An Improved Coupled Routing and Excess Storage (CREST) Distributed Hydrological Model and Its Verification in Ganjiang River Basin, China

The coupled routing and excess storage (CREST) distributed hydrological model has been applied regionally and globally for years. With the development of remote sensing, requirements for data assimilation and integration have become new challenges for the CREST model. In this paper, an improved CREST model version 3.0 (Tsinghua University and China Institute of Water Resources and Hydropower Research, Beijing, China) is proposed to enable the use of remotely-sensed data and to further improve model performance. Version 3.0 model’s runoff generation, soil moisture, and evapotranspiration based on three soil layers to make the CREST model friendly to remote sensing products such as soil moisture. A free water reservoir-based module which separates three runoff components and a four mechanism-based cell-to-cell routing module are also developed. Traditional CREST and CREST 3.0 are applied in the Ganjiang River basin, China to compare their simulation capability and applicability. Research results indicate that CREST 3.0 outperforms the traditional model and has good application prospects in data assimilation, flood forecasting, and water resources planning and management applications.


Review of Distributed Hydrological Model
Hydrological models have been successfully invented and applied for decades worldwide. Based on the structures and configurations of the model, hydrological models can be categorized into three types,

Traditional CREST Model-Version 2.x
The CREST model was initially designed to provide real-time regional and global hydrological predictions by simultaneously modeling over multi-basins with significantly cost-effective computational efficiency [23]. It is also applicable for small-to-medium-size basins at very high-resolutions [24,26]. CREST simulates the spatiotemporal variations of water fluxes and storage on a regular grid with the grid cell resolution being user-defined, thereby enabling multi-scale applications. The scalability of CREST simulations is accomplished through sub-grid-scale representation of soil moisture storage capacity (using a variable infiltration curve) and multi-scale runoff generation processes (using multi-linear reservoirs). The representation of the primary water fluxes such as infiltration and routing are physically related to the spatially variable land surface characteristics (i.e., vegetation, soil type, topography, etc.). The runoff generation process and routing scheme are coupled, thus providing more realistic interactions between lower atmospheric boundary layers, terrestrial surface, and subsurface water. The above flexible modeling features and embedded automated calibration algorithms make CREST a powerful yet cost-effective tool for distributed hydrological modeling and implementation at global, regional, basin, and small catchment scales [25].
The principles of the traditional CREST 2.x model can be found in relevant literature [23,24,26,33,34]. The brief principles and calculation processes of the CREST 2.x model are demonstrated in Figure 1 [25] and in the following paragraphs. As we can see in Figure 1a, the CREST 2.x model requires grid-based precipitation, potential evapotranspiration (PET), and watershed characteristics (such as DEM (Digital Elevation Model), flow direction, flow accumulation, slope, etc.) as forcing data or model inputs. After the canopy interception, infiltration, runoff generation, evapotranspiration (ET) computation, and cell-to-cell routing, the model generates discharge, actual ET (AET), soil moisture, runoff, etc. As demonstrated in Figure 1b, the model is composed of a variable infiltration curve-based runoff generation module, a saturated hydraulic conductivity-based runoff component separation module, a soil moisture computation module, and a cell-to-cell routing module.

Tension Water Storage Capacity Distribution Curve-Based Runoff Generation
For each grid cell, the total runoff generated by the CREST 3.0 is computed by using a tension water storage capacity distribution curve. CREST 3.0 supposes that the runoff generation has a close relationship with the amount of net rainfall and tension water storage. The runoff does not generate until the tension water storage reaches the field capacity. Before the runoff generates, the rainfall replenishes the tension water storage after deduction of the evapotranspiration. CREST 3.0 adopts a tension water storage capacity distribution curve, which is also adopted in previous versions of CREST model, to describe the sub-grid inhomogeneity of the tension water storage capacity. The equation of the distribution curve is [23]: where f/F represents the proportion of the pervious area of the grid cell whose tension water storage capacity is less than or equal to the value of WM'; WM' represents the tension water capacity at a point; WMM represents the maximum value of WM'; b is a model parameter that reflects the inhomogeneity degree of the tension water storage capacity in the vadose zone. The areal mean tension water capacity, WM, can be computed as where IM is the proportion of the impervious area of the grid cell. The total runoff, R, is generated on the pervious area of the grid cell as: Otherwise 2.2. Improved CREST Model-Version 3.0

Tension Water Storage Capacity Distribution Curve-Based Runoff Generation
For each grid cell, the total runoff generated by the CREST 3.0 is computed by using a tension water storage capacity distribution curve. CREST 3.0 supposes that the runoff generation has a close relationship with the amount of net rainfall and tension water storage. The runoff does not generate until the tension water storage reaches the field capacity. Before the runoff generates, the rainfall replenishes the tension water storage after deduction of the evapotranspiration. CREST 3.0 adopts a tension water storage capacity distribution curve, which is also adopted in previous versions of CREST model, to describe the sub-grid inhomogeneity of the tension water storage capacity. The equation of the distribution curve is [23]: where f /F represents the proportion of the pervious area of the grid cell whose tension water storage capacity is less than or equal to the value of WM'; WM' represents the tension water capacity at a point; WMM represents the maximum value of WM'; b is a model parameter that reflects the inhomogeneity degree of the tension water storage capacity in the vadose zone. The areal mean tension water capacity, WM, can be computed as where IM is the proportion of the impervious area of the grid cell. The total runoff, R, is generated on the pervious area of the grid cell as: Otherwise where P represents precipitation; EM represents potential evapotranspiration; W represents tension water storage; KC represents model parameter used to modify the potential evapotranspiration; and A is an index indicating the saturation degree of the grid cell, and is computed as

Three Soil Layers-Based Soil Moisture and Evapotranspiration Computation
In CREST 3.0, evapotranspiration is related to potential evapotranspiration through a three-layer soil moisture module depending on parameters KC, WUM, WLM, and C. Until the storage WU of the uppermost layer is exhausted, evaporation occurs at the potential rate, equal to KC times the potential evaporation as [23]: Upon exhaustion of the upper layer (capacity WUM), any remaining potential evapotranspiration is applied to the lower layer, but the efficiency is modified by multiplication by the ratio of the actual storage WL to the capacity storage WLM of that layer as When the lower layer storage WL is reduced to a proportion C of WLM, evapotranspiration is assumed to continue, but at a further reduced rate, and the evapotranspiration ED is given by

Free Water Reservoir-Based Separation of Three Runoff Components
The total runoff, R, is separated into three components including RS, overland runoff, RG, the ground water contribution, and RI, a contribution to interflow. The concepts of free water storage S and free water storage capacity SM are used. The latter is assumed to be distributed between zero and a point maximum SMM in a parabolic manner, over FR, that portion of the sub-grid which is currently producing runoff. The distribution curve of the SM in a grid cell is stated as [23] where f is the portion of the sub-grid area for which the free water storage capacity is less than or equal to SM' and EX is a model parameter that describes the inhomogeneity of the SM. The areal mean free water storage capacity SM may be used instead of SMM as a model parameter, and has the relationship Assume AU is an index indicating the free water storage of the grid cell and it is computed as and FR is computed as The interflow and ground water are computed as where KG and KI are discharge coefficients for ground water and interflow, respectively. The overland runoff is computed as: Otherwise 2.2.4. Four Mechanisms-Based Cell-To-Cell Routing CREST 3.0 adopts the four mechanisms-based cell-to-cell routing method to implement flow routing of overland flow, interflow, ground water, and river channel flow. For each upstream grid cell, the arrival time from the upstream grid cell to its downstream basin outlet grid cell is accumulated along the flow path based on the flow concentration time within each grid cell. The flow path is generated based on the eight flow direction method adopted in ArcGIS software. The flow concentration time within grid cell (i, j) is computed as [23]: where T i,j represents the flow concentration time within grid cell (i, j); L i,j represents the distance between the centers of grid cell (i, j) and its adjacent downstream grid cell; S i,j represents the slope from grid cell (i, j) to its adjacent downstream grid cell; K is defined as the runoff velocity coefficient. The value of K varies from cell to cell and within cells depending on the following four processes that it represents: for ground water runoff, K is set to a value representative of the ground water flow velocity; for interflow runoff, K is set to a value representative of the soil saturated hydraulic conductivity; for overland runoff, K corresponds to the land surface roughness; and for river channel flow, K is a channel velocity coefficient determined by the channel roughness and hydraulic radius [23]. The interflow and ground water flow move more slowly in response to relatively small values of K. In contrast, overland flow and river channel flow move faster and correspond to larger values of K. All values of K-which control the timing of peak flow-can be provided a priori using land cover maps, soil surveys, and channel cross-sections, but typically must be optimized through calibration [23].
The overland, interflow, and ground water runoff are generated in a hill slope grid cell and flow downstream along their flow paths according to the eight flow direction method. The runoff component retains its flow movement mechanism until it encounters the stream grid cell. After the runoff comes into the stream grid cell, it will flow downstream according to the river channel routing mechanism. The total flow concentration time of the upstream grid cell is accumulated along its flow path, meanwhile considering its flow movement mechanism.

Model Calibration
The CREST 3.0 model has 15 parameters, which can be found in Table 1. The model parameters are optimized using the SCE-UA optimization method developed by Duan [35][36][37][38][39][40][41][42][43][44]. The objective function of the parameter optimization is the Nash-Sutcliffe coefficient of efficiency (NSCE). Some parameters or state variables of the CREST 3.0 model have strong physical meanings, and their values must obey some constraints including KG + KI < 1, non-negative tension water storages, and KRF ≥ KOF ≥ KIF ≥ KGF. These constraints are implemented by punishment function method. The constraints are treated as punishment terms which are added to the objective function value if the constraints are conflicted.

Study Area and Data Description
In this study, we apply the original CREST 2.x and the improved CREST 3.0 models in daily streamflow simulation in the Ganjiang River basin, China to verify the performance of the improved model.
The Ganjiang River is a tributary of the Yangtze River and is the longest river in Jiangxi Province, China. It is located in the southeast part of China. The basin outlet is the Waizhou station, and the drainage area is 81,258 km 2 . The elevation of the Ganjiang River basin ranges from 11 to 1997 m, and the terrain varies significantly from hilly land to low hill. The climate of the Ganjiang River basin is mainly subtropical humid monsoon climate, and rainfall usually happens during April and June and flood frequently happens during May and July [45,46]. The DEM, slope, flow direction, and flow routing sequence maps of the Ganjiang River basin are demonstrated in Figure 2.
This study adopts hydro-meteorological data from 2003 to 2009 to carry out daily streamflow simulation. The rainfall data adopts the TRMM (Tropical Rainfall Measuring Mission) satellite-based remote sensing product 3B42V7. The original spatiotemporal resolutions of the TRMM 3B42V7 rainfall data are 0.25 • × 0.25 • and 3 h. We resampled the TRMM rainfall data into spatiotemporal resolutions of 1 km × 1 km and 1 day by using the nearest-neighbor method and daily rainfall accumulation method. The potential evapotranspiration (PET) data was downloaded from the global daily potential evapotranspiration dataset of the Famine Early Warning System (FEWS) of the USGS (The United States Geological Survey) and the spatiotemporal resolutions are 1 • × 1 • and 1 day. The spatial resolution of the PET data was resampled into 1 km × 1 km by using the nearest-neighbor method. The daily streamflow data was obtained from Waizhou station observations.

Comparison of Basin Outlet Discharge Simulations between CREST 2.x and 3.0
For the purposes of comparing the simulation capability of the CREST 2.x and 3.0 models, the measured and simulated basin outlet hydrographs are compared herein. The basin outlet streamflow simulation results are demonstrated in Figure 3. As we can see in Figure 3, the performance of CREST 3.0 is overall better than the 2.x version. The low flow simulations of the CREST 2.x are significantly

Comparison of Basin Outlet Discharge Simulations between CREST 2.x and 3.0
For the purposes of comparing the simulation capability of the CREST 2.x and 3.0 models, the measured and simulated basin outlet hydrographs are compared herein. The basin outlet streamflow simulation results are demonstrated in Figure 3. As we can see in Figure 3, the performance of CREST 3.0 is overall better than the 2.x version. The low flow simulations of the CREST 2.x are significantly underestimated. In contrast, the low flow simulations of the CREST 3.0 version are much better than the 2.x model. CREST 3.0 can simulate the flow recession process and base flow of the Ganjiang River basin with satisfactory accuracy. runoff generation and flow concentration processes of the ground water. Therefore, the base flow and ground water simulations of CREST 3.0 outperform the 2.x version. The addition of the base flow simulation functionality of the improved 3.0 version is of great significance to global climate change research, global land surface modelling, water resources planning and management, etc. In these applications, the base flow plays a very important role in enclose the water cycle balance. Therefore, the 3.0 version has good application prospects in these study areas.    x only considers one soil layer. Therefore, it generates soil moisture of the total vadose zone. CREST 3.0 separates the vadose zone into three soil layers and considers the runoff generation, soil moisture, and evapotranspiration computations of upper, lower, and deep soil layers. It can be observed that the simulated total tension water soil moisture of CREST 2.x (W) was smaller than the 3.0 generated W. This indicates that CREST 2.x may underestimate the total soil moisture of the vadose zone due to the exclusion of the deep soil layer. CREST 2.x separates runoff into overland flow and interflow according to saturated hydraulic conductivity, and can be recognized as only considering the upper and lower layer soil moisture. These soil moistures correspond to upper (WU) and lower (WL) layer We also compared model performances of CREST 2.x and 3.0 according to Nash-Sutcliffe coefficient of efficiency (NSCE), BIAS (a statistic is biased if it is calculated in such a way that it is only systematically different from the population parameter of interest), regression R 2 , and root mean square error (RMSE) [47]. The error statistics of CREST 2.x and 3.0 are listed in Table 2. The NSCE of CREST 2.x and 3.0 were 0.75 and 0.77, respectively. The BIAS of CREST 2.x and 3.0 were −9.2302% and 0.0003%, respectively. The R 2 of CREST 2.x and 3.0 were both 0.78. The RMSE of CREST 2.x and 3.0 were 852.60 and 827.21, respectively. These results indicate that CREST 3.0 overall outperformed the 2.x version.  In Figure 6, we show the simulated tension water storages of CREST 2.x (W) and 3.0 (WU + WL). We can see that the W and WU + WL simulations are closer to each other.  In Figure 6, we show the simulated tension water storages of CREST 2.x (W) and 3.0 (WU + WL). We can see that the W and WU + WL simulations are closer to each other. The variation tendency and absolute values are similar to each other. For high values, CREST 3.0 simulations are larger than the 2.x version. For low values, CREST 3.0 simulations are smaller than the 2.x version. For medium values, the simulations of the two models are similar to each other. Figure 7 shows the scatter plot of simulated tension water storages of CREST 2.x (W) and 3.0 (WU + WL) models. The regression R 2 value between the CREST 2.x and 3.0 are 0.8594, which indicates a high linear correlation relationship between W and WU + WL. The distribution of scatters is even and good. These results prove that the vadose zone total tension water storage simulations of the CREST 2.x model are similar to the sum of upper and lower soil layers tension water storage simulations of CREST 3.0. In Figure 6, we show the simulated tension water storages of CREST 2.x (W) and 3.0 (WU + WL). We can see that the W and WU + WL simulations are closer to each other. The variation tendency and absolute values are similar to each other. For high values, CREST 3.0 simulations are larger than the 2.x version. For low values, CREST 3.0 simulations are smaller than the 2.x version. For medium values, the simulations of the two models are similar to each other.  In this research, we do not have sufficient station-observed soil moisture data to further validate the effectiveness and accuracy of the model simulations. This is because the soil moisture data in China is not publicly available to researchers. As a substitute, remotely-sensed soil moisture data provided by satellites such as AMSR-E, SMOS, and SMAP become valuable in research activities. However, both station-observed and remotely-sensed soil moisture data have their own limitations when applied to real-world scenarios. The distribution of soil moisture observation stations in most areas of China is sparse, therefore, the representativeness of the station-based data is poor when interpolating to the whole basin. Remote sensing data can provide spatial distribution of the soil moisture, but the uncertainty issue of remote sensing data is severe for soil moisture variables. Therefore, it is suggested that the validation of soil moisture simulations must combine the advantages of the station observation and remotely-sensed data in future studies.  Figure 8 shows the areal mean actual evapotranspiration (AET) simulations of the CREST 2.x model. Most AET simulations are good and reasonable. However, there are some unreasonable AET peak values that we call "singular points". These singular points are much larger than other AET simulation results, and may be unreasonable. After comparing the computed AET with remote sensing retrieved PET used in this study, we found that these singular values are even larger than observed PET values. Therefore, we confirm that the evapotranspiration computation algorithm or codes of the CREST 2.x model need to be improved, and the singular values are unreasonable.  In this research, we do not have sufficient station-observed soil moisture data to further validate the effectiveness and accuracy of the model simulations. This is because the soil moisture data in China is not publicly available to researchers. As a substitute, remotely-sensed soil moisture data provided by satellites such as AMSR-E, SMOS, and SMAP become valuable in research activities. However, both station-observed and remotely-sensed soil moisture data have their own limitations when applied to real-world scenarios. The distribution of soil moisture observation stations in most areas of China is sparse, therefore, the representativeness of the station-based data is poor when interpolating to the whole basin. Remote sensing data can provide spatial distribution of the soil moisture, but the uncertainty issue of remote sensing data is severe for soil moisture variables. Therefore, it is suggested that the validation of soil moisture simulations must combine the advantages of the station observation and remotely-sensed data in future studies.

Comparison of Areal Mean Actual Evapotranspiration
Simulations between CREST 2.x and 3.0 simulation results, and may be unreasonable. After comparing the computed AET with remote sensing retrieved PET used in this study, we found that these singular values are even larger than observed PET values. Therefore, we confirm that the evapotranspiration computation algorithm or codes of the CREST 2.x model need to be improved, and the singular values are unreasonable. Figure 7. Scatter plot of simulated tension water storages of CREST 2.x (one soil layer) and 3.0 (sum of upper and lower soil layers) models. CREST 2.x and 3.0 Figure 8 shows the areal mean actual evapotranspiration (AET) simulations of the CREST 2.x model. Most AET simulations are good and reasonable. However, there are some unreasonable AET peak values that we call "singular points". These singular points are much larger than other AET simulation results, and may be unreasonable. After comparing the computed AET with remote sensing retrieved PET used in this study, we found that these singular values are even larger than observed PET values. Therefore, we confirm that the evapotranspiration computation algorithm or codes of the CREST 2.x model need to be improved, and the singular values are unreasonable. In order to compare AET simulation results of CREST 2.x and 3.0, we exclude the abovementioned singular values. Figure 9 demonstrates simulated evapotranspiration of CREST 2.x (excluding singular points) and 3.0 models. CREST 2.x can only generate total AET, and CREST 3.0 generates the AET of upper, lower, and deep soil layers. It can be seen the total AET of CREST 3.0 is In order to compare AET simulation results of CREST 2.x and 3.0, we exclude the above-mentioned singular values. Figure 9 demonstrates simulated evapotranspiration of CREST 2.x (excluding singular points) and 3.0 models. CREST 2.x can only generate total AET, and CREST 3.0 generates the AET of upper, lower, and deep soil layers. It can be seen the total AET of CREST 3.0 is larger than the total AET of CREST 2.x, which indicates that CREST 2.x may underestimate the total AET due to the exclusion of deep soil layer evapotranspiration. It also can be found that most AET of the upper soil layer is larger than the AET of the lower soil layer, and the AET of the deep soil layer is usually very small. Figure 10 shows the scatter plot of the simulated evapotranspiration of CREST 2.x (excluding singular points) and 3.0 models. The regression R 2 between CREST 2.x and 3.0 is 0.875, which indicates a high correlation relationship. The scatter distribution is also satisfactory and even. Figure  10 indicates that the evapotranspiration of CREST 3.0 is overall larger than the 2.x version. These results indicate that the exclusion of deep soil layer tension water storage and evapotranspiration computation leads the CREST 2.x to generate lower AET compared with the CREST 3.0 model. The validation of ET simulations is much more difficult than soil moisture variables. Unlike the soil moisture, the true values of ET for different underlying surfaces (e.g., bare soil and vegetationcovered land surface) are difficult to obtain. Observed ET data are usually potential ET (PET) of the water surface, rather than actual ET (AET). Precise estimation of the relationship between PET and AET for all kinds of underlying surfaces is also very challenging. Even though the study of ET is a hot issue in remote sensing and hydrology studies, a solidly reliable ET data set or product is still difficult to obtain, and the uncertainty issue in ET estimation is severe. Considering that the stationobserved ET data is not sufficient, remote sensing-generated products (e.g., MODIS MOD16, AVHRR, etc.) are better substitutes for researchers and engineers in real-world applications. Therefore, we will further validate the ET simulations by using remotely-sensed ET products in future research.  Figure 10 shows the scatter plot of the simulated evapotranspiration of CREST 2.x (excluding singular points) and 3.0 models. The regression R 2 between CREST 2.x and 3.0 is 0.875, which indicates a high correlation relationship. The scatter distribution is also satisfactory and even. Figure 10 indicates that the evapotranspiration of CREST 3.0 is overall larger than the 2.x version. These results indicate that the exclusion of deep soil layer tension water storage and evapotranspiration computation leads the CREST 2.x to generate lower AET compared with the CREST 3.0 model. The validation of ET simulations is much more difficult than soil moisture variables. Unlike the soil moisture, the true values of ET for different underlying surfaces (e.g., bare soil and vegetation-covered land surface) are difficult to obtain. Observed ET data are usually potential ET (PET) of the water surface, rather than actual ET (AET). Precise estimation of the relationship between PET and AET for all kinds of underlying surfaces is also very challenging. Even though the study of ET is a hot issue in remote sensing and hydrology studies, a solidly reliable ET data set or product is still difficult to obtain, and the uncertainty issue in ET estimation is severe. Considering that the station-observed ET data is not sufficient, remote sensing-generated products (e.g., MODIS MOD16, AVHRR, etc.) are better substitutes for researchers and engineers in real-world applications. Therefore, we will further validate the ET simulations by using remotely-sensed ET products in future research.

Analysis of Areal Mean Runoff Generation Area and Free Water Storage Simulations
Apart from grid-based streamflow, soil moisture, and AET, CREST 3.0 can also generate runoff generation area (FR) and free water storage (S) simulations for each grid cell. Figure 11 demonstrates simulated areal mean runoff generation area and free water storage of the CREST 3.0 model. Most of the FR are between 0.1 and 0.9. This result indicates that for large basins such as the Ganjiang River basin, the tension water storage cannot be totally saturated (FR = 1) or totally exhausted (FR = 0) for the whole basin. This conclusion is reasonable. For very large basins such as the Ganjiang River basin, the spatial distribution of rainfall intensity is usually significantly uneven and the rainfall usually cannot fall on the whole basin for long enough to saturate the whole watershed. Therefore, the areal mean FR cannot be 1. For large basins, the base flow plays an important role in drought periods. The drought season streamflow is stable and larger than 0. The free water reservoir can be supplied by base flow, and the free water storage of the whole basin cannot be totally exhausted.

Analysis of Areal Mean Runoff Generation Area and Free Water Storage Simulations
Apart from grid-based streamflow, soil moisture, and AET, CREST 3.0 can also generate runoff generation area (FR) and free water storage (S) simulations for each grid cell. Figure 11 demonstrates simulated areal mean runoff generation area and free water storage of the CREST 3.0 model. Most of the FR are between 0.1 and 0.9. This result indicates that for large basins such as the Ganjiang River basin, the tension water storage cannot be totally saturated (FR = 1) or totally exhausted (FR = 0) for the whole basin. This conclusion is reasonable. For very large basins such as the Ganjiang River basin, the spatial distribution of rainfall intensity is usually significantly uneven and the rainfall usually cannot fall on the whole basin for long enough to saturate the whole watershed. Therefore, the areal mean FR cannot be 1. For large basins, the base flow plays an important role in drought periods. The drought season streamflow is stable and larger than 0. The free water reservoir can be supplied by base flow, and the free water storage of the whole basin cannot be totally exhausted.

Analysis of Areal Mean Runoff Generation Area and Free Water Storage Simulations
Apart from grid-based streamflow, soil moisture, and AET, CREST 3.0 can also generate runoff generation area (FR) and free water storage (S) simulations for each grid cell. Figure 11 demonstrates simulated areal mean runoff generation area and free water storage of the CREST 3.0 model. Most of the FR are between 0.1 and 0.9. This result indicates that for large basins such as the Ganjiang River basin, the tension water storage cannot be totally saturated (FR = 1) or totally exhausted (FR = 0) for the whole basin. This conclusion is reasonable. For very large basins such as the Ganjiang River basin, the spatial distribution of rainfall intensity is usually significantly uneven and the rainfall usually cannot fall on the whole basin for long enough to saturate the whole watershed. Therefore, the areal mean FR cannot be 1. For large basins, the base flow plays an important role in drought periods. The drought season streamflow is stable and larger than 0. The free water reservoir can be supplied by base flow, and the free water storage of the whole basin cannot be totally exhausted. Figure 11. Simulated areal mean runoff generation area and free water storage of the CREST 3.0 model. Figure 11. Simulated areal mean runoff generation area and free water storage of the CREST 3.0 model. The free water storage, S, reflects the quantity of runoff generated during the runoff generation process to some extent. Figure 11 shows that when FR is increasing, S is still increasing; when FR is decreasing, S is also decreasing. This result indicates that when runoff generation area, FR, is large, the generated runoff is also large. The FR and S react with the same tendency in the process of runoff generation.

Conclusions
In this paper, we make four improvements to the traditional CREST 2.x distributed hydrological model and propose a new version 3.0. With the addition of three soil layers-based soil moisture and evapotranspiration computations, free water reservoir-based separation of three runoff components, and four mechanisms-based cell-to-cell routing, CREST 3.0 outperforms the traditional 2.x version in simulation accuracy, output information, and application fields. CREST 3.0 is also friendlier than the 2.x version in its handling of remote sensing retrieved data such as soil moisture. Furthermore, following conclusions can be made here: (1) CREST 3.0 resolves the high and low flow underestimation problem of the CREST 2.x model. Version 3.0 considers the overland flow, interflow, ground water flow, and river channel flow, and may be more adaptable to different kinds of watersheds and may perform better in flood forecasting applications than the 2.x version.
(2) CREST 3.0 solves the soil moisture and evapotranspiration under-estimation problem of the CREST 2.x model. The 3.0 version considers the soil moisture and evapotranspiration of upper, lower, and deep soil layers. CREST 3.0 performs better in the water quantity estimation and may perform better in water resources planning and management applications.
(3) CREST 3.0 separates the soil layer into three layers, and the depth of the upper soil layer can be configured to adapt to remote sensing retrieved soil moisture data such as AMSR-E, SMAP, SMOS, and Sentinel-1. The 3.0 version is capable of using the remotely-sensed soil moisture and AET data to assist model calibration and validation. The CREST 3.0 model is more suitable to data assimilation tasks than the 2.x version.
(4) Apart from grid-based streamflow, soil moisture, and AET, CREST 3.0 can also generate runoff generation area and free water storage for grid cells. These additional products can provide more useful information for scientific research and real-world engineering applications.