A Method For High-Precision Evaluation Of Precipitation Driven By Remote Sensing Data

China is short of water resources,with uneven distribution in spatial and temporal,and does not match the distribution of population,productivity,and land.High timeliness and refinement of water resource evaluation is the basis for improving the rational allocation of water resources in spatial and temporal,but the spatiotemporal resolution of current precipitation products cannot fit the needs of water resources evaluation.Therefore,it is of great significance to explore and construct the precipitation datasets with high precision and spatiotemporal resolution for water resources evaluation.Compared with gauge observations and radar precipitation data,satellite precipitation products have become an important source of precipitation data because of their wide coverage,real-time performance,and more information.However,due to the limitations of sensors,retrieval algorithms,and spatiotemporal sampling,the satellite precipitation products suffer from random errors and systematic bias.This study took the Luanhe River Basin as the research region,and used the Integrated Multi-Satellite Retrievals for Global Precipitation Measurement mission(IMERG)and environmental variables with high-spatial resolution from 2010 to 2019 as the data sources.In this study,a high precision evaluation method of precipitation driven by remote sensing data was studied from the following four aspects: identifying environmental variables affecting precipitation,constructing a self-optimizing deep learning model for satellite precipitation downscaling,proposing an efficient and high precision correction method for preliminary downscaled precipitation,and dynamic evaluation of watershed precipitation based on time series remote sensing precipitation.The main contents are as follows:(1)Identifying environmental variables affecting precipitationDue to the problems of linear assumption,collinearity of multiple independent variables,and undetectable interaction between factors existing in current methods,this paper proposed a quantitative analysis method based on the Geodetector model.The explanatory power of individual environmental factor and their interaction on the spatial distribution of precipitation were explored by using the factor detector,interaction detector,and ecological detector.The results of factor detector showed that the explanatory power of latitude,longitude,slope,aspect,NDVI,and elevation on the spatial distribution of precipitation was more than 1%,and the influence of latitude,longitude,and elevation is the most significant.In the interactive detection results,except for elevation-slope in February and NDVI-longitude in early July,the interaction effect of other factors was enhanced,and those were more than 10%,which was significantly improved compared with the single factor.The results of ecological detector showed that there were significant differences in the effects of different environmental factors on the spatial distribution of precipitation.The above results showed that the Geodetector model can quantitatively reflect the influence of environmental variables on the spatial distribution of precipitation,and can be used for the factors selection of spatial downscaling.(2)Constructing a self-optimizing deep learning model for satellite precipitation downscalingTraditional downscaling methods have problems such as lack of nonlinear fitting ability,poor modeling ability of small watershed and accurate time scale of deep learning methods,and lack of downscaling research in ten-day time scale.Based on the statistical relationship between IMERG and NDVI,elevation,slope,aspect,latitude,and longitude,a parameter efficient selfoptimizing convolutional neural network model for precipitation downscaling was constructed.The model performance and the variation of model parameters were analyzed at annual,seasonal,monthly and ten-day scales.The results showed that,compared with the original IMERG,the similarity indexes of annual,seasonal,and monthly downscaled results were more than 0.94,0.89,and 0.69,respectively,and the precipitation of ten-day scale also was effectively represented.Compared with the China Gauge-Based Daily Precipitation Analysis(CGDPA),the average similarity indexes of annual,seasonal,monthly,and ten-day downscaled results were 0.58,0.78,0.68,and 0.47,respectively.The similarity of model parameters increased with the depth of model layers,which meant that the proposed model had good convergence,and it is useful for precipitation downscaling.(3)Proposing an efficient and high precision correction method for preliminary downscaled precipitationAiming at the influence of homogeneous part of precipitation field on precipitation downscaling and the errors and multi-scaling problems of downscaling residuals correction methods,a High Accuracy Surface Modeling(HASM)based on Bayesian parameter optimization(Bayes-HASM)was proposed in this paper to correct preliminary downscaled precipitation.The results showed that after residuals correction by Bayes-HASM,the scatter point of the downscaled precipitation was concentrated near the 1:1 line.The accuracy of the downscaled precipitation dataset at annual,seasonal,monthly and ten-day scales was improved,especially in autumn and winter.For accuracy indexes,Correlation Coefficient(R)and Index of Agreement(IA)reached about 0.9,and R and IA indexes of some monthly and ten-day scales increased by more than 0.8.In all time scales,the root mean squared error(RMSE)of the downscaled precipitation dataset decreased to below 20 mm,and the absolute value of BIAS decreased significantly.The above results showed that this method can effectively improve the accuracy of preliminary downscaled precipitation and eliminate the influence of homogeneous part of precipitation field on statistical downscaling processes.(4)Dynamic evaluation of watershed precipitation based on time series satellite precipitationSince the original satellite precipitation products cannot satisfy the need of refined precipitation evaluation,the satellite precipitation downscaling and correction method proposed in this paper was used to generate the annual,seasonal,and monthly precipitation products with high precision from 2010 to 2019.It is used to evaluate precipitation in the Luanhe River Basin from both spatial distribution and temporal variation.It was found that the precipitation in the Luanhe River Basin had a great variation in annual scale,showing a trend of fluctuation and decreasing,and the spatial distribution showed decreasing from southeast to northwest.From2010 to 2019,precipitation in spring and summer was complementary,while precipitation in autumn and winter showed an obvious decreasing trend.The monthly precipitation in study region changed significantly,and the precipitation in January,February,and December was very little,mainly concentrated in the Plateau area.From March to June,the precipitation increased and the rainfall concentration area moved southward gradually.The precipitation reached the maximum in July and August,showing a decreasing trend from southeast to northwest.From September to December,the precipitation decreased and the precipitation concentration area moved northward.During 2010～2019,precipitation in January,March,June,July,September,October,and November showed a decreasing trend,while precipitation in other months showed an increasing trend.The results showed that the spatial downscaling method proposed in this paper can effectively improve the accuracy and spatial resolution of original satellite precipitation products,and the downscaled precipitation can support the dynamic evaluation of precipitation in the basin.