A New Cloud And Cloud Shadow Detection Method Supported By Land Type Product Dataset

As an important climate change factor,clouds alter the energy transfer process of solar radiation to affect the radiation budget between the Earth's surface and the atmosphere.Therefore,clouds play a critical role in the radiation energy balance.Although satellite remote sensing technology is currently a crucial means of Earth monitoring,clouds can blur remote sensing images and even block ground information,which leads to failures in correct information expression and the accurate inversion of atmospheric and surface parameters.On this basis,cloud detection is essential for remote sensing image processing and is significant for the improvement of both the quality and utilization ratio of remote sensing data.The threshold methods always use a unified empirical threshold to achieve cloud recognition of all pixels on an image,or a threshold based on image statistics.Due to the complexity of the surface structure cloud state and the instability of the threshold based on image statistics,it is always difficult for such methods to effectively achieve high-precision of cloud detection,especially for thin clouds and clouds over high-luminance areas.To solve this problem,this paper proposed a Land Cover-based Cloud Detection(LCCD)and Land Cover-based Cloud Shadow Detection(LCCSD)algorithm.And the GlobeLand30(The 30-meter Global Land Cover Dataset)is used as the land type database in the paper.The main content of this paper is as follows:(1)Land Cover-based Cloud Detection algorithm.The algorithm is supported by surface type products,dividing the ground surface into constant attribute surfaces and changing attribute surfaces.For wetland,water bodies,artificial surface,bareland and ocean,the reflectance varies little with the seasonal and latitudinal features,so they are regarded as a constant attribute surface.According to their single-band characteristics and the indices that indicate the band characteristics,the objects use their best bands to assign corresponding fixed thresholds.And for cultivated land,forest,grassland,and shrubland,the reflectance varies greatly at different latitudes,seasons,and vegetation types.Therefore,the land surface types need comprehensive consideration of the influence of spatial-temporal changes,giving corresponding proportions of bare soil and vegetation at different latitudes in different seasons,and then determining a reasonable threshold.(2)Land Cover-based Cloud Shadow Detection algorithm.The cloud shadow detection algorithm needs to use the prior database of clear sky pixels and cloud shadow pixels selected from the Landsat 8 OLI remote sensing data,built by visual inerpretion manually.The cloud shadow detection algorithm is generated based on the difference between the reflection spectrum of the cloud shadow and the clear sky surface through statistical analysis,selection of the optimal band,and cloud shadow probability calculation.In the process of algorithm generation,the threshold is changed by a certain step size from 0 to 1,and the shadow accuracy and misjudgment rate of the data pixel database are changed according to the threshold change.If there is a band with a shadow accuracy higher than 0.95 and a misjudgment rate lower than 0.1,then the band is defined as one of the optimal bands.After that,calculation of the cloud shadow probability in the pixel database at each wavelength was done,and then single-wavelength cloud shadowing probability function was fitted and the standard error was used to assign weights to different bands.The final cloud shadow probability results were weighted and the certain shadow results were obtained with the probability higher than 0.85(3)Algorithm accuracy evaluation.Finally,this paper uses Landsat 8 data to validate the cloud and cloud shadow algorithm.First of all,the cloud and cloud shadow detection results are compared with false color images by visual interpretation,and the results show that the method can effectively identify the clouds and cloud shadows in the images,especially for thin or broken cloud and virtual shadows.Then,60 sample areas were selected to evaluate the experimental results.The results show that the detection accuracy of cloud and cloud shadow detection over different surface types is good,the overall accuracy is above 0.95.In detail,the correct rate and overall accuracy of cloud are 0.8903 and 0.9685 respectively.The types of surface with the highest accuracy are oceans,forests,and wetlands.The correct rate and overall accuracy rate of cloud shadow are 0.8672 and 0.9519 respectively.The types of surfaces with the highest accuracy are artificial surface,bareland,and shrubland.