Flood forecasting in large karst river basin by coupling 1 PERSIANN CCS QPEs with a physically based distributed 2 hydrological model 3

Abstract. There is no long-term meteorological or hydrological data in karst river basins to a large extent. Especially lack of typical rainfall data is a great challenge to build a hydrological model. Quantitative precipitation estimates (QPEs) based on the weather satellites could offer a good attempt to obtain the rainfall data in karst area. What's more, coupling QPEs with a distributed hydrological model has the potential to improve the precision for flood forecasting in large karst watershed. Precipitation estimation from remotely sensed information using artificial neural networks-cloud classification system (PERSIANN-CCS) as a technology of QPEs based on satellites has been achieved a wide research results in the world. However, only few studies on PERSIANN-CCS QPEs are in large karst basins and the accuracy is always poor in practical application. In this study, the PERSIANN-CCS QPEs is employed to estimate the hourly precipitation in such a large river basin-Liujiang karst river basin with an area of 58 270 km2. The result shows that, compared with the observed precipitation by rain gauge, the distribution of precipitation by PERSIANN-CCS QPEs has a great similarity. But the quantity values of precipitation by PERSIANN-CCS QPEs are smaller. A post-processed method is proposed to revise the PERSIANN-CCS QPEs products. The result shows that coupling the post-processed PERSIANN-CCS QPEs with a distributed hydrological model-Liuxihe model has a better performance than the result with the initial PERSIANN-CCS QPEs in karst flood simulation. What's more, the coupling model’s performance improves largely with parameter re-optimized with the post-processed PERSIANN-CCS QPEs. The average values of the six evaluation indices including Nash–Sutcliffe coefficient has a 14 % increase, the correlation coefficient has a 14 % increase, process relative error has a 8 % decrease, peak flow relative error has a 18 % decrease, the water balance coefficient has a 7 % increase, and peak flow time error has 25 hours decrease, respectively. Among them, the peak flow relative error and peak flow time error have the biggest improvement, which are the greatest concerned factors in flood forecasting.The rational flood simulation results by the coupling model provide a great practical application prospect for flood forecasting in large karst river basins.



Introduction
The highly anisotropic karst water-bearing media and intricate hydraulic conditions make the karst flood process exhibit significant differences in time and space, which led to the laminar flow and turbulent flow transmute into each other in karst areas, and the flood events in karst river basins are more complicated compared with that of in non-karst area (Ford and Williams,2007;Goldscheider and Drew,2007).This makes it difficult to precisely simulate Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.and forecast the karst flood process based on a hydrological model in mechanism.It is a common practice that the karst water-bearing media should be simplified before build a model.For example, making karst river basin as a multiple and nested spatial structure; making the underground river as the intelligible river system in the model; cave as the anisotropic medium with a large vertical infiltration coefficient and porosity but small specific yield.Even so, it is still hard to quantify the spatial structure of the karst waterbearing media with a physics-mathematics model.And the karst flood simulation results usually have some errors that could not be ignored, which is the main problem in flood forecasting in karst river basins (Kovacs and Perrochet,2011).
Because the dynamic change of karst hydrological process and the hydraulic conditions of underlying surface are complicated and non-linear in karst area, which makes it hard to obtain the hydrogeology parameters, such as specific yield, hydraulic conductivity and aquifer transmissivity and so on.With the rapid development of remote sensing, GIS technology and hydrogeology, the technology of field work including the tracer tests (Birk et al.,2005;Doummar et al.,2012) and infiltration tests have made a significant progress.However, it is still a challenge to accurately simulate the laws of motion of the karst hydrological process in the karst water-bearing media with these experimental tests.So the traditional methods such as lumped hydrological models are not suitable for flood forecasting in karst area (Hartmann et al.,2013).Compared with the performance of lumped hydrological models, the physically based distributed hydrological models (PBDHMs) have some advantages for karst flood forecasting in mechanism.The PBDHMs divide the whole karst river basin into a series of small grid units named karst sub-streams, which could reflect the real rules of hydrological process and karst development characteristics precisely.
Therefore, it has a great application potential to improve the karst floods simulation and forecasting capability (Ambroise et al., 1996).Many PBDHMs have been proposed since the blueprint of the PBDHMs published by Freeze and Harlan (1969).The first full PBDHM is regarded as the SHE model published in 1987 (Abbott et al., 1986a, b).Shustert and White (1971) used the PBDHM as an attempt in karst area, in their research, the dissolved carbonate species were analyzed in the waters of 14 carbonate springs in the Central Appalachians.And these springs were classified into diffuse-flow feeder-system types and conduit feeder-system types.The PBDHMs have been achieved many good research results in karst area (Atkinson,1977;Quinlan and Ewers,1985;Quinlan et al.,2011;Duan and Miller,1997;Ren,2006;Liu et al.,2013;Zhang et al.,2007).
Since the regulation and storage capacity of the karst water-bearing media are weak.
When the accumulated rainfall exceeds the maximum drainage capacity of the channel during a heavy rain storm, then the karst immersion-waterlogging hazard is much more likely to appear in this situation.And the hazard will become more and more serious with the intensification of global extreme weather events.So some effective measures need to be taken Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.
to reduce the flood losses.For example, simulating and forecasting the karst flood process reliably with a PBDHM effectively, it is an important non-project measure for flood control.
However, there is no enough rain gauges as well as the long-term meteorological or hydrogeological data to build a PBDHM in karst river basin where belongs to ungauged basin.
Prediction in ungauged basins (PUB) is the theme of international hydrological decade, the core of which is runoff calculation (Li and Ren, 2009).Therefore, it is more difficult to forecast the flood events in karst river basin compared with that of in non-karst area.How to solve the problem of rainfall source is a key factor of the current karst flood forecasting.The quantitative precipitation estimates /QPEs, especially the satellite QPEs technology brings the possibility to obtain the reasonable rainfall data in karst area.But the current application of the QPEs is not mature enough, which makes the accuracy of QPEs as well as the effect of karst flood simulation and forecasting are not so good.
Although many scholars at home and abroad have done a lot of research with the QPEs technology, also achieved many accepted results.However, there are considerable uncertainty exists in the application, which makes the precision of the QPEs is low and the precipitation result by the QPEs is not satisfactory.Two effective measures could reduce the uncertainty of the QPEs results in karst area.One is to match the appropriate resolution of the model.
Because the resolution can affect the result of the QPEs directly: if the resolution is too low, then the grid units divided are coarse, which causes a considerable error in rainfall estimates; if the resolution is too high, the meteorological model structure is complicated and unstable.
Furthermore, the requirement of computation resources will increase exponentially with the raise of the model spatial resolution (Chen et al., 2017), which leads to huge calculation and low efficiency.So the appropriate model spatial resolution is extremely important for the results of QPEs.And the other is the current technology of QPEs still has some systematic errors existed due to the uncertainties in structure and mathematical algorithm.For this reason, In this study, a post-processed method is employed to revise the PERSIANN-CCS QPEs products, which makes the result of QPEs more credible and receivable.
There are many researches on PERSIANN-CCS QPEs (Yang et al 2007) at present.But most of them have been used in small non-karst watersheds.In this study, the PERSIANN-CCS QPEs is employed to estimate the rainfall data as an attempt in such a large karst river basin -Liujiang Karst River Basin (LKRB) with an area of 5.8*10 4 km 2 in Guangxi province, China.Watershed flood forecasting relies on a PBDHM for a computation tool, while the precipitation is the model's driving force (Li et al., 2017).It has the potential to improve the accuracy of karst flood forecasting by coupling PERSIANN-CCS QPEs with a PBDHM.And the PBDHM in this paper is Liuxihe model (Chen,2009) (Chen et al., 2016) is used to optimize the coupling model parameters in this paper, which could control the uncertainty of the parameter passing.

PERSIANN-CCS QPEs
The original PERSIANN system (Hsu et al., 1999) was based on geostationary infrared imagery and later extended to include the use of both infrared and daytime visible imagery, which is an automated system for precipitation estimation from remotely sensed information using artificial neural networks .The system for rainfall estimation under development at The University of Arizona and gets constantly stronger with the improvement of the technology (Soroosh et al.,2000).The fundamental algorithm of PERSIANN system is based on a neural network.And the network parameters could be optimized by an adaptive training characteristic, which makes the precipitation could be estimated from geosynchronous satellite at any time and place.
The Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System /PERSIANN-CCS (Yang et al., 2004;Hsu et al., 2007) is a patch-based cloud classification and rainfall estimation system from low Earth-orbiting Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.and geostationary satellites by using pattern recognition technology and computer imaging technology (Yang et al.,2007 ).Satellite-based precipitation retrieval algorithms use information ranging from visible (VIS) to infrared (IR) spectral bands of Geostationary Earth Orbiting (GEO) satellites and microwave (MW) spectral bands (Hsu et al., 2007).
The QPEs products of PERSIANN-CCS has been generated precipitation estimates at Rainfall estimation from the PERSIANN-CCS consists as the follow steps (Hsu, 2007): (1) IR cloud image segmentation, (2) Characteristic extraction from IR cloud patches, (3) Patch characteristic classification, (4) Obtain the rainfall estimation results of QPEs products, (5) Evaluate and revise the results of QPEs products.
In this paper, the PERSIANN-CCS QPEs real-time data used in LKRB from the current version of PERSIANN-CCS are available and downloadable online (http://hydis8.eng.uci.edu/CCS/).

Liuxihe model
Liuxihe model proposed by Yangbo Chen (Chen, 2009)of Sun Yat-Sen University, China is employed as the fully distributed hydrological model in this study, which is a physically based distributed hydrological model(PBDHM)mainly for catchment floods simulating and forecasting (Chen et al., 2011(Chen et al., , 2016(Chen et al., ,2017;;Li et al., 2017).Liuxihe model earn its name by first successful application in Liuxihe catchment, Guangdong province, China.There are three layers vertically, including the canopy layer, the soil layer and the underground layer in the model and the whole catchment is divided into a great number of grid cells horizontally by using the DEM, which are treated as a uniform basin, and the elevation, land cover type, soil type, and other model elements including rainfall-runoff, evapotranspiration and so on are calculated on the uniform basin.All cells are categorized into three types, namely hill slope cell, river cell and reservoir cell.

2.3The improvement of the Karst hydrological model
Liuxihe model has been applied successfully for floods forecasting in many river basins.However, all these basins are non-karst areas.This is the first time the model is used in karst river basin as an attempt in this study.And the structure of the model should be improved to suit the karst basins.So some effective measures should be taken before building the model.
Firstly, simplify the karst water-bearing media, including making karst basin as a multiple and nested spatial structure, underground river as the intelligible channel system in the model, cave as the anisotropic medium with a large vertical infiltration coefficient and porosity but small specific yield, and fault as the anisotropic medium with a vertical, large infiltration coefficient and specific yield.Secondly, the whole karst river basin will be divided into many small karst sub-basins by the theory of distributed hydrological model.Furthermore, the karst sub-basins will be divided into many karst hydrology respond units (KHRUs), which are generally independent of each other.The whole karst hydrological process including the storage and regulation process of the epikarst zone, the spatial interpolation of the precipitation, evapotranspiration and rainfall-runoff are all calculated on the KHRU.After that, these hydrological processes will be summarized in the karst sub-basins.Then the outlet flow will be formed through the river confluence among each karst sub-basin from upstream to downstream.Such a multi-structure distributed hydrological model could utilize various scale information effectively and make the best use of the observed meteorological, hydrological and geological data.
The whole karst river basin is composed by many small karst sub-basins, then the karst sub-basins will be divided into many KHRUs.And each KHRU has its own model characteristics such as the meteorological and hydrological characteristics as well as the karst development characteristics in this study.The KHRU is proposed to describe the spatial variation of the karst sub-basins.And make sure that the differences within the KHRUs are smaller than of among the KHRUs.Then the KHRU is divided into five layers vertically: the canopy layer, the soil layer, the epikarst zone, the bedrock and the underground river.The sketch map of the KHRU is as follow: The structure of the KHRU (Ren,2006)   In order to satisfy the applicability of the model in karst area, the epikarst zone as a distinctive structure of the KHRU is considered carefully in the model.An exponential decay equation is used to calculate the regulation and storage process of surface karst zone.The linear reservoir model is employed to describe the regulation process of the superficial karst fissure system.And the Muskingum routing method is used to calculate the convergence process of the karst underground river that will be summarized and converge to drainage outlet through the underground river system.
The karst hydrological process of the epikarst zone could be divided into rapid fissure flow and slow fissure flow.When the precipitation falls to the surface karst zone, it will fill the pores of the macro crack firstly.After all the pores are full-filled, means the macro crack is saturated.This part of saturated water content named rapid fissure flow will go directly into the underground river through the macro crack, and ignore the regulation and storage hydrological process of the macro crack in this study.The rest of the water content will enter the tiny pores in the surface karst zone, and the water content of rapid fissure flow could be described as the following equation: Where epi SW is the water content of the rapid fissure flow in the epikarst zone, inf Q is the infiltration water content, and crk V is the water content in the macro crack.
The slow fissure flow in the epikarst zone is calculated by an exponential decay equation (Ren, 2006) Where sep W is the water content from the epikarst zone to the underground river, The linear reservoir model is employed to calculate the regulation process of the superficial karst fissure system, and the base discharge is calculated by the hydraulic gradient of the KHRU (Neitsch et al.,2000) : the supplies quantity of the base discharge that converge to the karst conduit or underground river on the i and (i-1) day respectively, epi K is the saturated hydraulic conductivity of the epikarst zone, The Muskingum routing method is used to calculate the convergence process of the karst underground river in this study, the equation is as follows: Where '   O is the water storage content, O is the outlet flow of the river reach, x is the dimensionless proportion factor, I is the inflow discharge of the river reach, K is the slope of the correlation curve of the water storage content and the discharge.
The finite difference method is used to calculate the water balance equation and the Muskingum routing method: Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.
where, 0 1 2 0.5 t = 0.5 t+ 0.5 t = 0.5 t+ 0.5 t+ = 0.5 t+ If the parameter of the Muskingum routing method K and x could be determined for a karst underground river reach, then the value of the

C
, which means the karst flood forecasting lead time will be 2Kx, then the Muskingum routing method could be simplified as follows: One of the key problems of Muskingum routing method is to optimize the parameters -K and x in the practical application.The least square method is used in this study: Where E is the objective function between the observed water storage content and the simulated one, which makes only require least squares approximation with regard to functional value, 0 () Wj and 1 () Wj are the observed and simulated water storage content at j period respectively, , n is the total numbers of the observation periods, C is the absolute value of the water storage content.
In order to simplify calculating, making A=K * x, B=K * (1-x), then taking the partials with respect to A, B, C respectively: Then the values of A, B, C could be calculated as follows: The parameters of the Muskingum routing method could be optimized through the above equations.And after that, the convergence process of the karst underground river could be calculated by the Muskingum routing method in Liuxihe model.

Study area
Liujiang Karst River Basin (LKRB) in southern China is selected as the study area in this paper.It is the second largest tributary of Pearl River that covers three provinces including Guizhou, Guangxi and Hunan province.LKRB is the most developed karst area of China with a drainage area of 58270km 2 and a channel length of 1121 km.The carbonate rocks are widely distributed in the southwest of the basin, and the areas account for 33% of the whole watershed.LKRB is a typical karst-mountainous catchment with frequent flash flooding in the past centuries .The peak forest-plain area is the main karst landform on the ground, while the karst conduit and fissure are well-developed underground, also there are many complicated underground rivers and springs with large flow (Li, 1996).The karst waterbearing media is highly non-linear and heterogeneous, which makes it very difficult to simulate and forecast the karst hydrological process.
LKRB is in the sub-tropical monsoon climate zone with an average annual precipitation of 1400mm to 1700mm, and the precipitation distribution is highly uneven at spatial and temporal scale.The precipitation from April to September accounts for 75% to 80% of the annual precipitation.
After studied the karst geomorphology of LKRB, Wil1iams (1987) believed that the peak-cluster depression had developed into turreted peak-forest landforms after a long evolutionary process, which is equivalent to the late prime of life, and going into the old age of geomorphologic evolution as the tradition physiognomy theory by Davis (1912).The allogenic water especially the Liujiang river is the main driving force for the development of peak-forest landforms.Therefore, the peak-forest plains and valleys are often distributed in contiguous areas near the main trunk stream of the Liujiang river.And the main karst landform of LKRB is peak-forest plain, there are also some peak-cluster depressions and peak-forest valleys.

Rain gauges and the karst flood process
There   is downloaded from http: //landcover.usgs.gov(Loveland et al., 1991(Loveland et al., , 2000)), and the soil type       2).Calculating the average precipitation of these 23 rain gauges.

Property data
Where, 2 P is the average precipitation observed by these 23 rain gauges; j P is the precipitation by the j rain gauge; M is the number of rain gauges.
3).The precipitations observed by the adjacent rain gauges are used to revise the results of PERSIANN-CCS QPEs with the following equation.method in this paper is a feasible and necessary.And it could greatly improve the accuracy of the coupling model in karst flood simulation and forecasting.Furthermore, the revise factor could be preserved as an empirical value for the future flood forecasting in LKRB.

5.1hydrological model setup
The method combining DEM with stream network leads to a more accurate drainage network from surface runoff modelling (Li and Tao,2000), especially in karst area.In this study, according to the high resolution of 200m*200m for Liuxihe model in LKRB, the whole studied area is divided into 1,469,900 grid cells named the karst sub-basins by using the DEM.
The grid cells include 1,463,204 hill slope cells and 6,696 river cells.Then the karst subbasins will be divided into many karst hydrology respond units (KHRUs) further, the KHRU is as shown in Figure 1.The river system are divide into three-order by Strahler method (Strahler,1957) as shown in Figure 3.
Because of the sinkholes and karst depressions in karst watershed, as well as the systematic error of the DEM itself, there are many pits including the true and false pits in LKRB.Among them the true pits are the karst depressions and sinkholes, they usually have a certain scale with elevation difference.While the false pits are just few points with low elevation, which is due to the systematic errors of the DEM.So the true and false pits should be distinguished reliably before using DEM data to divide into the karst sub-basins.Firstly, finding out all the pits with low elevation, and connect them into a plane, then distinguish the true pits from the false ones according to the on-site topographic survey.Finally,keeping the true pits like the sinkholes and karst depressions unchanged but filling the false pits in the model.
The karst hydrology respond unit (KHRU) is introduced in this study to reasonably describe the spatial variability of the karst water-bearing media (as shown in Figure1).The spatial characteristic of every KHRU has definite physical meaning.So the calculation of the evapotranspiration, rainfall-runoff and parameter optimization on the KHRU is also physically based, which could truly reflect the differences of the underlying surface.After the division of the karst sub-basins and the KHRUs, the post-processed PERSIANN-CCS QPEs results will be as the input data for Liuxihe model to simulate and forecast the karst flood process.The performance of the coupling model could be improved reliably in this way.

Parameter optimization of coupling model
There are many parameters need to be optimized for a distributed hydrological model, as shown in Table .2,among them the parameters of soil water properties, the epikarst zone and the underground river are the most sensitive parameters for the coupling model in this study.
The parameters of the epikarst zone are the most complicated due to the anisotropy of the karst water-bearing media, which makes it hard to measure and calculate the hydraulic characteristics.According to the field survey, the epikarst zone is mainly developed on the hard surface of pure carbonate rock, especially on the Paleozoic limestone.The thickness and characteristics of the epikarst zone are different due to the different climate, topography and landforms.And the thickness of the epikarst zone is about 10 meters in the study area-LKRB.
The parameters of the coupling model are listed inTable 2.

Table 2. The parameters of the coupling model
The parameters of the Soil type like the saturated water content and field capacity are calculated through a software tool by the research result of Saxton (Saxton et al.,1986) .The statistical relation between the soil texture and the soil water could be queried easily in the software tool.And it has been effectively proved by many experiments (Servat and Sakho,1995), the calculated value of this method has a good fitting relation with the measured value.
Liuxihe Model has been deployed on a supercomputer system with parallel computation technology (Chen et al., 2011) .An improved PSO algorithm (Chen et al., 2017)   are 0.86, 0.86, 18%, 4% , 0.91, and -7 hours respectively (as shown in Table 3).This means parameters optimization with the improved PSO algorithm in this study is effective.From From Figure 11, it could be seen that the simulated flood discharges with the precipitation of rain gauge are better than that of the PERSIANN-CCS QPEs.And the simulated peak flows with the PERSIANN-CCS QPEs are lower than the observed ones.
However, the flood simulation effects with the post-processed PERSIANN-CCS QPEs make a great progress, the simulated flood processes fit the observation values reasonably, and simulated peak flows are much closer to the observation ones.It implies that the flood forecasting capability has been largely improved by the post-processed method of the PERSIANN-CCS QPEs.
To further compare the flood simulation results, six evaluation indices are calculated and listed in Table 3.It has been found that all the six evaluation indices of rain gauges are better than that of PERSIANN-CCS QPEs.And the indices of QPEs have been improved a lot with the post processed QPEs.The average value of Nash-Sutcliffe coefficient has a 7% increase, the correlation coefficient has a 8% increase, process relative error has a 6% decrease, peak flow relative error has a 14% decrease, the water balance coefficient has a 5% increase, and peak flow time error has 7 hours decrease, respectively.Among them, the peak flow relative  From the above results in Figure 12, it has been found that the simulated flood results with re-optimized parameters by the post-processed PERSIANN-CCS QPEs are much better than that of with the same parameter as rain gauge precipitation.The simulated flood discharge processes, especially the peak flows with the re-optimized parameter are closer to the observation values.To further compare the flood simulation results, six evaluation indices are calculated in Table 4, the average value of Nash-Sutcliffe coefficient has a 7% increase, the correlation coefficient has a 6% increase, process relative error has a 2% decrease, peak flow relative error has a 4% decrease, the water balance coefficient has a 2% increase, and peak flow time error has 18 hours decrease, respectively.What is more, comparing with the simulated flood results of the initial PERSIANN-CCS QPEs in Table 3, the average value of Nash-Sutcliffe coefficient has a 14% increase, the correlation coefficient has a 14% increase, process relative error has a 8% decrease, peak flow relative error has a 18% decrease, the water balance coefficient has a 7% increase, and peak flow time error has 25 hours decrease, respectively (as shown in Table 3 and Table 4).So it implies the re-optimized parameters with the post-processed PERSIANN-CCS QPEs for the coupling model is necessary and effective, which makes a better performance for the coupling model in karst flood simulation and forecasting.

Peak flow time error analysis
It is very important to accurately determine the flood peak flow time in karst area, which could offer enough response times for evacuation safely and rapidly before the flood disaster appears.From the above results in Figure 11, 12 and Table 3, 4, it has been found that all flood simulations have significant peak flow time errors, and all of them are negative, means the simulated flood peaks appeared earlier than the observed values.Among them the average peak flow time error with the precipitation of rain gauge is -7 hours, and that is -32 hours with the precipitation of the initial PERSIANN-CCS QPEs.It is an obvious error and could not be ignored in flood forecasting.While the average peak flow time error of the coupling model

Conclusion
There is no reliable precipitation data of rain gauges in many karst river basins.How to obtain the reasonable rainfall data for the hydrological model in flood forecasting is especially Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License. the results of QPEs compared with the observed precipitation by rain gauge have some relative errors, which causes the karst flood simulation results by the coupling model (coupling QPEs with a PBDHM) have uncertainties that affect the model's performance largely.So the results of initial QPEs could not be used directly to build the coupling model.
.The spatial resolution of Liuxihe model for LKRB is 200m*200m.And the PERSIANN-CCS QPEs products that the spatial resolution is 0.04°*0.04°scale and time interval is 30 minutes are employed to estimate the precipitation results for LKRB.The resolution of the PERSIANN-CCS QPEs must be downscaled to the same size as Liuxihe model before building the coupling model.The PERSIANN-CCS QPEs products after post-processed could offer the high-precision precipitation results for LKRB where lack of enough rain gauges.It can largely improve the model performance by coupling the post-processed PERSIANN-CCS QPEs with Liuxihe model.A modified PSO algorithm resolution 0.04°*0.04°scale and time interval 30 minutes since 2000.The output of PERSIANN-CCS QPEs has been downscaled at 200m*200m as the same spatial resolution as Liuxihe model in LKRB.The hourly precipitation data of the PERSIANN-CCS QPEs are collected and compared with the precipitation observed by rain gauges.
is employed to optimize the model parameters in this study, which can make the model's performance much better in flood forecasting in karst river basins.The observed meteorological, hydrological data and the development conditions of the karst underground river are used to optimize the model parameters.The terrain property data like the DEM, land use type and soil type can be downloaded freely from an open access databases on the website.The model is validated by observed karst flood events.All these factors of the model are physically based and rational to truly reflect the underlying surface of the karst basin.So it implied Liuxihe model could be used for real-time flood forecasting in karst river basins.
Figure 1.Sketch map of the KHRU In Figure 1.b, the three-dimensional space model of the KHRU in Liujiang Karst River Basin(LKRB) is built in the laboratory to better understand how groundwater move in the karst media and convert mutually with the surface river.Then the hydrological model could be built more visualized through this way.
of the epikarst zone, T  is the simulation time-step, perc TT is the attenuation coefficient, epi SAT is the saturation water content of the epikarst zone, epi FC is field capacity of the epikarst zone, and epi K is the saturated hydraulic conductivity of the Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.epikarst zone.
quantity of the aquifer on the i day(mm/d), seep W is the water flux through the bottom of the soil profile into the underground aquifer on the i day(mm/d), gw  is the delay time of the supplies(day).
Figure 2. The DEM and three-dimensional topographical map of LKRB.
are 68 rain gauges and 131 grid points of PERSIANN-CCS QPEs within LKRB and five karst flood events from 2008 to 2013 has been collected respectively.There is a flood event each year.The karst floods process in LKRB have typical characteristics: the flood peak flows usually exceed 10,000m 3 /s and expression of the multi-peaks flood process.A flood process usually lasts about 10 days, and the shortest flood event duration is only about 3 days, the longest is 25 days.The hourly precipitation data of rain gauges are collected in this study to compare with the results of PERSIANN-CCS QPEs.The rain gauges, grid points of PERSIANN-CCS QPEs and the Liuzhou river gauge that closes to the outlet of LKRB are shown in Figure3.

Figure 3 .
Figure 3. Sketch map of Liujiang River Basin(LKRB) Catchment property data for distributed hydrological model mainly include DEM, land use and soil types.These data are downloaded from an open access databases.The DEM is downloaded from the shuttle radar topography mission database at http://srtm.csi.cgiar.org(Falorni et al., 2005, Sharma et al., 2014).The downloaded DEM has an initial spatial resolution of 90m*90m, and after many model resolution tests, the most appropriate resolution has been confirmed as 200m*200m for Liuxihe model in LKRB.So the spatial resolution of the initial DEM is rescaled to 200m*200m in this study, which is a high resolution for Liuxihe model in LKRB.The DEM is shown in Figure 2(a).The land use type is downloaded from http://www.isric.org.The initial spatial resolutions of the land use type and soil type are 1000m*1000m.Both of them need to be rescaled to 200m*200m in this study.Figure 4 (a) is land use types and (b) is soil types.

Figure 4 .
Figure 4.The property data for Liuxihe model in LKRB

Figure 5 .
Figure 5. Precipitation pattern comparison of two precipitation products(2008), (a) is the average precipitation of rain gauges, (b) is the average precipitation of PERSIANN-CCS QPEs.

Figure 6 .
Figure 6.Precipitation pattern comparison of two precipitation products(2009), (a) is the average precipitation of rain gauges, (b) is the average precipitation of PERSIANN-CCS QPEs.

Figure 7 .
Figure 7. Precipitation pattern comparison of two precipitation products(2011), (a) is the average precipitation of rain gauges, (b) is the average precipitation of PERSIANN-CCS QPEs.

Figure 8 .
Figure 8. Precipitation pattern comparison of two precipitation products(2012), (a) is the average precipitation of rain gauges, (b) is the average precipitation of PERSIANN-CCS QPEs.

Figure 9 .
Figure 9. Precipitation pattern comparison of two precipitation products(2013), (a) is the average precipitation of rain gauges, (b) is the average precipitation of PERSIANN-CCS QPEs.
precipitation results based on PERSIANN-CCS QPEs after revised will be as input data for Liuxihe model to test its feasibility through the floods simulation.From the above procedure of the post-processed PERSIANN-CCS QPEs, it could be found that the revise factor-2 / PERSIANN CCS PP  is a key to make the results of PERSIANN-CCS QPEs much closer to the observed precipitation by rain gauges, means the systematic errors of the PERSIANN-CCS QPEs could be corrected effectively.So the post-processed Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.
is employed to optimize the parameters of the coupling model in this study.And the flood process for parameter optimization is the Flood 2009060908.The results of parameters optimization are shown in Figure 10, among them, (a) is the objective function evolution result, (b) is the parameters evolution result, and (c) is the simulated flood process by using the optimized model parameters.

Figure 10 .
Figure 10.Parameter optimization results with the improved PSO algorithm From the results of Figure 10(c), it could be found that the coupling model with initial model parameter values does not simulate the observed karst flood process satisfactorily, and compared with that, the parameters optimization with the improved PSO algorithm could largely improve the coupling model's performance.The simulated flood process is very close to the observed value.In order to test the parameters optimization effect with different precipitation sources, both the precipitation of the rain gauge and PERSIANN CCS QPEs are used to optimize the parameters of the coupling model.To compare with that, the simulated flood process of the

Figure 11 .
Figure 11.The flood simulation results of the coupling model with two precipitation products From the result of Figure 11, it could be seen that the simulated karst flood discharges with the precipitation of rain gauge are the best.And the simulated values are the closest to the measured values, especially the simulated flood peaks are satisfactory, which is the most concerned factor in real-time flood forecasting.The average values of six evaluation indices, including the Nash-Sutcliffe coefficient (C), correlation coefficient (R), process relative error (P), peak flow relative error (E) ,water balance coefficient (W), and peak flow time error (T)

Figure 11 ,
Figure 11, it may be seen that the karst flood simulation results with the initial PERSIANN CCS QPEs are not so satisfactory, and the performance of the model are worse than that of the rain gauge precipitation.While the flood simulation results of the coupling model with the post-processed PERSIANN CCS QPEs are much better, also the evaluation indices of the flood simulation have been largely improved.
error has the biggest improvement.It is obvious that the evaluation indices are improved substantially with the post-processed QPEs.So it implies the post-processed method for PERSIANN-CCS QPEs in this paper is feasible and effective.And coupling the postprocessed PERSIANN-CCS QPEs with Liuxihe model has the potential to improve the model performance in flood simulation and forecasting in LKRB.Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.6.2 Effects comparison of different model parameters The performance of the coupling model makes a big difference with different parameters.There are two different sets of model parameter values in this study, one is the parameters with the precipitation of rain gauge, and the other is the parameters with the post-processed PERSIANN-CCS QPEs.The post-processed PERSIANN-CCS QPEs are used to re-optimize the coupling model parameters again to test its necessity.And the flood simulation results with two different sets of model parameters are shown in Figure 12.

Figure 12 .
Figure 12.Coupled flood simulation results with the same parameter as the rain gauge precipitation and re-optimized parameter with the post-processed PERSIANN-CCS QPEs Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-438Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 5 September 2018 c Author(s) 2018.CC BY 4.0 License.with the post-processed PERSIANN-CCS QPEs precipitation and re-optimized parameters is also -7 hours.It makes a great difference.It has been found that both the average peak flow time errors of Liuxihe model with the precipitation of rain gauge and the coupling model with the precipitation of the post-processed PERSIANN-CCS QPEs and re-optimized parameters are -7 hours (as shown in Table 4).So it implies the peak flow time error is -7 hours for the coupling model in LKRB, means the actual time of the flood peak may be 7 hours later, which is very important in flood forecasting and equivalent to a 7 hours long lead time for evacuation safely.There are two reasons for the peak flow time errors.One is the systematic error of the coupling model itself.And that could be reduced by improving the model structure and function as well as the reliable precipitation by PERSIANN-CCS QPEs and parameters optimization.The other is due to the karst development laws and the characteristics of karst water-bearing media, which can regulate the rainfall process during floods.The karst depressions and other karst negative landforms in the upstream regions can hold back and store some large floods.What is more, the karst fissures can also slow down the floods rate.These factors can play a crucial role in natural flood detention and peak clipping.So the response times of flood peak flow to rainfall increased, and the observed flood peak times lagged behind.In comparison, the simulated flood peak flows appear ahead of time.The rainfall process from the sky to the ground and finally converge to the outlet of the basin has passed through the surface karst zone, the karst conduit and fissure as well as the underground river.And the karst development laws and the characteristics of karst waterbearing media have obvious influence on the rainfall-runoff process during the whole hydrological process, which makes the response time of flood peak flow to rainfall increases, and the simulated flood peak flow by the coupling model appears earlier.It implies there is a lead time for evacuation safely in flood forecasting.The flood peak flow time has a very close relationship with the floods rate, and the floods rate is very important to determine the key factors of the karst conduit, the underground river and other hydrogeological parameters.The sensitive parameters in this paper such as the underground river parameters (as shown in

Tables 868 Table 1 .
Precipitation pattern comparison of two precipitation products