Comparative Study On Temperature Simulation In Loess Plateau Based On Different Machine Learning Methods

Under the background of global warming,the simulation of near-surface temperature has become an important research content.It is of great significance and function to study the change of near-surface temperature for analyzing the evolution of climate and predicting the future climate change.The near surface temperature data are mainly obtained by meteorological stations.In the vast areas with complex terrain,meteorological stations are rare,so it is difficult to obtain the spatial distribution of near surface temperature data by interpolation or extrapolation of meteorological stations.Machine learning has a good advantage in temperature simulation,but different machine learning methods have different simulation accuracy in different regions.Therefore,the loess plateau with less meteorological stations is used to simulate the near surface temperature data of spatial continuous distribution by using remote sensing data and machine learning method.Therefore,this paper selects 13 variables of surface temperature data,longitude and latitude,altitude,slope,slope direction,normalized vegetation index,solar radiation,surface albedo and precipitation for MODIS 4 time phases as input data of three machine learning models:deep neural network(DNN),random forest(RF),support vector machine(SVM).the simulation results of three near-surface temperature are obtained,and the simulation results of three models are verified by using the measured data of meteorological stations.The optimal model for simulating near-surface temperature in loess plateau is compared.In addition,the temperature re analysis data of ERA-5?ERA-Interim and CSFR compare the optimal model temperature simulation results in change trend,spatial distribution and error.Finally,the near surface temperature of the Loess Plateau in 2003~2017 is obtained by using the random forest model with high precision.The main conclusions are as follows:(1)Compared with other temperature simulation methods,the random forest model has higher accuracy and an average absolute error of 0.91?.RMS error is1.06?,and the stability of the model in different regions is better.Compared with other temperature simulation methods,the input factors of the model are easy to obtain and have better universality.(2)The surface temperature data obtained by the satellite at night is very important to the near surface temperature simulation results.The surface temperature data obtained by the satellite at night should be selected as an important parameter to simulate the near surface temperature.This paper provides the basis for the study of the near surface temperature using the night surface temperature data.(3)The monthly average change trend of random forest simulation results is more consistent with the change trend of measured stations.The CSFR reanalysis data have obvious overestimation of the near surface temperature as a whole,and the ERA-5 and ERA-Interim temperature reanalysis data also have different overestimations.(4)The near surface temperature data simulated by random forest method are higher in spatial resolution and more reasonable in transition between values than CSFR?ERA-5 and ERA-Interim temperature reanalysis data.The applicability of random forest in simulating near surface temperature and the stability of the model are verified.(5)The highest value of temperature is distributed in the central part of the Loess Plateau,and the low temperature is distributed in the western and southern plateau mountains.The temperature is greatly affected by altitude;from the seasonal point of view,the average temperature of spring and summer in the Loess Plateau generally fluctuates As a result,the average temperature in winter is relatively stable,and the temperature rise is not large,and it shows a downward trend in autumn.From 2003 to2017,the average annual temperature in the Loess Plateau showed an upward trend year by year,from the spatial distribution of the annual average temperature.From the above point of view,the trend of warming in some areas is more obvious.