Research On Downscaling Of Precipitation Data From Multi-source Remote Sensing In Qinba Mountains, Shaanxi

Regional precipitation is a key component of regional water cycle and energy cycle.It is of great significance to accurately grasp the temporal and spatial distribution information of precipitation for the study of regional climate,ecology,hydrology and meteorological processes.Shaanxi Qinling-Daba Mountains area is one of the regions where precipitation is relatively complex and heavy rain disasters occur frequently in mountainous areas.Due to the interaction of complex topography and monsoon climate,precipitation in mountainous regions has strong spatial and temporal heterogeneity.Improving the monitoring ability of precipitation is one of the important problems of precipitation forecast and water resources exploitation and utilization in Shaanxi Qinling-Daba Mountains area.In this paper,the applicability of multi-source remote sensing precipitation observation in study area is compared and verified by using rainfall observation information such as ground rainfall observation station.The downscaling models of various satellite precipitation data based on linear,local linear and nonlinear are constructed in order to obtain 1km spatial resolution refined remote sensing precipitation data in order to reveal the temporal and spatial variation characteristics of mountain rainfall.First of all,select different remote sensing precipitation products,evaluate the precipitation products with good precision in the study area and analyze their error sources.Then,the suitable precipitation interpretation factor is selected to construct the descending scale regression model,and the high resolution remote sensing precipitation products are obtained through the residual correction.Finally,the downscaling results of each downscale model are compared and verified by using the precipitation measured at meteorological stations.The main conclusions of this paper include the following three parts:(1)The applicability of different precipitation products in Shaanxi Qinling-Daba Mountains area is quantitatively analyzed by using a variety of statistical indexes.a.In the daily time scale,TRMM and CMORPH on the precipitation of R2 was about 0.3.The accuracy is better than PERSIANN.The detection ability of TRMM for daily precipitation events is slightly better than that of CMORPH,PERSINN is the worst.The precision of daily precipitation of GPM is better than that of TRMM.It was the most obvious in winter.The daily precipitation event detection ability of GPM is also slightly better than that of TRMM,especially in the Guanzhong basin area.In addition,the satellite precipitation data have the best evaluation accuracy for autumn precipitation,which is slightly worse in spring and summer than in autumn,and is the worst in evaluating the precipitation process in winter.b.In the monthly time scale,the precision of TRMM is the best in the three satellite precipitation products of TRMM,CMORPH and PERSIANN,and its R2 is 0.85,Compared with the monthly precipitation products of TRMM and GPM from 2014.4-2016,the monthly precipitation precision evaluation R2 is 0.84 and 0.83,especially,the estimation accuracy of precipitation in winter is only 0.39 and 0.24.The GPM error of precipitation estimation is lower than that of TRMM.The estimation accuracy of satellite precipitation products in monthly time scale is generally higher than that in daily time scale.(2)It is helpful to improve the precision of the downscaling model by selecting the appropriate precipitation interpretation factor as the input variable of the downscale regression model.The precision of model downscaling depends not only on the regression relation of the model,but also on the input variables of the model.In the construction of MLR downscale model,the optimal precipitation interpretation variable combination,Lon,Lat,LSTDN and NDVI are selected as the input variable of the model by using the ArcGIS "exploratory regression" tool.In the construction of GWR downscaling model,the ArcGIS "Geo-weighted regression" tool is used to evaluate all possible combinations of input candidate interpretation variables,and finally the optimal precipitation interpretation variable LSTDN is selected as the unique input variable of the model.In the construction of downscaling model of machine learning,the exhaustive search method is used to evaluate the combination of all subsets of input variables,and finally the best combination of input variables,Lon,Lat,NDVI and LSTDN are selected.(3)The lkm spatial resolution satellite remote sensing precipitation products are obtained from the downscaling data of various downscaling models.The downscale precipitation results are verified by meteorological station data.a.Comparing the downscaling results of MLR model and GWR model based on linear regression in May and August 2015.The TRMM downscaling results of GWR model and the actual rainfall fitting coefficient R2 are 0.77 and 0.57.The GPM downscaling results and the actual rainfall fitting coefficient R2 are 0.69 and 0.63.Compared with the GWR model,the global linear regression MLR model has low downscaling accuracy and is not suitable for obtaining high resolution remote sensing precipitation data.b.Comparing the downscaling results of four kinds of nonlinear regression models based on machine learning in May and August 2015,it is found that the RF model based on ensemble algorithm has the best accuracy.The precision index R2 of TRMM downscaling results is 0.72 and 0.69 respectively.The precision index R2 of GPM downscaling results is 0.63 and 0.68.The CART model and k-NN model ranked second after RF,The SVM model has the worst precision and is not suitable for high resolution remote sensing precipitation data.c.The residuals correction of downscaling results is helpful to improve the precision of downscaling results.After residual correction,the precision evaluation index R2 of GWR model and four kinds of machine learning downscaling model are improved to some extent,and the error evaluation indexes,Bias,RMSE and MAE are reduced to different degrees.In general,the downscaling analysis of satellite precipitation data not only improves the spatial resolution of precipitation data,but also improves the accuracy of the data to a certain extent.Based on four kinds of remote sensing precipitation data,this paper analyzes the applicability of satellite precipitation data in the study area,taking Shaanxi Qinling-Daba Mountains area as a typical research area.Aiming at TRMM and GPM data,a perfect downscaling method is proposed.The remote sensing precipitation data with high spatial and temporal resolution have been obtained and and method is innovated.