Study On Fine Mapping Of Air Temperature In Manas River Basin In Xinjiang With Sparse Samples

China has a vast territory and diverse landforms,and temperature prediction has always been a difficult and hot issue in complex terrain areas where there are no meteorological stations or where meteorological stations are sparsely distributed.Xinjiang lies deep inland,and snow-melting water resources is an important source of runoff in the basin.It is very important to study the snow-melting simulation model and timely and effectively determine the snow depth,snow water equivalent and other factors for regional climate simulation and disaster prevention and reduction.As the basic input data of snow hydrological model,meteorological data is the premise and guarantee of model research.In this paper,aiming at the preparation of air temperature spatial data in the simulation of snow cover-snow melting process in the Manas River basin in the middle part of Tianshan Mountains in Xinjiang,China,three models are selected to predict the temperature in the Manas River basin where there are few meteorological stations.The main work is as follows:(1)firstly,the air temperature environmental variables in winter and spring(2015.11-2016.4)were analyzed by least square correlation analysis,and the optimal factor set of latitude,altitude,slope,aspect and NDVI was determined by collinearity detection.(2)the spatial interpolation model of monthly mean temperature based on geographically weighted regression Kriging(GWRK)and generalized regression neural network(GRNN)is constructed.In this paper,80% of the observation data from 139 stations in the selected area are used as training data to train the GRNN model in different months.the regional air temperature spatial interpolation model of GRNN model for 6 months in winter and spring is established,and the remaining 20% sample data are used for error test and accuracy evaluation.(3)introducing the knowledge of machine learning,establishing the hourly temperature prediction model based on the long-term and short-term memory network(LSTM),training and predicting the future hourly temperature data,and combining with GRNN to prepare the hourly temperature distribution map of the study area.According to the above research results,the following conclusions are drawn:(1)based on the analysis of the results of monthly air temperature interpolationprediction model,it is found that both GWRK and GRNN models can fit the monthly mean temperature in the study area with high precision,and the fitting accuracy of GRNN model is higher than that of GWRK.(2)the analysis results of hourly air temperature prediction model show that the time series characteristics of air temperature can be accurately obtained based on LSTM,and the prediction can achieve higher accuracy in the case of sparse samples,which provides a new idea for the study of air temperature in areas with complex terrain and sparse stations.At the same time,the LSTM-GRNN combination model can also provide a high-precision hourly temperature prediction distribution map.