Research On Cloud Detection Algorithm Of Optical Remote Sensing Image

Optical remote sensing data can be widely used in different fields such as phenology detection,ecosystem monitoring and urban expansion.However,optical remote sensing images are susceptible to interference from clouds and cloud shadows during the imaging process,and cloud and shadow pollution are important factors that limit the application of Landsat data.At the present,remote sensing image data usually include cloud and cloud shadow mask,such as the quality assessment band（QA）in the Landsat products,which is generated by the mainstream cloud and shadow detection algorithm Fmask（Function of Mask）.However,it is found that cloud mask products have cloud and cloud shadow missed detection and some wrong detection phenomenon.In order to improve the cloud and shadow detection accuracy of Landsat cloud mask,a new method was developed to improve the original QA band and further promote its application.At the same time,we further analyzed the impact of improving the accuracy of cloud mask products on subsequent related applications.Taking cloud pollution image restoration as an example,we analyzed the impact of different cloud mask products on the recovery accuracy of polluted pixels.In view of the above problems,the research work of this paper is as follows:This study developed a cloud and shadow detection algorithm（Auto-PCP for short）on the basis of the original QA band.We assumed that the new method would add additional time features and combine the information provided by the original QA band to further improve the quality of the QA band.The algorithm mainly has the following four steps: using the multi-temporal information provided by the reference image to construct the cloud index and cloud shadow index,respectively,to identify the initial cloud pixel and shadow pixel;The initial detection results can be refined by matching the initial cloud pixel and shadow pixel according to the geometric space positions of cloud and shadow pixel in the image.Some misdetected pixels in the original cloud mask were marked to further refine the detection results.The cloud and shadow detection performance of Auto-PCP was tested using the "L8 Biome" cloud and shadow validation dataset.The QA band processed by Auto-PCP（"QA＿Auto-PCP"）was compared with the QA band generated by other three methods,including the original Landsat QA band,the Landsat QA band modified by expanding two pixels at the edge of the cloud and cloud shadow,and the QA band generated by the multi-temporal cloud detection algorithm ATSA.The quantitative accuracy of cloud and shadow detection is evaluated by using cloud and shadow visual interpretation samples.The research results show that the Auto-PCP algorithm has higher cloud and cloud shadow detection accuracy than the other three algorithms,and other methods have large missed detection errors or misclassification errors in some images.Compared with the original QA band,the average F1-score coefficient value of Auto-PCP algorithm is improved by about0.81%.At the same time,the effects of different cloud mask products on the recovery accuracy of contaminated pixels were analyzed.Experimental results from real and simulated Landsat cloud pollution images show that the visual evaluation effect of cloud removal through QA＿Auto-PCP band is the best,and the RMSE value of the cloud restoration result is the smallest and the SSIM value is the largest.For example,in the restoration results of the MNSPI cloud removal method in the North China Plain,the RMSE value of the QA＿Auto-PCP restoration result is 0.02855,and the RMSE values of the remaining three methods range from 0.02957 to 0.03399.The QA＿Auto-PCP band can improve the accuracy of cloud recovery because the Auto-PCP algorithm improves the accuracy of cloud mask products.In addition,the auto-PCP algorithm is easier to implement and uses the same data removal as the cloud without the need for additional image collection.It shows that the Auto-PCP algorithm can further improve the detection quality of Landsat cloud mask and promote the application of Landsat data.