| The detection of clouds at different altitudes in the atmosphere is meaningful for many satellite remote sensing applications.Depending on different remote sensing data,different cloud detection systems have their own advantages and disadvantages.For example,the cloud detection algorithms based on visible light data can achieve high cloud detection accuracy,but it is difficult to detect thin clouds and clouds in the night sky.Although the infrared data is highly sensitive to clouds,cloud detection based on infrared data is still a challenging work because the infrared data is susceptible to background fields and its own resolution.This thesis relies on the Cross-track Infrared Sounder(CrIS)and the Visible Infrared Imaging Radiometer Suite(VIIRS)carried on the US Suomi National Polar-Orbiting Partnership(SNPP)satellite,and proposes a cloud detection method that combines a cloud detection index developed from CrIS physical features and three machine learning algorithms.In 2017,a method for solving the cloud detection index using the channel characteristics of the SNPP CrIS was proposed and proved to be able to detect the presence of thin clouds.However,in this method,only part of the infrared channels is used to qualitatively describe the existence of the cloud,without quantitative analysis.To solve the above drawbacks,we design a new quantitative cloud detection method in this thesis.The specific research contents are as follows:According to the characteristics of SNPP CrIS full spectral resolution(FSR)channels,we design a pairwise pairing method using full resolution channels,and select the channel pairs that can be used for cloud detection.Besides,in order to avoid the impacts from ground surface,sunshine and other background fields on CrIS data,a partitioning and time-division processing method is proposed in this thesis.In each region,the brightness temperature collected by paired channels are used to calculate the FSR cloud detection index(FCDI).In order to realize the combination of FCDI and machine learning,we propose a method of using VIIRS cloud product(CP)detection results to add cloud tags to CrIS data.VIIRS CP is a cloud detection product based on VIIRS data,which is widely used due to its high detection accuracy.In the machine learning method,three classifiers such as: Extreme Learning Machine(ELM),Support Vector Machine(SVM),and Multilayer Layer Perceptron(MLP)are selected to carry out binary classification of FCDI.Then the results are analyzed quantitatively and optimized.The current experimental results show that the detection accuracy of FCDI cloud detection method in the test set has reached about 78%.And through statistics,it is found that about 50% of the false positive samples contain a considerable amount of water vapor and ice,indecating that this part of the samples may not meet the conditions of clear sky.It may contain some optical thin cloud that cannot be effectively recognized by optical instruments,so the actual cloud dectection accuracy rate is higher than 78%.In this thesis,the feasibility of using FCDI for cloud detection and advantages of this method are verified,which provides a new idea for cloud detection using infrared data in the future. |
