Bayesian Rainfall Assimilation And Runoff Processes Analysis By Artificial Intelligence

To explore mechanisms of runoff generation and flood routing,accurate and reliable rainfall information,as the key input to different runoff components,is necessary especially under the background of global climate change and rapid urbanization.Due to the improvement in rainfall measurements and estimations,rainfall information synthesized from multiple sources becomes an effective way to take advanteages from various rainfall products and detect the spatial and temporal distribution of rainfall.In this study,a novel multiple precipitation merging framework based on the Bayesian method with gamma distribution is proposed.A total of seven scenarios are designed by using three types of satellite rainfall products,namely TMPA(TRMM Multi-satellite Precipitation Analysis)3B42,PERSIANN(Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks)and CMORPH(NOAA CPC Morphing technique).To demonstrate the feasibility and reliability of the proposed Bayesian merging framework,gauge measurements collected from 2010 to 2015 are assimilated with above seven senarios individually at spatial resolutions of 0.25°,0.1°,0.05°,0.025° and 0.01° in Jinhua River basin,Zhejiang Province.Except for accurate rainfall information,potential controlling factors such as antecedent soil moisture,vegetation coverage,topography and soil type can also be the key factors affecting the rainfall runoff mechanism.Reliable rainfall-runoff process can help to improve the hydrological forecasting effectively and manage water resources scientifically under the situation of vulnerable eco-system and stressful population.On the assumption that different hydrograph patterns are resulted from various rainfall-runoff processes,this study attempts to identify the first-order factor in the discharge dynamics.To make full use of the powerful learning ability and error tolerance of the artificial neural network(ANN)for data mining in complicated runoff generation and flood routing situations in Jinhua River,the analytical framework can be constructed by the combination of self-organizing map(SOM)and back-propagation neural networks(BPNN).The SOM is used for classifying the potential controlling factors and the BPNN is utilized to build the hydrological model for each cluster.Based on the cross-validation skill,flood forecasting models before and after classifications are assessed to obtain the best hydrological model for different flood patterns.This analysis can be a general framework for the identification of the key factor in runoff process.Even though the evaluation of hydrological forecasting usually focuses on peak flow periods,which are highly related to human safety and water power generation,the low flow periods significantly affect the water demand,water quality and sediment accumulation.In general,Boussinesq equation is always used for the analysis of recession limbs in flood events.The lower boundaries of the recession slope curves in short and long time are fitted in power function,of which the parameters can be linked to hydraulic parameters.Besides,it is a complementary material for limited underground experimental data in a basin scale.This study summaries the derivation of solutions for Boussinesq equations and identifies different recession mechanisms by investigating the recession slope curves in two contrasting basins,namely Jinhua River and Longquan River in Zhejiang Province,China.The summary of main procedures and results are listed as follows:(1)Three commonly used probability distribution functions,namely lognormal,gamma and Pearson III distributions,are evaluated with the empirical distribution of gauge measurements.Then,the best fitting function is used to modify the traditional Bayesian merging method.Seven assimilation scenarios are designed according to the combinations of TMPA 3B42,PERSIANN and CMORPH and the proposed Bayesian method with gamma distribution is introduced for rainfall merging.The performance of all merging scenarios gradually decreases with the increase of spatial resolution,however the seven scenarios in both temperal and spatial series perform consistently at different scales,indicating that the merging method is effective and insensitive to spatial resolutions.(2)The comparison of seven nonlinear merging cases at five spatial resolutions shows that:The performance of the merging product synthesized from PERSIANN,CMORPH and gauge measurements outperforms other merging scenarios,suggesting the compatibility of PERSIANN and CMORPH products is higher than that of TMPA 3B42;the contribution of satellite products to different merging scenarios becomes larger as the spatial resolution increases,especially for the identification of the occurrence and trend of rainfall.(3)Data mining for the heterogeneous data,including rainfall,runoff,antecedent soil moisture at different depths and normalized difference vegetation index(NDVI),is implemented by the ANN method to identify the main rainfall-runoff process and key factors.Classification based on the features related to runoff regimes can significantly improve the flood forecasting performance of data-driven model,indicating that the combination of SOM and BPNN models for identifying the controlling factors during rainfall runoff processes is feasible and applicable.(4)After classification,the results of flood forcesting at different forecasting lead times demonstrate that:With the increase of forecasting lead times,the dominant factor can be found out,nevertheless,this benefit will decrease once the lead time is beyond the time domain of the controlling factor.Different hydrograph patterns are dominated by different factors.The flood patterns ?-? are dominated by soil moisture around 255cm underground,NDVI and rainfall,respectively.Furthermore,the dominant factor may change with time even in the same hydrograph category.For example,the controlling factor changes from rainfall to soil moisture around 255cm underground for flood pattern ? when the forecasting lead time increases.(5)The optimal parameters in Boussinesq equation for flood events at short-term and long-term recession limb in both Jinhua River and Longquan River can be determined according to the scatter cloud plot.The recession slope in Jinhua River is much slower than that of Longquan River.On the other hand,the recession slope curve in an individual flood event in these two catchments is also fitted by Boussinesq equation and a linear relationship between the two parameters loga and ? is significantly confirmed.Moreover,the linear function between the two parameters is unique in different catchments.This research focuses on the whole process from the beginning of rainfall to the end of flood and includes some significant techniques to improve flood forecasting performance.Furthermore,the linear relationship in recession limbs can be applied for hydrological forecasting in ungauged basins.