A Cloud And Cloud Shadow Detection Method Supported By Surface Reflectance Dataset For Landsat 8 OLI Imagery

Clouds and cloud shadows are widely present in remote sensing optical images,which not only affect the quality of optical images,disturb and block land surface monitoring,but also seriously affect remote sensing applications such as vegetation monitoring,land use analysis and change monitoring.Clouds and cloud shadows have an inestimable impact on the reliability of the inversion results,especially in the process of quantitative analysis.The fixed threshold algorithm can’t respond according to the type of underlying surface in time.Their working mechanisms are relatively complex and computationally complicated,and sometimes has poor results on cirrus and their shadows.In view of this,in this paper,a cloud and cloud shadow detection method supported by surface reflectance dataset for Landsat 8 OLI imagery is proposed,it uses high-quality surface reflectance data as a reference to implement cloud and cloud shadow detection through a threshold calculation model.By analyzing the influence of cirrus on the top of atmosphere(TOA)of different bands,it is found that the cirrus has a severely influence on the TOA of the blue band,and when cirrus present the reflectance difference between the near infrared and blue bands will be significantly reduced.Aiming at this feature,this paper uses the 6S radiation transmission model to simulate the variation range of the TOA under different observation geometries and atmospheric conditions in clear sky,and performs multiple regression analysis on the maximum values of TOA changes in different observation geometries.Then the threshold calculation models for blue band and for the difference between the blue band and the near-infrared band is obtained.In the cloud detection,if the TOA of blue band is larger than the threshold,and the TOA difference between the near infrared and the blue band is smaller than the threshold,the pixel is determined to be a cloud pixel,otherwise a clear sky pixel.The initial cloud results are further optimized to obtain the final cloud results.Comparing the TOA of cloud shadows over different underlying surfaces in the visible and near-infrared bands,it is found that the spectral signatures of cloud shadows can vary with the signal characteristics of the underlying surface,leading to high spectral variability in cloud shadows.Based on land surface reflectance,the possible range of TOA reflectance for each clear pixel can be estimated using the radiative transfer equation under different atmospheric conditions,cloud shadow threshold calculation models for four bands of visible and near infrared is respectively fitted.If a pixel has a TOA reflectance smaller than the minimum value of the possible range for the clear condition,it can be identified as being shadow-covered.Different strategies are adopted for water and land,only use blue band to detect cloud shadow over water,and the intersection of the four band results is taken as the cloud shadow detection result over the land.Using the coexistence of cloud and cloud shadow to remove the outlier in the preliminary cloud shadow detection result.One hundred and twenty-five Landsat 8 OLI scenes covered by various surface types(e.g.,vegetation,water,soil,city,and desert)were selected to qualitative analysis and quantitative evaluation of the algorithm.Comparing the results of the proposed algorithm with those of the object-based cloud shadow detection algorithm(Fmask)recently developed for Landsat images,the results show that the most significant improvement is for cirrus and broken clouds detection,the proposed algorithm performs generally better than Fmask for cirrus and broken clouds detection and has more accurate detection of cirrus cloud on the edge of thick clouds,there is less omission and commission.For cloud shadow detection,the results of proposed paper has obvious advantages,and are more accurate and stable.Overall,the cloud and cloud shadow detection results of the proposed algorithm are significantly higher than the Fmask algorithm.Using the manual masks to validate the proposed algorithm,the results show that the detection accuracy of the proposed algorithm for different types of clouds over different underlying surfaces is overall high.The average produce’s accuracy and average user’s accuracy are 0.9282 and 0.7200,respectively.There are fewer omission,and the dilate operation leads to commission increase.The cloud shadow detection algorithm can not only accurately identify thick cloud shadows,but also has high detection accuracy for broken and cirrus cloud shadows.The average Kappa coefficient of the cloud shadow detection algorithm is 0.8202,which is in good agreement with the visual interpretation results.The average ommision error and the average commission error are 0.1074 and 0.1950,respectively.There are less misclassification and missing pixels.The experimental and verification results show that the proposed algorithm has obvious advantages in the Landsat 8 OLI cloud and cloud shadow detection,and can be extended to multiple types of satellite data.