Reconstruction Of Snow Cover Information With High Spatial And Temporal Resolution At Watershed Scale Based On Machine Learning

Snow monitoring with fine temporal and spatial resolution has important guiding significance for watershed scale snow water resources management and sustainable utilization,natural disaster assessment and early warning in pastoral areas.The rapid development of satellite remote sensing technology provides an effective platform for watershed scale snow monitoring.How to effectively mine the important snow information hidden in massive remote sensing data is a major challenge to improve the performance of snow monitoring.How to improve the time resolution of snow cover data is the key to the snowpack monitoring of the watershed;For snow depth retrieval,how to cooperate with multi-source observation data and improve the spatial resolution of snow depth data is an urgent problem to be solved in the accurate management of snow water resources at the watershed scale.Therefore,taking Kaidu River Basin,an important animal husbandry base in Northern Xinjiang,as an example,satellite remote sensing big data and machine learning technology were used to realize the cloud removal algorithm by using snow grain size to fill the snow cover at the watershed scale,and realize the downscaling snow depth retrieval algoithm based on neural network,and the fine resolution snow data was applied to the research on the spatio-temporal trend of waterscale snow information.(1)Snow grain size filling and cloud removal based on space-time extra tree.Based on the spatial distribution characteristics of snow grain size in the Kaidu River Basin,we also design "Spatio-temporal" dimensional data that can fully characterize the geomorphological features of the Kaidu River Basin,construct and train a space-time extra tree to simulate the non-linear mapping relationship between multidimensional inputs such as elevation,slope,aspect,landcover,Spatio-temporal dimensional information and snow grain size in the Kaidu River Basin,and to conduct a daily snow grain size filling study under the cloud layer.Through model training under different data missing rates,the maximum cloud coverage applicable to the model in this study is determined to be 70%,and the importance score determines the optimal combination of input factors to estimate the snow grain size under the cloud layer,and the snow grain size that meets the threshold range is judged as snow,to achieve the effect of snow discriminating under the cloud.Ultimately,this study completed cloud removal of 66.75%of the snow products from 2000 to 2020.Compared with the MOD10A1 snow product,the annual average cloud coverage rate decreased from 52.46% to 34.41%,while the snow coverage rate increased from 21.52% to 33.84%,which greatly improved the time resolution of snow cover data without reducing the accuracy of snow identification.(2)Downscaled snow depth retrieval based on deep neural networks.The passive microwave bright temperature data and snow cover data,geographic and meteorological background field information with synoptic downscaling are used as the inputs,and the in-situ snow depth from meteorological stations in the northern Xinjiang are used as true values to explore the complex nonlinear relationship between the inputs and snow depth by training a deep neural network model to obtain a snow depth retrieval model with high accuracy and 500 m resolution for the northern Xinjiang-Kaidu River Basin.The experimental results show that the overall root mean square error of the snow depth obtained by the method is about 8 cm,which is better than the existing snow depth products(9.38?10.55 cm).In addition,by comparing and evaluating the subsurface and elevation distribution of meteorological stations with the overall geography of the Northern Xinjiang,the model is considered to have good applicability for snow depth estimation in the Northern Xinjiang-Kaidu River Basin,and can improve the spatial resolution of snow depth at the watershed scale.(3)Spatio-temporal trends of fine resolution snow information at the watershed scale.Snow information with fine temporal and spatial resolution generated by the above methods were applied to the study of the distribution characteristics and trend of snow phenology and snow reserves at the watershed scale.The overall characteristics of the snowpack in the Kaidu River basin show that the snow onset date becomes earlier,and delay in the snow end day,results in the increasing of snow cover days and the maximum continuous snow cover days.The snow depth shows a slightly increasing trend,while the maximum snow depth increases more strongly compared to the mean snow depth,while in the areas where the trend decreases,the maximum snow depth decreases more significantly compared to the mean snow depth.The snow area in the Kaidu River Basin showed a trend of decreasing and then increasing,while the snow storage showed a significant increasing trend.In addition,this study also further analyzes the distribution characteristics and trends of temperature and precipitation in the Kaidu River Basin from 2000 to 2020 were further analyzed,and structural equation models were used to quantify the extent to which each meteorological trend characteristic and geographic element contributed to the changes in snowpack parameters.The altitude has a significant impact on the trends of snow cover days,and the influence of precipitation and temperature trends were not reflected.In general,a watershed-scale snow reconstruction algorithm was developed to improve the temporal resolution of snow cover,and a watershed-scale downscaled snow depth retrieval algorithm to improve the spatial resolution of snow depth.The fine spatial and temporal resolution snow information is applied to the study of snow parameters trends in the Kaidu River Basin,and it is that the overall trend of snow cover,snow depth and snow storage in the basin increase from 2000 to 2020.