Research On Polarization Remote Sensing Cloud Detection Method Based On Machine Learning

Cloud is an important regulator of the earth-atmosphere coupling system.It is closely related to the earth's water vapor cycle,radiation budget,and climate change.In the process of retrieving the physical properties of aerosols and observing the ground object information from remote sensing images,cloud detection and removal must also be completed.Accurate cloud detection is one of the key steps in retrieving aerosols and obtaining ground object information.Aiming at the problem of inaccurate cloud detection over the bright surface of the polarization remote sensing experience threshold method,which is susceptible to subjective factors,this paper proposes a machine learning cloud detection algorithm combining active and passive polarization remote sensing satellites.Compared with the official POLDER3 algorithm,the algorithm in this paper is more sensitive to thin clouds over bright surfaces and can perform cloud detection more effectively.This method can also provide new ideas for the cloud detection method of my country's Gaofen-5 satellite DPC polarization load..The specific method is as follows:First,based on the high-precision cloud vertical data of the active remote sensing CALIOP polarization load and the multi-angle polarization data of the passive remote sensing POLDER3 load,this paper analyzes the cloud detection sensitivity of the polarization load-related channel.Then use their collaborative observation data to construct a multi-characteristic data set combining active and passive polarization remote sensing.The longitude and latitude of the data set match the bright surface desert.Among them,many features include the detection characteristics of the vertical direction of the CALIOP cloud;POLDER3's 490 nm,670nm and 865 nm waveband polarization reflectivity,443 nm,670nm and 865 nm waveband reflectivity,670 nm and 865 nm reflectivity ratio and 763 nm and 765 nm reflectivity ratio,multi-angle Observation characteristics.Secondly,the PSO-BP cloud detection model and the BP-Adaboost cloud detection model were designed and built,and the multi-feature data set combining active and passive polarization remote sensing was substituted into the model for training.The model test results show that the fitting coefficient of the PSO-BP cloud detection model is 0.928.When the particle swarm is iterated to the optimization,the prediction accuracy rate is0.935;when the number of weak classifiers of the BP-Adaboost cloud detection model is30,the cloud The minimum mean square error of the detection model is 0.08,and the optimal prediction accuracy rate is 0.94.Finally,use the MOD09A1 surface reflectance product to build a typical monthly average reflectance database,and explore the impact of the highlighted surface type on cloud detection.After substituting the conventional POLDER3 level 1 data into the machine learning cloud detection model,the results show that the consistency between the PSO-BP method test results and the official POLDER3 cloud detection product is 86.7%,and the consistency with the MODIS cloud detection product is 92.4%;BP-The consistency of the Adaboost method test results with the official POLDER3 cloud detection product is 87.2%,and the consistency with the MODIS cloud detection product is 93.5%.The study of the optical characteristics of the undefined pixels of the official method and the sensitive cloud pixels of the machine learning method found that the optical thickness of these special cloud pixels are mostly small,which confirms that most of the special cloud pixels are thin cloud pixels and also validates machine learning.The effectiveness of the method in the detection of thin cloud pixels over the bright ground surface.