Research and application of parameter estimation method in hydrological model based on dual ensemble Kalman filter

In hydrological models, the parameter is the quantity describing the physical characteristics of the underlying surface, which is usually time invariant. With the development of hydrological science, some hydrological models have become more accurate in describing physical processes, but also more complex and contain more parameters (Zhao 1992; Arnold & Fohrer 2005). Most of the parameters are not easy to obtain directly through the basin measurement, thus, to obtain the accurate parameters, we need a long series of hydrological data, which require optimization to calibrate the parameters. For example, the impervious area ratio Im of the Xin'anjiang model can now be obtained by analyzing the land use of the watershed through remote sensing satellite images, but civil remote sensing satellites will not monitor the target watershed in real time, so the obtained data generally lags behind, resulting in the reduction of the accuracy of this parameter. What is more, the deep evapotranspiration coefficient parameter Ci is related to the proportion of deep-rooted plants in the watershed area. This parameter can only be obtained through artificial ground measurement, so it is a heavy workload to obtain the latest data for large watersheds. Therefore, a long series of rainfall–runoff data are needed to calibrate the parameters by using the optimization algorithm, so as to obtain more accurate parameters. The underlying surface condition of a basin usually changes with time, and there are some deviations in the data used for parameter calibration. These problems will be ignored when using fixed model parameters (Pathiraja et al. 2016). In addition, parameter calibration cannot be performed in ungauged basins. Usually, an interpolation algorithm (Wagner et al. 2012) or the parameter of a similar basin (Micovic & Quick 1999; Reichle & Walker 2002) is used to solve this problem. The above conditions may make the parameters deviate from the optimal value to a certain extent and reduce simulation accuracy. Therefore, it is necessary to apply the real-time parameter estimation method to the hydrological model to correct the parameters.

The convincing performance of data assimilation (DA) in real-time correction of hydrological models has attracted the attention of hydrologists. DA is a development process used to optimally combine model simulations and observations with appropriate modeling. Many DA methods have been applied in hydrology. Among them, the most commonly used methods are variational method, particle filters (PFs) and the Kalman filters (KFs). Compared with other methods, KFs are easier to implement and can produce even better results with lower computational requirements. In addition, KFs are very flexible in coupling with hydrological models and have more derived variants than any other method.

Kalman proposed the Kalman filter in 1958 (Kalman 1960), which is similar to the least-squares method in that it aims to minimize the error and only applies to linear models. In reality, hydrological models are usually non-linear. When the Kalman filter is applied to the non-linear model, filter divergence is often caused by error accumulation (Mirzaei & Roumeliotis 2007). For the non-linear model, Kalman proposed an extended Kalman filter (EKF), which has been successfully applied to the biological field (Svrcek et al. 1974). However, the EKF may still diverge under the condition of strong non-linearity. When Evensen applied EKF to the multilayer quasi-geostrophic model, he found the reason for the divergence of the EKF, and proposed the ensemble Kalman filter (EnKF), which eliminated the divergence phenomenon caused by the infinite error growth in the EKF due to an overly simplified closure in the error covariance equation (Evensen 1994). Later studies also reported EnKF was better than the EKF in performance (Reichle & Walker 2002). Burgers et al. (1998) improved EnKF and proposed that one should add random perturbations with the correct statistics to the observations and generate an ensemble of observations that is then used to update the ensemble of model states. In the following ten years, there were still many researchers trying to improve EnKF (Anderson 2001; Bishop et al. 2001; Whitaker & Hamill 2002; Tippett et al. 2003; Vrugt et al. 2005; Clark et al. 2008). The earliest EnKF is applied to state estimation (Houtekamer & Mitchell 1998, 2001). This method can not only estimate state variables but also use observations to estimate the value of imprecisely known model parameters (Anderson 2001).

For the first time, Moradkhani et al. (2004) applied EnKF in the field of hydrology, combined the parameter ensemble and variable ensemble to form a dual-parameter-state ensemble, used observations to estimate model parameters, and proved its applicability and practicability in the prediction. EnKF has been widely used in the field of hydrology and water resources. There are three ways to use EnKF: (1) assimilate observations to estimate the state (Reichle & Walker 2002; Komma et al. 2008; Xie & Zhang 2008; Fu et al. 2014; Lu et al. 2016); (2) assimilate observations to estimate the parameters (Pathiraja et al. 2016; Lei et al. 2019); (3) assimilate observations to estimate the state and parameters (Vrugt et al. 2005; Lu et al. 2013; Xie & Zhang 2013; Liu et al. 2017). Previous studies considered the parameters as a whole to study their characteristics in the estimation process or to study the effect of parameters corrected by the EnKF on simulation results. There are few studies that focus on the effects of selecting different parameters for estimation and setting different initial ensemble distributions on the estimation process.

In this study, according to the sensitivity of parameters and the function of parameters in the hydrological model, five representative parameters of the Xin'anjiang model are selected as the research object. This study is divided into two parts. The first part is numerical simulation, which mainly studies the characteristics of EnKF when different parameters are selected for estimation from two aspects: the initial distribution of parameters' ensemble and the function of parameters in the model. According to these rules, some suggestions are put forward to improve the efficiency of EnKF in the hydrological model. The second part is the application, in which EnKF is applied to a new basin: Jinjiang River Basin, China, and it is verified if the EnKF method can solve the problem of prediction accuracy decline caused by the change of underlying surface, the lack of data and other reasons, and the correctness of some suggestions put forward in the first part is further verified.

As mentioned above, this study selects five representative Xin'anjiang model parameters to study the applicability and practicability of the ensemble Kalman filter (EnKF) in hydrological models.

The Xin'anjiang model, which was originally used in flood prediction of Xin'anjiang hydropower station in China, was developed by Zhao and was officially published in 1980 (Zhao 1992). After decades of improvement, it has become one of the most influential hydrological models in the world, and has been widely used in streamflow simulation, flood prediction, and other aspects worldwide (Jayawardena et al. 2006; Yang et al. 2011; Rahman & Lu 2015). It is a conceptually based model based on the conception of hillslope hydrology. In this paper, the most commonly used improved form of the Xin'anjiang model, the three-source Xin'anjiang model, is used for research. It divides the whole watershed into many sub-watershed units. The parameters of the model have clear physical meaning, and the parameters are basically independent of each other. Compared with the past lumped hydrological models such as the two-source Xin'anjiang model, hymod model and SCS model, the three-source Xin'anjiang model is more detailed and more in line with the real situation in terms of evaporation, soil water, runoff division, and slope confluence. The infiltration module of the model calculates the total runoff generated by rainfall according to the concept of excess storage runoff. Water in the soil is divided into tension water (water below the field capacity) and free water (water above the field capacity), and their regulation and storage effects are considered, respectively. In the evaporation calculation, the model divides the tension water into upper layer, middle layer, and lower layer to calculate the evapotranspiration, respectively. In this model, the total runoff is divided into saturated surface runoff, soil water runoff, and groundwater runoff by using a simplified free water storage reservoir. In terms of routing calculation, the unit hydrograph method is generally used for surface runoff, and the linear reservoir method is used for routing of soil flow and underground flow. The inputs of the model are precipitation and potential evaporation. The precipitation is directly obtained from the hydrological station, and the potential evaporation is calculated from the meteorological data.

As there are many parameters in the Xin'anjiang model, the dynamically dimensioned search (DDS) algorithm, which is suitable to deal with the problem of parameter optimization in large calculations, is selected to calibrate the model parameters in the application. The objective function is Nash–Sutcliffe efficiency coefficient (NSE). The DDS algorithm was proposed by Tolson & Shoemaker (2007). This algorithm does not need to adjust the algorithm parameters because it has fast convergence speed and easily avoids local optimization. When dealing with the parameter optimization problem in large calculations, the performance of this algorithm is better than the traditional global optimization algorithms such as SCE. The DDS algorithm and its improved algorithm are widely used in model parameter calibration with more undetermined parameters (Huang et al. 2014; Yen et al. 2016).

The global sensitivity of the model parameters was analyzed using the LH-OAT method. The accuracy of streamflow simulation is evaluated by NSE, Kling–Gupta efficiency (KGE) and bias. In the numerical simulation experiment, the correlation coefficient is selected as the correlation index between the parameter ensemble and streamflow ensemble simulations at a certain time step. Some algorithms and evaluation indexes used in the study will be described in detail below.

The LH-OAT sensitivity analysis method is a global sensitivity analysis method combining the Latin hypercube sampling (LH) and the one-factor-at-a-time algorithm (OAT) proposed by Griensven et al. (2006).

The sensitivity analysis steps are as follows:

• (1)

  Using LH sampling, the space is divided into n layers, and each layer is randomly sampled once to generate n LH sampling sets (including m parameters).

• (2)

  According to OAT, each parameter is changed slightly, so that each sampling set can obtain m group parameters.

• (3)

  n×(1+m) group parameters are substituted into the model for calculation, and an objective function is selected to evaluate the calculation results.

• (4)
  Evaluate the sensitivity of each parameter according to the following equation:

  

  (1)

  where M [] is the objective function; f_i is the minor change in parameter, which is directly proportional to the range of parameter values; e_i is the ith parameter, i[1, p]; p is the total number of parameters; RS_ij is the relative sensitivity of parameters considering the range of parameter value; and j is a LH sampling set, j[1, n].

The Kalman filter is an unbiased estimation based on minimizing the variance of the analysis error under the assumption that the noise is Gaussian white noise in a linear system. The estimation process is as follows:

The noise with a Gaussian distribution is called Gaussian noise, and its power spectral density is a uniform distribution, which is called Gaussian white noise. In the ensemble Kalman filter, the initial values of the parameter and state ensembles are Gaussian white noise sequences. In this study, the approximate Gaussian white noise sequence is used as the initial value of each parameter ensemble. The generation steps are as follows:

• (1)
  Generate uniformly distributed random numbers between 0 and 1 according to the following equations:

  

  (13)

  and

  

  (14)

  where rnd_i is the random number between 0 and 1, mod() is the rounding function, m =2¹⁶, r_i is the random seed, and r₀ can be any value less than m.
• (2)
  Generate the Gaussian white noise sequence according to the following equation:

  

  (15)

  where n is sufficiently large, the approximation of n=12 is quite good; μ is the mean of noise, σ is the variance of noise, and y is the element of noise sequence.

The Jinjiang River is a tributary of Ganjiang River, a first-order tributary of the Yangtze River in China. It flows into the left bank of Ganjiang River, with a length of 294 km, a drainage area of 7,650 km², and an average annual flow of 222 m³/s. The forest coverage of the Jinjiang River Basin is over 70%. The rainfall is rich, with an average annual precipitation of 1,400–1,700 mm. The rainstorms are mainly from May to July. In this paper, the main study area is the drainage basin of Gaoan hydrological station, which is located on Jinjiang River, with an area of 6,215 km² (Figure 1). The elevation of the catchment is higher in the northwest, where most tributaries of the Jinjiang River originate, ranging from 18 to 1,096 m (Yin et al. 2018). There is no large reservoir in the basin, and it is less disturbed by human activities than other basins. It is a representative natural basin, and it can obtain better streamflow simulation results without a coupling reservoir operation model or other models; therefore, it is suitable to study the adaptability and practicability of EnKF as the hydrological model itself.

The Danjiang River is a tributary of Hanjiang River in the north of the middle reaches of the Yangtze River, with a drainage area of 16,812 km² and an average annual precipitation of 400–1,100 mm. There are dense vegetation and primitive forests in the basin. Another study area of this paper is the drainage basin of Jingziguan hydrological station, which is located on the Danjiang River, with an area of 7,060 km² (Figure 1). The elevation of the basin is 163–1,970 m, and the river flows from the northwest to the southeast.

In this study, the study area of the numerical simulation part is Jinjiang River Basin; the main study area of the application part is Jinjiang River Basin, and the Danjiang River Basin is used to provide the initial distribution mean of parameters ensemble in the case of lack of data in Jinjiang River Basin.

The daily precipitation observation data of 20 precipitation stations from 2001 to 2009 and the hydrological station daily streamflow observation data from 2001 to 2009 were used to simulate the streamflow of the Gaoan hydrological station. The daily precipitation observation data of 14 precipitation stations from 2001 to 2009 and the daily streamflow observation data of the Jingziguan hydrological station from 2001 to 2009 were used to simulate the streamflow of the Jingziguan hydrological station. The Thiessen polygon method (Yen et al. 2016) was used to interpolate the precipitation data into the sub-basin areas. Other data including temperature, wind speed, relative humidity, and solar radiation were obtained from the observation data of three weather stations in or around the basin from 2001 to 2009 provided by China Meteorological Bureau. The DEM (http://www.resdc.cn/data.aspx?DataID=123), vegetation index grid data (http://www.resdc.cn/data.aspx?DataID=257), and spatial distribution grid data of three soil textures (http://www.resdc.cn/data.aspx?DataID=260) were provided by the resource and environmental science data center of the Chinese Academy of Sciences, with a resolution of 1 km×1 km (Figure 2). Flow length grid data and slope grid data were obtained by hydrological analysis of the DEM.

In this section, the drainage basin of Gaoan hydrological station is selected as the study area. In numerical simulation, the randomly generated parameters are used as the true values of the Xin'anjiang model parameters. These true values will be used in the model to generate an ideal observed runoff. In the numerical simulation test, the true values of some parameters are set as unknowns, and the ideal observed runoff is input into the ensemble Kalman filter algorithm to test whether the algorithm can accurately find the true values of these parameters and, at the same time, evaluate the performance of the algorithm when using different parameters as unknowns. The two-year daily streamflow data generated by the observation precipitation data of the Gaoan station drainage basin from 2008 to 2009 are selected for the observation streamflow for the numerical simulation experiment. Table 1 shows the true value and range of the model parameters.

When there are many undetermined parameters in the model, no matter the optimization algorithm or data assimilation algorithm used to optimize the parameters of the model, to ensure the efficiency of the algorithm, all parameters are not input into the algorithm for calculation at the same time, but the parameters that have a greater impact on the simulation results are selected for calculation. The global sensitivity of the parameters is an important index to evaluate the effect of the parameters on the simulation results. In this study, the LH-OAT global sensitivity analysis method is used, and the range of each parameter is divided into 200 layers. Sensitivity analysis is performed for 13 parameters of the Xin'anjiang model, and the sensitivity analysis results are shown in Figure 3. According to the global sensitivity and the role of parameters in the model, this study selects five parameters I_M, PETM, B, C_i, and K_i for numerical simulation. These parameters directly affect the main processes of the Xin'anjiang model. The parameter Im is a parameter in the process of saturation excess runoff, which determines how much water will penetrate into the soil, thus affecting the surface runoff, soil runoff, and underground runoff respectively; parameter B is also a parameter in the process of saturation excess runoff, which reflects the heterogeneity of soil water storage capacity in the basin. Runoff occurs first where the soil water capacity is low. Therefore, this parameter affects the total amount of the runoff and the time of runoff generation in the basin; parameter PETM is a parameter of evapotranspiration process, which directly affects the evaporation of soil water and the total amount of the runoff in the basin; parameter Ki is the parameter of water distribution process, which determines the proportion of free water outflow in soil to soil flow in a certain period of time, and directly affects the total discharge of soil flow; the parameter CI is a parameter of the routing process, which determines the routing time of the flow in soil.

This section analyzes three factors that affect the time required for the parameter ensemble to stabilize under the condition of single-parameter estimation. They are the mean of Gaussian white noise used to generate the initial parameter ensemble (hereinafter referred to as the mean initial distribution value), the variance in Gaussian white noise used to generate the initial parameter set (hereinafter referred to as the initial distribution variance), and the roles of parameters in the model.

We change the initial distribution means of I_M, PETM, B, C_i, and K_i and initial distribution variance is fixed as (bl2-bl1)×0.1, where bl1 is the lower limit of parameter value and bl2 is the upper limit of parameter value. Each parameter is estimated separately, and other parameters are set as the true value.

The ensemble number is set to 100, the standard deviation of the model error is set to 0.1, and the standard deviation of the observation error is set to 0.1. Parameter updates are performed at every time step except for special instructions. It is set that when the elements in the ensemble at a certain time reach the following conditions, the ensemble is considered to reach stability:

The difference between the maximum and the minimum in the ensemble is less than (bl2-bl1)×0.01, and no longer greater than (bl2-bl1)×0.01.

The difference between the ensemble mean at a certain time and the ensemble mean at the end of the estimation process is less than (bl2-bl1)×0.01 and no longer greater than (bl2-bl1)×0.01.

The difference between the parameter ensemble mean and the parameter true value is less than (bl2-bl1)×0.01.

According to Figure 4(a)–4(e), when there is only one parameter to be estimated, the closer the initial distribution mean is to the true parameter value, the shorter the time step required for the parameter ensemble to stabilize. For the parameters of PETM, C_i, and K_i, the true values coincide with the initial distribution mean, which needs the shortest time step to stabilize, but there are some exceptions. First, it can be seen from Figure 4 that when the initial distribution mean is close to the upper and lower limits, the time to stabilize does not conform to the above conclusion. This is because there are elements exceeding the upper or lower limits specified by the model in the initial parameter ensemble; therefore, it is necessary to correct these elements to return to the range, and that affects the estimation process, which is also the reason that the time step required for I_M to stabilize when the initial distribution mean is equal to the true parameter value is not the shortest. Second, it can be seen that the time step required for B to stabilize when the initial distribution mean is equal to the true value is not the shortest. This is because parameter B only affects the streamflow when the precipitation is not equal to 0. When the precipitation is 0, only the model error and observation error effect the value of in Equation (4), which makes the parameter update in a random direction and changes the time step required for B to stabilize. To further verify this conclusion, on the basis of the original solution, parameter B is updated only when precipitation is not equal to 0. The calculation results are shown in Figure 4(f). It can be seen that the time required for the parameter ensemble to stabilize is basically the same as when the parameter is updated at all times, and except for the part where the initial distribution mean is close to the upper and lower limits, all the experimental results meet the conclusion that the closer the initial distribution mean is to the true parameter value, the shorter the time step required for the parameter ensemble to stabilize.

We change the initial distribution variance of I_M, PETM, B, C_i, and K_i and the initial distribution mean is fixed as (bl2+bl1)/2. Each parameter is estimated separately, and other parameters are set as the true value. Other settings are the same as the previous section.

It can be seen from Figure 5 that when the initial distribution variance is less than the blue line, with the increase in initial distribution variance, the time step required for different parameter ensembles to stabilize generally demonstrates the three trends of increasing, decreasing, and increasing first and then decreasing.

is the matrix of Kalman gain of the m-th parameter at time t, is the vector of the value of the m-th parameter at time t in, and is the mean of ensemble of the m-th parameter value at time t.

When the initial distribution variance increases, the uncertainty of the initial ensemble increases, and both and increase, and according to Equation (21), the Kalman gain will also increase. According to Equation (11), larger Kalman gain makes the algorithm believe less in the old predicted value, making it easier for the parameters to approach the true value.

In Figure 5(c), it can be seen that the time step required to stabilize parameter B increases with the increase in initial distribution variance because the true value of B is 0.249109, and the initial distribution mean of B is set to 0.25, which is very close to the true value. Low uncertainty of the parameter ensemble is required for rapid stabilization; therefore, increasing the initial distribution variance makes it more difficult for the parameter ensemble to stabilize. In Figure 5(a), it can be seen that the time required for I_M to stabilize decreases with the increase in initial distribution variance. The true value of I_M is 0.0177124, which is quite different from the mean of initial ensemble. Therefore, to make the parameter ensemble stabilize in a short time, it is more important to shorten the time of the parameter ensemble to close to the true value than to reduce the uncertainty, thus increasing the variance of the initial distribution will reduce the stabilization time step. In Figure 5(b) and 5(d), the true values of PETM and Ci are neither around the mean of the initial distribution nor quite different. When the initial distribution variance is too small, because it is difficult for the parameter ensemble to approach the real value, the time step required to stabilize is long. When the initial distribution variance is too large, the uncertainty of the ensemble is still high, and the time step required to stabilize is also long. When the initial distribution variance is moderate, the parameter ensemble can not only approach the true value rapidly, but also reduce the uncertainty quickly, so PETM and Ci can stabilize after a short time step. To further verify the above conclusion, the parameter PETM is estimated with the initial distribution variance of 0.002, 0.02, and 0.04. Figure 6 shows the estimation process of the first 200 time steps of PETM. In these three cases, after 61, 29, and 22 time steps, the difference between the ensemble mean and the true value is less than (bl2-bl1)×0.02. Based on the above time steps, after 69, 82, and 103 time steps, the parameter ensemble stabilizes, and the total time steps to stabilize are 130, 111, and 125, respectively.

In Figure 5(e), the position of the true value of parameter Ki is similar to that of PETM and C_i, but it generally decreases with the increase in initial distribution variance, because the condition K_i + K_g < 1 of the Xin'anjiang model affects the process of parameter estimation. Although it may not be reasonable to do so in the physical model, to further verify the above conclusion, the condition of K_i +K_g <1 is ignored in the Xin'anjiang model, and then the same calculation is performed for K_i. It is found that the time step required to stabilize decreases first and then increases with the increase in initial distribution variance (Figure 5(f)), which conforms to the above conclusion.

In Figure 5, when the initial distribution variance is greater than the blue line, the ensemble elements exceeding the upper and lower limits will be corrected back to the parameters allowed by the model, which means that the uncertainty can no longer be increased, so the estimation process is basically similar, and it will stabilize at a similar time step.

In conclusion, the most favorable initial distribution variance for the parameter ensemble stabilization depends on the difference between the ensemble mean and the true value of the parameter. When the initial distribution mean is close to the true value, the small variance in the initial distribution is conducive to making the ensemble stabilize. When the initial distribution mean is far from the true value, the large initial distribution variance is conducive to making the ensemble stabilize. When the initial distribution mean is not close to or far from the true value, the moderate variance in the initial distribution is conducive to making the ensemble stabilize. However, in practice, we cannot know the true value of the parameter, so it is difficult to improve assimilation efficiency by determining an initial distribution variance.

To improve simulation accuracy, it is preferred to use high sensitivity parameters for assimilation, although they are not all suitable. It can be seen from Figures 4 and 5 that in all cases, the time step required for high sensitivity parameter I_M to stabilize is less than 30 steps. I_M is the parameter with the shortest stabilization time step. However, the time step required for the high sensitivity parameter PETM to stabilize is similar to the low sensitivity parameter K_i, which is longer than that for medium sensitivity parameters B and C_i. To analyze the causes of the above phenomena, the high sensitivity parameters I_M and PETM are selected, the initial distribution mean value is fixed as (bl2+bl1)/2, the initial distribution variance is fixed as (bl2-bl1)×0.01, each parameter is estimated separately, and other parameters are set as the true value. Other settings are the same as the previous section. The results are shown in Figure 7.

According to Equation (21), EnKF updates the parameters according to the covariance matrix between the parameter and streamflow simulation ensembles at time t. Covariance is an index to evaluate the correlation between two ensembles. The higher the correlation between parameters and streamflow, the more accurate the result of parameter estimation is, and the more easily it stabilizes. It can be seen from Figure 7(d) that in the estimation process of PETM, the correlation coefficient is close to 1 at some time steps, indicating that the parameter is almost linearly related to the streamflow, but the range and mean of the parameter ensemble are almost unchanged. The reason is that PETM at time t not only affects the streamflow at time t, but also affects the streamflow at time t+1, t+2…., which is very common in hydrological models. However, EnKF does not consider such a situation, resulting in the innovation that in Equation (5) may be 0. For example, at some time steps, some elements in the ensemble are not real values; however, because the variables such as capacity of soil water are not reasonable due to the effect of parameters of non-true values at other time steps, the error of streamflow is 0, so the innovation is 0, and the parameter update is not performed. In addition, the effect of parameters of other time steps may reduce the correlation between the parameter and the simulation of streamflow, making it more difficult for the parameter ensemble to stabilize. On the contrary, I_M only affects the streamflow at the current time, so it can estimate the parameter more accurately when the correlation coefficient is close to 1 and can stabilize rapidly. It should be noted that I_M could only be updated correctly when the precipitation was not 0. Although the precipitation is 0 at the time steps of 1–9, the reason why the parameter ensemble changes at the time steps of 1–9 in Figure 7(a) is that there are model observation errors causing the parameter ensemble to update in a random direction.

In conclusion, it is necessary to analyze not only the sensitivity of parameters, but also the functions of parameters in the model, and then select the suitable parameters for estimation.

In this section, we analyze the time step required for the parameter ensemble to stabilize under different initial distribution means and different initial distribution variances in the case of two-parameter estimation, three-parameter estimation, and four-parameter estimation. In additon, we discuss the different points of multi-parameter estimation and single-parameter estimation and the interaction among parameters during multi-parameter estimation. We set the following solutions:

• A(1): Select parameters I_M and C_i and change the mean of their initial distributions together. The variance of their initial distributions is fixed to (bl2-bl1)×0.1. The two parameters are estimated together, and the other parameters are set to the true value.

• A(2): On the basis of A(1), parameter B is added. The initial distribution mean of B is fixed to (bl2+bl1)/2, and the initial distribution variance is fixed to (bl2-bl1)×0.1. These three parameters are estimated together, and the other parameters are set as the true values.

• A(3): On the basis of A(1), parameters B and K_i are added. The initial distribution mean of B and K_i is fixed to (bl2+bl1)/2, and the initial distribution variance is fixed to (bl2-bl1)×0.1. These four parameters are estimated together, and the other parameters are set to the true value.

• B(1): Select parameters C_i and PETM and change their variance of initial distribution together. The initial distribution mean is fixed to (bl2+bl1)/2. The two parameters are estimated together, and the other parameters are set to the true value.

• B(2): On the basis of B(1), parameter B is added. The initial distribution mean of B is fixed to (bl2+bl1)/2, and the initial distribution variance is fixed to (bl2-bl1)×0.1. The three parameters are estimated together, and the other parameters are set as the true values.

• B(3): On the basis of A(1), parameters B and I_M are added. The initial distribution mean of B and I_M is fixed to (bl2+bl1)/2, and the initial distribution variance is fixed to (bl2-bl1)×0.1. The four parameters are estimated together, and the other parameters are set to the true value.

Other settings are the same as the previous section.

In Figure 8(a) and 8(b), by adding the parameters to be estimated, the overall parameters still tend to take the shortest time to stabilize when the initial distribution mean value is closer to the true value, although this feature becomes less obvious. In Figure 8(c), only C_i and PETM are estimated. As in single-parameter estimation, the ensemble of C_i has the feature that the time required to stabilize first decreases and then increases with the increase in initial distribution variance. When the variance is approximately 0.02, the time step required to stabilize is the shortest, but by adding estimation parameters, this characteristic becomes less obvious. PETM in Figure 8(d) has a similar situation. Equation (21) is a calculation of the Kalman gain established in the cases of single-parameter and multi-parameter estimations, which indicates that the calculation of Kalman gain of each parameter will not consider whether other parameters are reasonable in the case of multi-parameter estimation, thus, it may cause mutual interference between parameters. Comparing Figures 4–8, it can be seen that the mutual influence of these parameters may be favorable or unfavorable to the parameter estimation, but it is unfavorable in most cases.

It can be seen from Figure 8(a) that adding other parameters together with I_M for multi-parameter estimation has little effect on the time required for parameter I_M to stabilize. As analyzed in the previous section, I_M only affects the streamflow of current time step and I_M is highly sensitive. This parameter has a high correlation with the streamflow when the precipitation is not zero, and the time step required to stabilize is very short, thus, adding other parameters will hardly affect the estimation process of I_M. C_i in Figure 8(b) is moderately sensitive, which directly affects the interflow and thus affects the streamflow of other time steps. After adding B, which is more sensitive than C_i and stabilizes more easily in single-parameter estimation, the time step required for C_i to stabilize increases slightly, while after adding K_i, which is less sensitive than C_i and does not stabilize easily in single-parameter estimation, the time required for C_i to stabilize remains almost unchanged. In conclusion, the correlation between streamflow and the parameters that required shorter time steps to stabilize in the case of single-parameter estimation is not easily affected in the case of multi-parameter estimation. Therefore, this parameter ensemble may stabilize first in the case of multi-parameter estimation, while the time step required for other parameters to stabilize may increase. However, it may be easier to disrupt the correlation between other parameters and streamflow after adding the parameters which are highly sensitive and do not stabilize easily, thereby significantly reducing the time step required for other parameters to stabilize. However, this is not absolute, because updating EnKF occurs locally, and the high global sensitivity of parameters does not mean that the local parameter sensitivity is also high, although it can represent a general trend. In Figure 8(c), the time required for C_i to stabilize after adding PETM and B on the basis of parameter C_i increases significantly, and in Figure 8(d), the time required for PETM to stabilize after adding C_i and B on the basis of parameter PETM conforms to the above conclusions. In addition, it can be seen from Figure 8(c) and 8(d) that the time required for C_i and PETM to stabilize after adding I_M is not obvious, because although I_M, as a highly sensitive parameter, more easily disrupts the correlation between other parameters and streamflow, the parameter itself can stabilize after a short time step and will not affect the estimation process of other parameters.

In conclusion, to improve the simulation accuracy of the hydrological model and ensure the performance of the ensemble Kalman filter, it is suggested that: (1) high sensitivity parameters that required shorter times to stabilize in the case of single-parameter estimation should be selected as the estimation parameters in the case of multi-parameter estimation; (2) low sensitivity parameters have little effect on improving the simulation accuracy and are difficult to stabilize, so it is not necessary to select them for estimation; (3) parameters with high sensitivity that need a long time step to stabilize due to various reasons should not be estimated. Although such parameters can still stabilize in numerical simulation, they may not in application.

In this section, the main study area is Jinjiang River Basin and two cases are designed to evaluate the applicability and practicability of the ensemble Kalman filter, and further verify some conclusions drawn in numerical simulation: (A) ‘there is a long series of hydrometeorological data in the basin, but due to the change of underlying surface or other reasons, the optimization algorithm may not get the optimal parameters of the forecast period’ and (B) ‘there is a lack of hydrometeorological data in the basin’. In (A), the result of parameter calibration by optimization algorithm is used as the assimilation basis, and in (B), the results of parameter calibration by optimization algorithm in a similar basin, Danjiang River Basin, are used as the initial distribution mean of parameter ensemble in assimilation. In this study, the similarity of the two basins is evaluated in terms of topography, vegetation, and soil. Each index is shown in Table 2. The mean of each pixel of grid data is taken for slope, vegetation index, and soil texture content. The two basins are located in a similar latitudinal zone, experience monsoon influence, have a humid climate and vegetation index of 0.8 indicating dense vegetation coverage (http://www.resdc.cn/data.aspx?DataID=257), and have similar area, max flow length, and slope. The soil texture is dominated by sand content in the two basins, and the silt content is basically the same; however, the soil viscosity in the Gaoan Basin is higher. The analysis shows that the underlying surface of the two basins is generally similar, and the parameters can be transferred.

The solution is as follows:

• A(1): The daily streamflow data of the Gaoan hydrological station from 2001 to 2007 are used to calibrate the parameters by the DDS algorithm, and the parameters obtained from the calibration are used to simulate the station streamflow from 2008 to 2009. The calibration period is from January 2008 to June 2008, and the validation period is from July 2008 to December 2009. Only the simulation results of the validation period are evaluated.

• A(2): The mean of the initial distribution of I_M, PETM, B, C_i, and K_i is set as the result of the calibration, and the variance in initial distribution is fixed as (bl2-bl1) × 0.1. Each parameter is estimated separately, and other parameters are set as the calibration result. The calibration period is from January 2008 to June 2008, and the validation period is from July 2008 to December 2009. Only the simulation results of the validation period are evaluated.

The ensemble number is set to 100, the standard deviation of the model error is set to one-tenth of the mean of the ensemble of simulation streamflow, and the standard deviation of the observation error is set to one-tenth of the observation. Parameter updates are performed at every time step. When the elements in the ensemble at a certain time reach the following conditions, the ensemble is considered to stabilize:

• (1)

  The difference between the maximum and the minimum in the ensemble is less than (bl2-bl1) × 0.05, and no longer greater than (bl2-bl1) × 0.05.

• (2)

  The difference between the ensemble mean at a certain time and the ensemble mean at the end of the estimation process is less than (bl2-bl1) × 0.05, and no longer greater than (bl2-bl1) × 0.05.

Note: since the true value of the parameter cannot be known in practice, the condition that the difference between the ensemble parameter mean and the true parameter value is less than (bl2-bl1) × 0.01 is ignored.

• A(3): I_M, B, C_i, and K_i are estimated together. Other settings are the same as in A(2).

• B(1): The daily streamflow data of the Jingziguan hydrological station from 2001 to 2007 are used to calibrate the parameters by DDS algorithm. The Nash of the calibration period is 0.839. The parameters obtained from the calibration are used for the drainage basin of the Gaoan hydrological station to simulate the streamflow from 2008 to 2009. Other settings are the same as in A(1).

• B(2): Set the mean of the initial distribution of I_M, PETM, B, C_i, and K_i as the result of calibration of the Jingziguan hydrological station, and the variance in initial distribution is fixed as (bl2-bl1) × 0.1. Each parameter is estimated separately, and other parameters are set as a result of calibration of the Jingziguan hydrological station. Other settings are the same as A(2).

• B(3): I_M, B, C_i, and K_i are estimated together. Other settings are the same as B(2).

In Table 3, in A(2), except for PETM proposed in a previous section, which is not suitable for assimilation and K_i, which is a low sensitivity parameter, the other three parameters can stabilize after assimilation and compared with A(1), the simulation is optimized (as shown in Figures S1–S5 in the Supplementary material, A(1) is the blue line in all the figures). In A(3), due to the interaction of parameters in multi-parameter estimation, some parameters are not reasonable when they stabilize, so although NSE and KGE increase (as shown in Figure S6 in the Supplementary material and bias in Table 3), the absolute value of bias also increases significantly. In solution A, the results of streamflow simulation after assimilation are generally better than that after parameter calibration by the DDS algorithm, which shows that the parameters obtained in the calibration period due to the underlying surface change or other reason are not fully suitable for the validation period, while EnKF can effectively adjust the parameters and improve the streamflow simulation accuracy.

In B(2), except for the solution of estimating C_i, the simulation of the solution of estimating other parameters is not much better than that of B(1), and some indexes become worse instead (as shown in Figures S7–S11 in the Supplementary material, B(1) is the blue line in all the figures). This is because solution A has performed parameter calibration, and the values of other parameters are reasonable except for those being estimated, thus, the estimation parameters more easily approach the true value, but solution B uses parameters of similar basins, some of which may be inconsistent with the underlying basin surface, so it is difficult to find the true value of estimation parameters. In B(3), four parameters are estimated together, which reduces the parameters that may not be consistent with the underlying basin surface, and the two indexes are significantly improved (as shown in Figure S12 in the Supplementary material).

In conclusion, in the basin where long data series can be used for parameter calibration, it is suggested to select a small number of parameters with high sensitivity that are suitable for assimilation as the estimation parameters to avoid interference among parameters in the assimilation. In the ungauged basin, it is suggested to use the parameters of a similar basin and select as many parameters suitable for assimilation as the number to be estimated to reduce the parameters inconsistent with the underlying basin surface.

In this study, the five representative parameters of I_M, PETM, B, C_i, and K_i of the Xin'anjiang model are selected as the research objects to analyze the characteristics of parameter estimation using the EnKF in numerical simulation and application.

After the numerical simulation analysis, the following conclusions are obtained: (1) the closer the mean of the initial parameter ensemble distribution is to the true value, the easier it is to stabilize. Generally, the parameter after calibration is closer to the true value than the default value. Therefore, for basins with long series of data, the method of parameter calibration to determine the initial distribution mean can improve the assimilation efficiency. (2) The most favorable value of the variance in the initial parameter ensemble distribution for the parameter ensemble to stabilize depends on the distance between the ensemble mean value and the true parameter value. In practice, it is almost impossible for us to know the true value of parameters, so it is difficult to improve assimilation efficiency by determining an initial distribution variance. (3) The parameters, which are highly sensitive like PETM, but are difficult to stabilize because the parameters at a certain time will affect the streamflow simulation at other times and greatly influence the correlation between other parameters and streamflow, are not suitable for assimilation. This conclusion is further verified in the application part. (4) Similar to I_M, it has high sensitivity and needs a short time step to stabilize in the case of single-parameter estimation, which is not easily affected by other parameters in the case of multi-parameter estimation. (5) Some conclusions obtained in the case of single-parameter estimation may not be applicable in the case of multi-parameters due to the interaction between parameters.

The application of EnKF in the Jinjiang River Basin shows that it can solve the problem of streamflow simulation accuracy degradation caused by the change in underlying surface, the lack of hydrometeorological data, and other reasons. In the basin where long series of data can be used for parameter calibration, it is suggested to select a small number of parameters with high sensitivity and assimilation suitability as the parameters to be estimated. In the ungauged basin, it is suggested to use the parameters of a similar basin and select as many parameters suitable for assimilation as the parameters to be estimated. The study area selected in this study is close to the natural state, and the underlying surface is minimally affected by human activities. The better application of EnKF in the basin does not show that the method is still applicable to the basin with large-scale reservoir and large-scale water diversion facilities. The applicability study of such a basin may be a future topic of investigation.

This paper was jointly supported by National Key R&D Program of China (2017YFC0406004).

The first two authors contributed equally to the work.

All relevant data are available from an online repository or repositories.