Cloud Detection Method Based On Spectral Area Ratio For Landsat 8 Complex Background Conditions

Landsat 8 images have been widely used in many fields,including rock and mineral identification,extraction of mineral alteration and faults information,and temperature retrieval.However,clouds in remote sensing images may greatly affect the accuracy of inversion of various parameters and reduce the utilization of data.For example,in monitoring the dynamic changes of glaciers,it is likely to take glaciers as clouds mistakenly because both have properties of high reflection and low temperature.Therefore,accurate cloud detection is of great significance for using remote sensing images.Based on a thorough research of other methods,this dissertation proposes a cloud detection method for identifying Landsat 8 images.The dissertation begins with an analysis of the spectral signatures of various objects and clouds from the perspectives of laboratory spectrum and actual images and divides clouds into two categories:thick clouds and thin clouds.In terms of the spectral signatures,it is relatively easy to determine the thresholds of thick clouds with consistent spectral signatures under different underlying surface conditions;by contrast,as the spectral signatures of thin clouds are subject to the influence of underlying surface conditions,their thresholds are hard to determine and require separate settings according to image features.The radiation spectral signatures show that the emissivity of clouds is greater than that of ground objects in visible light but less in thermal infrared.Based on this,spectral area is used to describe the total amount of radiation and to construct a pixel-based Spectral Area Ratio(Sratio).In order to differentiate underlying surfaces,Normalized Difference Vegetation Index(NDVI)and Sratio are used to construct a two-dimensional scatter plot of images.In response to the difficulty of selecting thin cloud thresholds and different demands for different research purposes,three confidence coefficients-high,medium,and low levels-are adopted for cloud detection by setting separate thresholds according to the characteristics of each image.Overall,the following progress and conclusions have been made:(1)In view of the band-setting characteristics of Landsat 8,the Spectral Area Ratio method is used to represent the total pixel radiation intensity through the spectral area and highlight the differences between the underlying surfaces and clouds with ratios.Experiments have shown that this method can distinguish clouds from different underlying surfaces to address the inadequate utilization of spectral information in the original band threshold method.(2)A two-dimensional scatter plot based on Normalized Difference Vegetation Index(NDVI)and Spectral Area Ratio(Sratio)is constructed.This scatter plot can distinguish different underlying surfaces,thus solving the problem of selecting thresholds in the univariate frequency distribution curve under complicated underlying surface conditions.(3)The cloud detection thresholds are adjustable to meet the demands of cloud detection for different research purposes and improve the utilization of images,thus making up for the deficiency of fixed thresholds and poor universality in traditional methods.(4)The proposed method in this dissertation has been respectively validated by the regional images of high reflection lava,urban and coastal zones,glaciers and frozen soils.Along with Fmask Cloud Detection Method,QA Cloud Mask Method,and ACCA + Cirrus Cloud Detection Method,it has been tested in nine different experimental areas for accuracy evaluation and analysis.The results have shown that this method has a strong advantage both in visual effects and overall detection accuracy(improved by about 10%).