Thin Cloud Removal From A Single Remote Sensing Image Using NSCT Based Fusion

Clouds in the remote sensing images are often classified into two classes,namely thick cloud and thin cloud,according to their optical thickness.Thin clouds obstruct the earth observation of most optical wavelength satellite sensors,and their contamination significantly compromises remote-sensing image usability.The removal of thin clouds from a single image remains a challenging task given the lack of supplementary information and the complexity of ground features in images.In this study,we analyzed features of the thin cloud in remote sensing images on the spatial and frequency domain.Learning from the studies of single image dehazing in the image processing area,we proposed a novel method to remove thin clouds and restore ground information from a single visible remote-sensing image.Two image processing methods,namely,dark channel prior and mask dodging,are selected to derive initial inputs.Fusion based on nonsubsampled contourlet transform(NSCT)is applied to effectively remove thin clouds and preserve ground information.Special rules based on directional features of the NSCT are set to address the artificial feature attenuation in cloud-contaminated images.Based on the design,we experimented with Sentinel-2 true color images to examine the proposed method in various scenes.Qualitative and quantitative assessment demonstrates that the proposed method achieves the target of removing thin cloud and halo caused by thin clouds from a single image and restores ground information effectively.As the method is not designed for the sensor of Sentinel-2 and it involves few parameters,the method can be applied to optical images of other sensors.In the future work,the difference of visible bands' response to the thin cloud will be considered,and the fusion rules will be redesigned to achieve better thin cloud removal results.