground objects can be blurred and cloud information can be maintained.Both qualitative and quantitative evaluation indicates that the algorithm can effectively remove thin clouds,restore surface details,and is insensitive to filter size.In Landsat-8 real data experiment,the extension algorithm had better coefficients in five of the seven bands.When the data range is limited,block effects may appear after large-scale filtering,but the algorithm still maintains good cloud removal performance.Considering the data range and operation time in practical application,we should choose the label data with blurred ground features and without block effect.The quantitative evaluation of the algorithm generally requires a cloud-free verification image at least one revisit period apart.Then,this thesis selects the Landsat-8 images in the overlapping area of two adjacent paths with closer time intervals,which can shorten the time-lapse days to seven or nine.Results showed that the closer time interval does not mean more similarity between images.Thus,it is feasible to use reference images that are a revisit period apart.