Research On Snow Cloud Removal Method Based On Machine Learning And Analysis Of Spatiotemporal Change

As an important component of the cryosphere and one of the active factors,the sensitivity of snow cover allows it to quickly capture changes in meteorological factors such as temperature and precipitation,which profoundly affects the hydrological conditions of the entire region.As one of the three major snow-covered areas in China,northern Xinjiang is also affected by unfavorable factors such as global warming,so snow monitoring in the region is particularly important.However,due to the lack of remote sensing images caused by cloud occlusion,this paper mainly carries out research on cloud removal methods based on temporal-spatial filtering and machine learning,and generates cloudless snow products.On this basis,the trend and correlation between meteorological factors in the study area are analyzed.The main research contents are as follows:(1)Using the Google Earth Engine(GEE)cloud platform,a hybrid method of time-space cloud filtering is built to de-cloud the daily snow cover data of MODIS from 2003 to 2020 hydrological years(June of that year ～ June of the following year)in northern Xinjiang to realize integrated cloud removal in the cloud platform;Relative to the true value of snow cover in the weather station,the total accuracy of the generated cloudless snow cover products reach 91.47%.(2)In order to further provide a higher precision cloud removal algorithm,a machine learning cloud removal algorithm based on XGBOOST is constructed based on factors such as terrain and geography,and considering the differences between terrain and geography in different river basins,only the daily snow cover data of MODIS from 2000 to 2020(March ～May)in the Kaidu River Basin in northern Xinjiang is declouded.The results show that the accuracy of cloud removal is high,and the Kappa coefficient can be as high as 0.9863,which has obvious advantages over other mainstream machine learning algorithms.(3)Based on the above snow products generated after cloud removal,through the research and analysis of snow cover,it is found that 46.3% of the areas with snow cover days(SCD)of more than 60 days in northern Xinjiang in winter;In January,the proportion of snow cover(SCP)peaks.SCD and snow onset day(SOD)in western Xinjiang shows an upward trend,while eastern Xinjiang shows a downward and postponed trend.Overall,the trend of snow end day(SED)is on an upward trend.In terms of correlation,SCD has a positive correlation with temperature.The negative correlation with precipitation only appears in autumn and winter.In addition,the snow melt in the Kaidu River Basin is most obvious in the large and small Youledusi Basin,and the most obvious snow melt area is mainly distributed at 2200～3400 meters,and the snow melt season is postponed.Based on the above research and analysis,the hybrid method of time and space can remove cloud on a large scale,but the accuracy of declouding is slightly lower than that of the cloud removal method based on machine learning.At the same time,the cloud removal method of machine learning is more suitable for the snow melting stage.There is no obvious change trend in SCD,SOD and SED in northern Xinjiang;Temperature affects SCD more than precipitation.