Abstract
Event-based hydrological models are extensively adopted for the estimation of design floods and in operational flood forecasting frameworks. However, an accurate estimation of the initial hydrologic condition (IHC) is essential in enhancing the predictive capability of an event-based hydrological model. Hence, in this study, IHCs of an event-based conceptual model are estimated using two different methods: (1) assimilation of observed variables such as streamflow and soil moisture using an ensemble Kalman filter and (2) states obtained from the continuous model calibrated using four different calibration metrics. The observed flood events at the Jagdalpur catchment are simulated using a conceptual hydrologic model setup at two spatial resolutions (lumped and semi-distributed). The results of the study demonstrate that IHCs estimated by the continuous models perform better than those obtained through data assimilation. The performance of semi-distributed event-based models was found to be outperforming their lumped counterparts demonstrating the advantage of increased model resolution. The states obtained from the continuous models calibrated using Nash–Sutcliffe Efficiency (NSE) are performing well in initialising the event-based models. The median efficiency of the semi-distributed event-based model (based on states from the NSE calibrated continuous model) is 0.91 and 0.77 during calibration and validation periods, respectively.
HIGHLIGHTS
Methods to estimate the initial hydrologic condition (IHC) to initialise event-based models were evaluated.
Streamflow assimilation in both lumped and semi-distributed models led to improved simulations.
Soil moisture assimilation yielded slightly better predictions in the semi-distributed model.
Semi-distributed event-based model, initialised by IHCs extracted from the corresponding continuous model, is outperforming other models.
INTRODUCTION
Flood is one of the most destructive natural disasters around the world that can cause catastrophic impacts on day-to-day operations (Alfieri et al. 2017; Manohar Reddy & Ray 2023). Real-time flood forecasting systems with sufficient lead time are considered as the most efficient and effective way to mitigate flood risk (Yatheendradas et al. 2008). Flood forecasting based on the integration of meteorological and hydrological models has shown markedly better improvement in recent times due to the advancements in computational power, remote-sensing based observations, and improved understanding of hydrological processes (Das et al. 2022). In the context of a flood modelling framework, event-based models are often considered as sound alternatives to continuous models as they are easier to calibrate and require only data at the event scale (Tramblay et al. 2012). However, the efficiency of both continuous and event-based hydrologic predictions, irrespective of the model structure, is highly influenced by the optimal simulation of model states/initial hydrological conditions (IHCs) (often considered as soil moisture) within the modelled watershed (Crow & Ryu 2009; Loizu et al. 2018). Consequently, multitudes of modelling efforts were reported so far on the accurate estimation of the magnitude of floods, particularly on minimising the uncertainties associated with the estimation of IHCs (Brocca et al. 2008, 2009, 2010; Berthet et al. 2009; Tramblay et al. 2010; Alvarez-Garreton et al. 2014, 2015).
In general, a separate method is needed to estimate the initial states of the event-based models that are separately estimated by establishing relationships between IHC and external predictors. For instance, Antecedent Precipitation Index (API), computed by cumulating rainfall values of preceding days, is often used as a predictor in models based on the Soil Conservation Service Curve Number method (SCS-CN). Various predictors were reported so far in literature to accurately estimate the IHC for event-based models, such as piezometric levels (Coustau et al. 2012), baseflow (Franchini et al. 1996; Longobardi & Villani 2003), Antecedent Discharge Index (Tramblay et al. 2012), outputs from continuous models (Huang et al. 2016; Yao et al. 2019; Hegdahl et al. 2020), and in situ or remote-sensing based observables (Meng et al. 2017; Bahramian et al. 2021). Real-time measurements or climatology-based values can be directly used for updating if the simulated model states reliably represent the observable model states (Berthet et al. 2009). Moreover, it is also worth noting that the relationship between the model states and the external predictors is determined by the model structure and uncertainty in the input data.
Data assimilation (DA) techniques are widely employed to reduce the uncertainties associated with inputs, model states, and output variables (McLaughlin 2002). DA updates the model states by optimally combining the information from observations and model simulations, leading to improved estimates of the initial states of a hydrological model (Sun et al. 2016). Observed datasets used to update the states of a hydrological model include streamflow (Seo et al. 2003), soil moisture (Brocca et al. 2010), snow-covered area, snow water equivalent (Andreadis & Lettenmaier 2006; Clark et al. 2006), and satellite observations of soil moisture and discharge (Andreadis et al. 2007). Discharge data is the most frequently assimilated variable since it contains the collective information of all other hydrological states (Clark et al. 2008). However, due to the difficulty in obtaining observations in real-time, many studies have used satellite-based soil moisture for DA (Alvarez-Garreton et al. 2015; Baguis & Roulin 2017; Abbaszadeh et al. 2020). In the majority of DA studies, soil moisture observations are rescaled into the model space prior to assimilation, which influences the assimilation efficiency (Crow & Van Loon 2006; Crow & Reichle 2008). For instance, a poor rescaling method coupled with incorrect assumptions of the observational and model structural errors deteriorates the performance of DA (Tugrul Yilmaz & Crow 2013).
A recent study by Nayak et al. (2021) showed that the efficiency of DA in reliable simulation of streamflow depends on the model structure. Various model structures are available based on runoff generation mechanisms such as physics-based, conceptual, and data-driven, and based on the spatial resolution of the model such as lumped, semi-distributed, and distributed. However, in real-time operational applications, a trade-off between input data, model structural complexities, and computation costs is necessary to issue reliable and timely flood warnings (Butts et al. 2004). Although physics-based, fully distributed models account for the spatial heterogeneity of meteorological forcing and physical features within the basin leading to better simulations, they are data intensive and demand more computational resources (Young 2002). On the contrary, the utilisation of conceptual hydrologic models (CHMs) in operational streamflow forecasting has been amplified over the past few years because their computational efficiency and simplicity proved their effectiveness in streamflow forecasting (Hapuarachchi et al. 2011). Many of the DA studies were reported on the lumped versions of conceptual hydrological models and only a few were reported on semi-distributed continuous models (Berthet et al. 2009; Pathiraja et al. 2012; Alvarez-Garreton et al. 2014, 2015; Loizu et al. 2018). Previous studies reported that accounting for spatial variability in CHMs enhances model simulations (Das et al. 2008; De Lavenne et al. 2016). Hence, it is essential to understand the relevance of DA-estimated initial states in improving the performance of an event-based conceptual hydrological model.
In scenarios where the initial states for event-based models are estimated through their continuous counterparts, the choice of calibration metrics plays a key role. This is due to the fact that the commonly used calibration objective functions have the tendency to emphasise on particular flow segments of a hydrograph and this bias gets translated into the model states simulated by the continuous models (Mizukami et al. 2019). Hence, this study is intended to assess the performance of a lumped and a semi-distributed event-based conceptual model with initial conditions estimated using DA and their continuous counterparts corresponding continuous models calibrated using different calibration objective functions.
STUDY AREA AND HYDROMETEOROLOGICAL DATA
Study area
Hydrometeorological data
The daily streamflow data at Jagdalpur, from 1965 to 2018, is available at the India-WRIS portal. The observed flood events are separated from the streamflow data using the warning level obtained from the Central Water Commission appraisal report on Flood Forecasting and Warning Network Performance (CWC 2018). In this study, the daily gridded meteorological forcing datasets are obtained from India Meteorological Department (IMD) at a spatial resolution of 0.25° × 0.25° (precipitation) and 1° × 1° (minimum and maximum temperature) (Srivastava et al. 2009; Pai et al. 2014). For assimilation of soil moisture, root zone soil moisture is obtained from the Global Land Data Assimilation System (GLDAS), Catchment Land Surface Model, V2.0, where the data are available from 1948 to 2014.
METHODOLOGY
Hydrological model
. | . | Parameter . | Description . | Range . |
---|---|---|---|---|
Semi-distributed GR4J | GR4J | X1 | Production reservoir capacity (mm) | 1 to 1,500 |
X2 | Inter-catchment exchange coefficient (mm/day) | −10 to 5 | ||
X3 | Routing store capacity (mm) | 1 to 500 | ||
X4 | Unit hydrograph time constant (days) | 0.5 to 4 | ||
Muskingum routing | Average wave celerity in the stream network in m/s | 0 to 5 | ||
Dimensionless weighting factor | 0 to 0.5 |
. | . | Parameter . | Description . | Range . |
---|---|---|---|---|
Semi-distributed GR4J | GR4J | X1 | Production reservoir capacity (mm) | 1 to 1,500 |
X2 | Inter-catchment exchange coefficient (mm/day) | −10 to 5 | ||
X3 | Routing store capacity (mm) | 1 to 500 | ||
X4 | Unit hydrograph time constant (days) | 0.5 to 4 | ||
Muskingum routing | Average wave celerity in the stream network in m/s | 0 to 5 | ||
Dimensionless weighting factor | 0 to 0.5 |
Implementation of DA
Ensemble Kalman filter (EnKF)
In Equation (4), represents the analysis of model states posterior to the update, is an operator to covert model states to the observational space with dimensions , represents the vector of observations with dimensions ( denotes the number of observations), K represents the Kalman gain, and R is the observation error covariance matrix with dimensions . It should be noted that each ensemble member is updated individually using Equation (5). In general, in the implementation of EnKF, each of the ensemble member is updated using vectors of observations. The values of for each ensemble member are used to generate vectors of observations by sampling from a distribution with observations as mean and R as variance.
Bias correction of the GLDAS root zone soil moisture data
Forecast error
To address input uncertainty, meteorological datasets such as precipitation and evapotranspiration are stochastically perturbed. To achieve this, Clark et al. (2008) recommend perturbing daily meteorological observations using multiplicative stochastic noise at each time step. A first-order autoregressive model is employed to ensure the temporal correlation of time-variant forcings and physical consistency. To generate 100 ensemble meteorological forcing datasets, a temporal decorrelation length of 1 day and 2 days were used for rainfall and evapotranspiration, respectively, with a fractional error parameter set to 0.65. For a more detailed explanation of the perturbation method, refer to Clark et al. (2008).
Observation error
Following the study by Piazzi et al. (2021), the error parameter was set to 0.1. To prevent the underestimation of error variances at lower values of observations, the minimum threshold for defining the error variance was assumed to be the 10th percentile value of observations . Following Thirel et al. (2010), the variance is evaluated proportionally to for observed values below .
Model calibration and validation
Performance evaluation
Performance evaluation measure . | Range . | Performance rating . |
---|---|---|
NSE | 0.75–1.00 | Very good |
0.65–0.75 | Good | |
0.50–0.65 | Satisfactory | |
PBIAS | <± 10% | Very good |
±10% to ±15% | Good | |
±15% to ±25% | Satisfactory | |
PEPF | <± 15% | Very good |
±15% to ±30% | Good | |
±30% to ±40% | Satisfactory | |
PETP | <± 10% | Very good |
±10% to ±15% | Good | |
±15% to ±30% | Satisfactory |
Performance evaluation measure . | Range . | Performance rating . |
---|---|---|
NSE | 0.75–1.00 | Very good |
0.65–0.75 | Good | |
0.50–0.65 | Satisfactory | |
PBIAS | <± 10% | Very good |
±10% to ±15% | Good | |
±15% to ±25% | Satisfactory | |
PEPF | <± 15% | Very good |
±15% to ±30% | Good | |
±30% to ±40% | Satisfactory | |
PETP | <± 10% | Very good |
±10% to ±15% | Good | |
±15% to ±30% | Satisfactory |
RESULTS
Performance evaluation of continuous streamflow simulation
The accuracy of the streamflow simulated by the continuous lumped and semi-distributed GR4J models is evaluated using a set of metrics as described in Section 3.4. The performance evaluation metrics for the calibration and validation periods are tabulated in Tables 3 and 4, respectively. From Table 3, it can be noticed that the performance of the R4MS4E calibrated lumped GR4J model is good at simulating the peak flow values. Higher values of logNSE and Skill Score of the lumped GR4J model with streamflow assimilation indicate its better performance in terms of matching the observed flow frequencies and also in terms of simulating the low flows. The PBIAS of the KGE calibrated lumped model is found to be low attributable to the bias factor () in the formulation of KGE. Furthermore, the PBIAS values of the R4MS4E calibrated model show its overestimation bias, whereas the other models slightly underestimate the flows (< −10%). In the case of semi-distributed GR4J, a significant improvement is observed in both open-loop (without assimilation) and assimilated models in terms of all chosen evaluation criteria. The R4MS4E values in semi-distributed GR4J models were reduced by an average of 18.9% in all models, when compared to their lumper counterparts. The PBIAS values of streamflow and soil moisture assimilated models along with NSE and KGE calibrated open-loop models were found to be less than ±5%, whereas logNSE and R4MS4E calibrated models were found to be underestimating and overestimating the flows, respectively. It can also be noticed that the performance of KGE calibrated models in terms of logNSE and Skill Score is low (–0.07 and 0.45) showing its inability to capture low flows. In summary, the performance of the semi-distributed GR4J model with soil moisture assimilation is poor in continuous simulation of streamflow. The semi-distributed GR4J model, with streamflow assimilation and model calibrated with NSE, logNSE, and KGE, comparatively performs well in terms of chosen evaluation criteria in the calibration period. Similar performance can be observed in the validation period for both lumped and semi-distributed GR4J models (Table 4). This indicates better transferability of the calibrated parameters and better assimilation efficiency in the validation period.
. | R4MS4E . | logNSE . | Skill score . | PBIAS . |
---|---|---|---|---|
Lumped GR4J | ||||
Soil moisture assimilation | 543.47 | −0.75 | 0.65 | −12.12 |
Streamflow assimilation | 407.5 | 0.9 | 0.89 | −10.61 |
NSE calibrated | 345.23 | 0.74 | 0.67 | −3.08 |
KGE calibrated | 360.22 | 0.66 | 0.63 | −0.7 |
logNSE calibrated | 402.33 | 0.87 | 0.79 | −11.79 |
R4MS4E calibrated | 312.36 | 0.7 | 0.78 | 40.22 |
Semi-distributed GR4J | ||||
Soil moisture assimilation | 426.07 | 0.06 | 0.64 | 5.15 |
Streamflow assimilation | 340.38 | 0.92 | 0.85 | 4.55 |
NSE calibrated | 238.59 | 0.87 | 0.66 | −1.41 |
KGE calibrated | 291.23 | −0.07 | 0.45 | −0.73 |
logNSE calibrated | 380.12 | 0.91 | 0.86 | −22.58 |
R4MS4E calibrated | 251.94 | 0.73 | 0.66 | 29.05 |
. | R4MS4E . | logNSE . | Skill score . | PBIAS . |
---|---|---|---|---|
Lumped GR4J | ||||
Soil moisture assimilation | 543.47 | −0.75 | 0.65 | −12.12 |
Streamflow assimilation | 407.5 | 0.9 | 0.89 | −10.61 |
NSE calibrated | 345.23 | 0.74 | 0.67 | −3.08 |
KGE calibrated | 360.22 | 0.66 | 0.63 | −0.7 |
logNSE calibrated | 402.33 | 0.87 | 0.79 | −11.79 |
R4MS4E calibrated | 312.36 | 0.7 | 0.78 | 40.22 |
Semi-distributed GR4J | ||||
Soil moisture assimilation | 426.07 | 0.06 | 0.64 | 5.15 |
Streamflow assimilation | 340.38 | 0.92 | 0.85 | 4.55 |
NSE calibrated | 238.59 | 0.87 | 0.66 | −1.41 |
KGE calibrated | 291.23 | −0.07 | 0.45 | −0.73 |
logNSE calibrated | 380.12 | 0.91 | 0.86 | −22.58 |
R4MS4E calibrated | 251.94 | 0.73 | 0.66 | 29.05 |
Ideal values for different model settings are bolded.
. | R4MS4E . | logNSE . | Skill Score . | PBIAS . |
---|---|---|---|---|
Lumped GR4J | ||||
Soil moisture assimilation | 438.33 | −0.86 | 0.69 | −14.73 |
Streamflow assimilation | 337.84 | 0.9 | 0.91 | −10.7 |
NSE calibrated | 280 | 0.83 | 0.77 | −3.34 |
KGE calibrated | 325.5 | 0.78 | 0.73 | −0.94 |
logNSE calibrated | 363.18 | 0.85 | 0.86 | −10.24 |
R4MS4E calibrated | 256.75 | 0.6 | 0.66 | 45.7 |
Semi-distributed GR4J | ||||
Soil moisture assimilation | 341.77 | 0.36 | 0.69 | 17.03 |
Streamflow assimilation | 286.17 | 0.92 | 0.87 | 4.75 |
NSE calibrated | 224.92 | 0.89 | 0.84 | 4.44 |
KGE calibrated | 316.53 | 0.36 | 0.53 | 2.67 |
logNSE calibrated | 368.96 | 0.89 | 0.75 | −22.05 |
R4MS4E calibrated | 253.09 | 0.59 | 0.53 | 35.62 |
. | R4MS4E . | logNSE . | Skill Score . | PBIAS . |
---|---|---|---|---|
Lumped GR4J | ||||
Soil moisture assimilation | 438.33 | −0.86 | 0.69 | −14.73 |
Streamflow assimilation | 337.84 | 0.9 | 0.91 | −10.7 |
NSE calibrated | 280 | 0.83 | 0.77 | −3.34 |
KGE calibrated | 325.5 | 0.78 | 0.73 | −0.94 |
logNSE calibrated | 363.18 | 0.85 | 0.86 | −10.24 |
R4MS4E calibrated | 256.75 | 0.6 | 0.66 | 45.7 |
Semi-distributed GR4J | ||||
Soil moisture assimilation | 341.77 | 0.36 | 0.69 | 17.03 |
Streamflow assimilation | 286.17 | 0.92 | 0.87 | 4.75 |
NSE calibrated | 224.92 | 0.89 | 0.84 | 4.44 |
KGE calibrated | 316.53 | 0.36 | 0.53 | 2.67 |
logNSE calibrated | 368.96 | 0.89 | 0.75 | −22.05 |
R4MS4E calibrated | 253.09 | 0.59 | 0.53 | 35.62 |
Ideal values for different model settings are bolded.
Performance evaluation of event-based models
In this study, the performance of the lumped and semi-distributed GR4J models in simulating the historical flood events using the IHC estimated from different methods (through data assimilation and continuous models) has been tested. Berthet et al. (2009) suggested that, for evaluating the predictive performance of event-based models, careful selection of evaluation criteria is essential which emphasises the match between magnitude, volume, and timing observed, as well as the simulated flood hydrographs. In addition to quantitative measures, a visual inspection of the match between observed and simulated flood hydrographs is necessary. Accordingly, four quantitative performance assessment metrics were selected namely, NSE (objective function), Percentage Error in Peak Flow (PEPF), PBIAS, and Percentage Error in Timing to Peak (PETP), to evaluate the simulated flood response.
Performance of event-based models using IHC obtained through DA
The aforementioned statistics are computed for both lumped and semi-distributed event-based models with IHC estimated by assimilating soil moisture and streamflow (Table 5). In Table 5, the statistics are computed for each flood event using the observed flow and ensemble mean of simulations. The median value of the statistics computed for the events used for the calibration and validation periods separately are given in Table 5. The median NSE value of the event-based lumped model based on soil moisture-based data assimilation (SM-DA) and streamflow-based data assimilation (Q-DA) is higher than 0.5 in the calibration period, indicating satisfactory performance. However, the performance of continuous simulations from the lumped model based on SM-DA and Q-DA in simulating the flood hydrographs during the calibration period is found to be low (median NSE < 0.4). In the validation period, the performance of all lumped models based on DA is not satisfactory in terms of the median NSE value (<0.5). Additionally, the median NSE values in the validation period suggest that the temporal transferability of the calibrated parameters is poor, leading to poor simulations.
Assimilation variable . | NSE . | PEPF . | PBIAS . | PETP . | NSE . | PEPF . | PBIAS . | PETP . |
---|---|---|---|---|---|---|---|---|
Lumped | Calibration | Validation | ||||||
SM-EVT | 0.56 | −29.37 | −12.5 | −19.46 | 0.28 | −48.58 | 0 | −34.94 |
Q-EVT | 0.6 | − 22.64 | −6.25 | −16 | 0.39 | − 37.91 | 0 | − 29.96 |
SM-CONT | −0.04 | −59.45 | −16.67 | −46.8 | −0.95 | −77.94 | −15.48 | −67.82 |
Q-CONT | 0.39 | −47.16 | 0 | −18.13 | 0.46 | −44.63 | 0 | −14.93 |
Discretized | ||||||||
SM-EVT | 0.82 | − 4.45 | 0 | − 6.25 | 0.44 | −39.93 | 0 | −23.54 |
Q-EVT | 0.64 | −14.7 | 0 | −14.13 | 0.57 | −26.33 | 0 | −27.16 |
SM-CONT | 0.41 | −46.92 | 0 | −19.54 | 0.3 | −50.83 | 0 | −25.56 |
Q-CONT | 0.42 | −14.3 | 14.29 | 12.04 | 0.38 | −16.66 | 15.48 | 9.47 |
Assimilation variable . | NSE . | PEPF . | PBIAS . | PETP . | NSE . | PEPF . | PBIAS . | PETP . |
---|---|---|---|---|---|---|---|---|
Lumped | Calibration | Validation | ||||||
SM-EVT | 0.56 | −29.37 | −12.5 | −19.46 | 0.28 | −48.58 | 0 | −34.94 |
Q-EVT | 0.6 | − 22.64 | −6.25 | −16 | 0.39 | − 37.91 | 0 | − 29.96 |
SM-CONT | −0.04 | −59.45 | −16.67 | −46.8 | −0.95 | −77.94 | −15.48 | −67.82 |
Q-CONT | 0.39 | −47.16 | 0 | −18.13 | 0.46 | −44.63 | 0 | −14.93 |
Discretized | ||||||||
SM-EVT | 0.82 | − 4.45 | 0 | − 6.25 | 0.44 | −39.93 | 0 | −23.54 |
Q-EVT | 0.64 | −14.7 | 0 | −14.13 | 0.57 | −26.33 | 0 | −27.16 |
SM-CONT | 0.41 | −46.92 | 0 | −19.54 | 0.3 | −50.83 | 0 | −25.56 |
Q-CONT | 0.42 | −14.3 | 14.29 | 12.04 | 0.38 | −16.66 | 15.48 | 9.47 |
Note that SM represents soil moisture assimilation and Q represents streamflow assimilation. EVT represents simulations from the event-based model, whereas CONT represents simulations from continuous models.Ideal values for different model settings are bolded.
Performance of event-based models using IHC obtained through continuous models calibrated with different objective functions
Calibration objective function . | NSE . | PEPF . | PBIAS . | PETP . | NSE . | PEPF . | PBIAS . | PETP . |
---|---|---|---|---|---|---|---|---|
Lumped | Calibration | Validation | ||||||
KGE | 0.69 | −25.65 | 0 | 9.34 | 0.55 | −24.72 | 0 | 5.33 |
R4MS4E | 0.67 | −25.96 | 0 | −2.61 | 0.55 | −25.58 | 0 | − 0.35 |
logNSE | 0.49 | −30.63 | 0 | −12.57 | 0.51 | −30.1 | 0 | −14.15 |
NSE | 0.7 | −26.57 | 0 | 5.58 | 0.57 | −26.43 | 0 | −2.86 |
Discretized | ||||||||
KGE | 0.85 | 2.99 | 0 | 4.23 | 0.65 | −12.88 | 0 | − 2.27 |
R4MS4E | 0.79 | −13.67 | 0 | 3.86 | 0.68 | −23.32 | 0 | −10.03 |
logNSE | 0.77 | −11.36 | 0 | 2.57 | 0.56 | −24.83 | 0 | −8.42 |
NSE | 0.91 | − 3.25 | 0 | 0.44 | 0.77 | −13.21 | 0 | −8.73 |
Calibration objective function . | NSE . | PEPF . | PBIAS . | PETP . | NSE . | PEPF . | PBIAS . | PETP . |
---|---|---|---|---|---|---|---|---|
Lumped | Calibration | Validation | ||||||
KGE | 0.69 | −25.65 | 0 | 9.34 | 0.55 | −24.72 | 0 | 5.33 |
R4MS4E | 0.67 | −25.96 | 0 | −2.61 | 0.55 | −25.58 | 0 | − 0.35 |
logNSE | 0.49 | −30.63 | 0 | −12.57 | 0.51 | −30.1 | 0 | −14.15 |
NSE | 0.7 | −26.57 | 0 | 5.58 | 0.57 | −26.43 | 0 | −2.86 |
Discretized | ||||||||
KGE | 0.85 | 2.99 | 0 | 4.23 | 0.65 | −12.88 | 0 | − 2.27 |
R4MS4E | 0.79 | −13.67 | 0 | 3.86 | 0.68 | −23.32 | 0 | −10.03 |
logNSE | 0.77 | −11.36 | 0 | 2.57 | 0.56 | −24.83 | 0 | −8.42 |
NSE | 0.91 | − 3.25 | 0 | 0.44 | 0.77 | −13.21 | 0 | −8.73 |
Ideal values for different model settings are bolded.
SUMMARY AND CONCLUSIONS
This study evaluates the performance of event-based models based on the IHC/model states using different methods. The current study on event-based models led to the following insights:
- (i)
Streamflow assimilation (Q-DA) in both lumped and semi-distributed models led to improved simulations at the catchment outlet during both calibration and validation periods. In contrast, assimilating soil moisture (SM-DA) did not enhance continuous streamflow simulations in the lumped model, but showed a slight improvement in the semi-distributed model.
- (ii)
Semi-distributed models resulted in enhanced streamflow simulation in both SM-DA and Q-DA, when compared to their lumped counterparts. This result is in line with the findings of Alvarez-Garreton et al. (2015).
- (iii)
Compared to simulations produced by continuous models with DA, the ensemble spread of simulations generated from the event-based model using IHCs obtained through DA is lower. However, simulations based on SM-DA have a higher ensemble spread than those based on Q-DA.
- (iv)
In event-based models with initial conditions generated through assimilation, SM-DA-based semi-distributed models performed exceptionally well in the calibration period exploiting the advantage of model spatial resolution.
- (v)
The event-based semi-distributed GR4J initialised by IHCs extracted from the corresponding NSE calibrated continuous model is found to be outperforming other models in terms of capturing the magnitude, timing, and volume of the observed flood events.
In conclusion, the findings suggest the potential benefits of assimilating streamflow and employing a semi-distributed model structure. However, limitations such as overfitting of parameters and uncertainties in channel characteristics may affect model performance (Khakbaz et al. 2012; Jehn et al. 2019). Future research can explore the integration of rainfall spatial variability in a lumped model to exploit its computational advantage (Zhou et al. 2021). Additionally, further investigation is needed to evaluate the efficiency of the combined assimilation of soil moisture in IHC estimation and to assess the suitability of EnKF for assimilation in various hydrological models. Overall, this study provides valuable insights for practical applications and calls for large sample investigations using multiple model structures (Gupta et al. 2014).
ACKNOWLEDGEMENT
This work has been funded by the Indian Institute of Tropical Meteorology under Monsoon Mission Phase-II (Project no IITM/MM-II/NIT/2018/IND-6/).
DATA AVAILABILITY STATEMENT
Data cannot be made publicly available; readers should contact the corresponding author for details.
CONFLICT OF INTEREST
The authors declare there is no conflict.