Cloud Detection Based On Deep Learning And Using S-NPP CRIS FSR Data

Cloud detection is an important and challenging problem in satellite remote sensing.Accurately detecting clouds and eliminating the cloud-contaminated data can effectively ensure the accuracy of the numerical weather forecasting system.In 2020,Tian et al.designed and derived the Full Spectrum Resolution Cloud Detection Index(FCDI)using the data of two instruments,the Cross-track Infrared Sounder(CrIS)and the Visible Infrared Imaging Radiometer Suite(VIIRS)on the US Suomi National Polar-Orbiting Partnership(S-NPP)satellite.This method has been proven to detect clouds quantitatively,but the research on datasets,channel pairing and training models is not perfect,resulting in a low final detection accuracy.This thesis further explores based on this method,redesigns the channel pairing method,establishes a more accurate dataset,and proposes a Feature Refinement Attention network(FRAnet)based on deep learning and attention mechanism.The main contents of this study are as follows.1.According to the channel characteristics of the full spectrum resolution channels of the CrIS instrument,a new algorithm is designed for channel pairing.Specifically,the cloud sensitivity level pairing conditions are relaxed,and the standard deviation of the simulated brightness temperature of the channel is no longer used as the filter condition,but the root mean square error(RMSE)of the linear fitting result is used to ensure that the channel pair has high linear fit under clear sky.Further,different RMSE thresholds are set according to the height to represent the physical characteristics of different band combinations.The resulting channel pairs have the potential to detect clouds located at different heights.2.During the construction of the data set,fully consider the impact of seasonal changes,and use the data of 37 days in four seasons as the training set to cover the brightness and temperature changes in different seasons.Increase the amount of linear fit data to obtain more accurate fitting coefficients.At the same time,the experience-based algorithm of Tian et al.to make cloud tags is improved.This thesis performs precise spatiotemporal collocation of CrIS data and VIIRS data to improve the accuracy of tags.3.To fully exploit the cloud detection potential of infrared channel pairs and further improve the accuracy of cloud detection,this thesis proposes a feature refining attention network based on deep learning and attention mechanism,namely FRAnet.The proposed FRAnet mainly includes channel expansion structure and feature enhancement structure.The channel expansion structure expands the input information by extracting the crosscorrelation information of the channel pair through the convolutional network.The feature enhancement structure assigns greater weight to important features through the attention mechanism,and pays attention to the relevant features with greater cloud representation ability.The network can utilize a large amount of sample data to mine the internal relationship between infrared channel pairs and achieve high-precision cloud detection.226 infrared channel pairs are obtained in this thesis,and the cloud detection accuracy of FRAnet is improved from 80.10% of Tian et al.to 87.81%.Among them,the detection accuracy of multi-infrared channel pairs is 5% higher than that of single channel pair,and the detection accuracy of FRAnet is more than 1% higher than that of the learning model used in the industry.The experimental results show that the 226 infrared channel pairs obtained in this thesis have great cloud representation ability,and the FRAnet model can effectively detect clouds.