Research On Cirrus Cloud Detection Method Under Highlighted Surface Based On Deep Learning

The classical methods of cirrus detection,such as spectral threshold,are easily influenced by subjective experience,which leads to the low accuracy of cirrus detection at the highlighted surface.For cirrus detection based on deep learning,it takes a long time to obtain the training data set,the cost is high,and the labeled pixel results are subjective,so it is difficult to form a high-quality large-scale training data set and train a model with practical value,which is not conducive to the application of deep learning in cirrus detection.Aiming at these two problems,this paper proposes a deep learning polarization remote sensing cirrus cloud detection method based on radiation simulation data.Compared with the official product of POLDER3 data,the method proposed in this paper is more sensitive to the detection of cirrus clouds under the bright surface.This method can provide a new idea for China’s Gaofen-5 series satellites.The general method is as follows:First of all,this paper will study the vector radiative transfer model(MYSTIC)embedded in LibRadtran to accurately simulate the atmospheric ceiling radiation characteristics of clear sky and cirrus clouds under different land surface types,and build a polarization remote sensing simulation data set to provide reliable training data for polarization remote sensing cirrus cloud detection based on deep learning.Secondly,this paper will carry out the research on the cirrus detection algorithm of polarized remote sensing based on deep learning method,including the cloud detection algorithm model based on GA-BP and the cloud detection algorithm model based on random forest.Using the above-mentioned deep learning model,we will train the simulation data set of polarized remote sensing,focus on mining the deep features that can be used for cirrus detection in multidimensional information such as intensity radiation,polarized radiation and multi-angles,obtain more complex detailed features,and suppress noise information,so as to overcome the false detection of polarized remote sensing in ice,snow and other highlighted surface areas.Finally,in order to effectively verify the effectiveness of the cirrus detection algorithm designed in this paper,this paper will use the actual remote sensing data to test the established cirrus detection algorithm model,and compare it with the traditional spectral threshold method to verify the effectiveness of the cirrus detection method.The results show that the cloud detection algorithm based on GA-BP is more accurate than the cloud detection algorithm based on random forest,and the consistency with MODIS official products is 86.685%,while the consistency between POLDER3 official products and MODIS official products is only 64.818%.Compared with POLDER3 official product,the experimental results of cloud detection algorithm based on GA-BP are more consistent with MODIS official product,which shows that this method can be effectively applied to polarization remote sensing cloud detection.