Research On Thin Cloud Removal Of Remote Sensing Image Based On Generative Adversarial Network

Multispectral remote sensing data has the characteristics of high spatial and temporal resolution,strong intuition,and large amount of information.It is widely used in marine supervision,agriculturzal investigation,environmental monitoring,military reconnaissance and other fields.However,the imaging process is easy to be affected by cloudy weather,which results in a part of the information on image to be obscured and seriously interferes with the effective application of the image.According to whether the ground surface is completely obscured by the cloud,the researchs of cloud removal for remote sensing image can be divided into two directions: thick cloud removal and thin cloud removal.Since the reflectance of the ground objects under the thin cloud still partially retains,it is more important to recover the information under the thin cloud from the image itself.Limited by the fact that cloud concentration is difficult to accurately estimate,there are still some problems in traditional thin cloud removal methods of remote sensing image,such as incomplete cloud removal and large loss of ground surface information.The existing cloud removal methods based on deep learning can solve this problem well,but they depend on the synthetic paired training examples,which cannot fully match the real data distribution,resulting in the difficulty to apply to the actual cloud removal work of remote sensing image.For this reason,this paper aims to abandon the supervised learning strategy based on data labeling pairs,build a "cloudy-clear" image conversion architecture under unsupervised deep learning,and convert the thin cloud removal task of remote sensing image into a conversion task of the ground surface information between cloudy and clear imaging states.In this paper,a method of removing the thin cloud with cloud features and image color perception for remote sensing image is proposed,which realizes the high fidelity reconstruction of the ground information under thin cloud in the image.The main content of the paper is summarized as follows:(1)A "cloudy-clear" image conversion architecture has been formed under unsupervised learning.In order to reduce the cost of training set,this paper extends the training object to the data domain dimension,constructs the image transformation model based on Cycle-Consistent Adversarial Networks,designs a low-cost training set creation method based on the real remote sensing image,and implements a complete unsupervised training framework to realize the conversion between thin cloud images and clear images.This image conversion architecture provides new ideas for the research of thin cloud removal under deep learning.(2)A thin cloud removal method under the "cloudy-clear" image conversion architecture is proposed.This paper designs a thin cloud feature extraction algorithm based on Generative Adversarial Networks,proposes an improved training scheme with thin cloud features,constructs image color consistency constraints in HSV color space.Based on the above conditions,this paper forms a cloud removal method with cloud features and image color perception ability,which can reduce the loss of ground information as much as possible and realize the accurate reconstruction of the ground surface under thin cloud at the same time.Finally,the effectiveness of this method is verified by experiments.(3)A comprehensive comparative experiment of thin cloud removal methods is designed,and application case tests are performed.Based on the proposed method,and comparing the classic thin cloud removal methods,this paper designs a comparative experiment.The cloud removal effect of the image is evaluated from three perspectives: visual effect,image quality and spectral characteristics.The remote sensing analysis cases such as vegetation coverage estimation and water extraction are performed,proving the effectiveness of the method,which has a positive role in improving the utilization of remote sensing image.