Spatial Downscaling Of Satellite-based Precipitation Data Over The Yangtze River Economic Belt

As an important meteorological parameter,precipitation is not only a key forcing factor in hydrological cycle,but also an important input in various hydrological and meteorological models.Therefore,it is of great significance to obtain high-precision precipitation data with high spatial-temporal resolution for many aspects of research.The rain gauges data cannot reflect the spatial characteristics of precipitation,especially in mountainous areas where rainfall stations are rare.In recent years,spatially continuous grid precipitation data can be obtained by remote sensing technology,so as to overcome the limitation of rain gauges data.Currently,the Tropical Rainfall Measurement Mission(TRMM)provides the TRMM products with the spatial resolution of 0.25 °,and the latest generation measurement mission of Global Precipitation Measurement(GPM)provides global IMERG(integrated multi satellite retrievals for GPM)products with the spatial resolution of 0.1 °.The resolution of satellite precipitation data is too low(?10 km or lower),which makes it unable to match the high resolution requirements of some studies when applied to the hydrological and meteorological studies.In this paper,aiming to produce the precipitation products with 1 km resolution,we used the downscaling technology for TRMM 3B43 data and the fusion research for GPM IMERG data.The main contents are as follows:(1)Spatial downscaling of TRMM 3B43 by Multivariate Adaptive Regression SplinesSome current downscaling methods(e.g.,geographically weighted regression,GWR;random forest,RF)have the discontinuous problem especially for “ blocky artifacts” phenomenon existing in the downscaled results,which was unlikely to occur in nature.In this study,we introduced a new method termed as Multivariate Adaptive Regression Splines(MARS)to address this problem.In this study,geolocation(longitude,latitude),Digital Elevation Model(DEM),daytime land surface temperature(LSTD),nighttime land surface temperature(LSTN),and four remote sensing indexes were taken into account as environmental explanatory variables.Monthly TRMM 3B43 precipitation datasets were downscaled from 0.25° to 1 km during 2006-2013 over the Yangtze River Economic Belt.Residual correction and geographical differential analysis(GDA)calibration were also implemented to improve the results.MARS separated the study area into sub-regions by splitting the auxiliary variables and constructed a smooth spline model with the selected variables in each sub-region.Thus,MARS can avoid the “blocky artifacts” phenomenon of downscaled result with its smooth modeling ability.Meanwhile,the downscaled TRMM data showed the reasonable spatially varying pattern and denoised result.According to the validation results,the final downscaled results using MARS were more accurate than other used methods GWR,RF and reduced bias compared with the original TRMM.In addition,the results of the variable importance revealed that the geolocation variables had the highest importance in MARS model,followed by DEM.(2)Merging daily IMERG and gauge precipitation data by Gaussian Process RegressionThis study takes the Hubei Province as the study area.The daily IMERG satellite precipitation data with rain gauge data are merged on July 19,2016 and auxiliary variables,such as longitude,latitude,DEM,brightness temperature data,and interpolation data,are considered.In this study,two fusion schemes,point-to-surface fusion method and bias-correction fusion method,are proposed.Three algorithms are selected including adaptive spline multiple regression,random forest and Gaussian process regression.The results show that the point-surface fusion method is superior to the bias-correction fusion method,and the fusion results by Gaussian process regression are better than other two algorithms.The fusion results based on the Gaussian process method show reasonable change details,which accord with the spatial pattern of precipitation.The spatial resolution of the fusion result has been improved from 0.1° to 1 km,and the accuracy has been greatly improved compared to the original IMERG data.This research has certain significance for the fusion of multi-source precipitation data with high temporal resolution.