| As an important part of the cryosphere,snow cover is highly sensitive to climate change,and has significant effects on the water cycle process and ecosystem structure.Snow cover is one of the important fresh water resources,runoff from snowmelt in spring usually accounts for 10-15%of the annual runoff in the watershed,and snow melt water solves the problem of fresh water shortage for about 1/6 of the world’s population.At the same time,snow cover is one of the main sources of global natural disasters,snowmelt flood seriously endangers the production safety of the residents in the lower reaches of the high-altitude mountain basin.Therefore,accurate inversion of snow depth and simulation of the hydrological process of snowmelt runoff have important practical influence on the development and utilization of local water resources and the prevention of spring snowmelt flood disaster.Snowmelt runoff in high-altitude mountain catchments mainly depends on snowmelt water supply.A large number of research results have shown that snow depth is more suitable for watershed scale runoff simulation than other snow data.Traditional snow depth data are collected by manual measurement or snow depth observation stations,but it is difficult to directly obtain snow depth in remote and high-altitude mountains.With the development of remote sensing technology and the accumulation of remote sensing data,it is possible to obtain large scale snow depth.The snow cover information obtained from optical remote sensing data has high spatial and temporal resolution,but is greatly affected by cloud and fog.Passive microwave remote sensing data is not affected by clouds and fog,it is convenient to obtain snow data,but usually its spatial resolution is coarse.Therefore,based on passive microwave data AMSR-2 and optical remote sensing data SSE mod FSC,combined with DEM data and CGLS LC-100 land cover type data,this thesis gives full play to the advantages of multi-source remote sensing data.Xtreme Gradient Boosting(XGBoost)model combined with Space Environment information to build the space environment extreme gradient regression tree(SE XGB)model,the data set(SE XGB SD)of 500 m daily snow depth in arid northwest China from 2013 to 2021is obtained by simulation.According to the above research results,taking the Xiying River basin of the Qilian Mountains in the arid region of Northwest China as an example,based on the snow remote sensing data(SSE mod FSC snow proportion data and SE XGB SD snow depth data),combined with the ERA5 atmospheric reanalysis data,this thesis used Extremely Randomized Trees(ERT)model and Artificial Neural Network(ANN)model to simulate snowmelt runoff in Xiying River basin.The results of this thesis show that:(1)The SE XGB spatial dynamic downscaling snow depth inversion algorithm constructed in this thesis has a remarkable inversion effect,and the snow information obtained is more consistent with the snow distribution law in space,which significantly reduces the lumpiness,and compared with other algorithms,the inversion accuracy and time consumption of this algorithm are significantly improved.Using the measured data of the weather station to verify,the evaluation coefficient R~2 between the measured snow depth data retrieved based on the SEXGB algorithm is 0.798,the RMSE is 2.81 cm,and the MAE is 0.90 cm.The results show that the algorithm has high accuracy in retrieving snow depth.(2)The ERT model and ANN model can better simulate the runoff in the area lacking observation data.The NSE of ERT model is 0.701,the RMSE is 6.228 m~3/s,the PBIAS is 4.903%;the NSE of the ANN model is 0.748,the RMSE is 4.554 m~3/s,and the PBIAS is 8.329%.The results show that the machine learning model has high precision and is an effective method to simulate watershed scale runoff.(3)The comparison of runoff simulated by ERT model and ANN model introduced snow cover remote sensing data shows that snow cover remote sensing data can improve the simulation accuracy of the machine learning model for runoff to a certain extent,especially in the snowmelt period.Taking the snowmelt period in Xiying River basin as an example,the NSE of ERT model increased by 0.099,the RMSE decreased by 0.369m~3/s,and the PBIAS decreased by 1.689%.The NSE of ANN model increased by 0.207,the RMSE decreased by 0.700 m~3/s and the PAIAS decreased by 1.103%.Snow remote sensing data were introduced into ERT model and ANN model to effectively improve the simulation accuracy of snowmelt runoff. |
