Effects of Digital Elevation Model Resolution on Watershed-Based Hydrologic Simulation

Abstract

This paper investigated effects of Digital Elevation Model (DEM) resolution on flow simulation by applying the HSPF watershed model to three U.S. watersheds with different topographic conditions. Each watershed was delineated automatically and manually with four DEMs of the resolution ranging from 3.5 to 100 m. Results indicated that the simulated flow decreased with lowering DEM resolutions due to the reduction in the delineated drainage area particularly in low gradient watersheds. The DEM resolution impact was minimal when the manual method for watershed delineation was applied. The parameter uncertainty was found to be substantially greater than the resolution uncertainty in two out of three tested watersheds, indicating that the calibration of water balance parameters can alleviate the adverse effects of coarse DEM resolution for watersheds with high to moderate gradients. The findings are important to reducing the uncertainty, caused by DEM resolutions, in watershed modelling results, serving as guidelines for watershed modelling-based water resources management.

Introduction

Hydrologic simulation plays a pivotal role in understanding hydrologic processes and managing water resources. Particularly, flow simulation is needed not only for flood management but also for water quality management and restoration. The accuracy of flow simulation with watershed models is highly affected by model input data (Chaplot 2005) since the primary source of uncertainty in hydrologic modelling is caused by input data and the model simulated flow is dominated by the DEM input (Patil et al. 2011). Therefore, setting up a watershed model with physically meaningful parameters is crucial to accurate characterization of the watershed hydrology (Fonseca et al. 2014). Uncertainties are always involved in the estimation of the parameters (i.e. parameter uncertainty) since their direct measurements are not always possible (Gallagher and Doherty 2007). One of the major inputs for any hydrologic process simulation is topographic data. These data are inputted as raster layers, called Digital Elevation Models (DEMs), into watershed models to mathematically describe the topography of a drainage area. Extensive efforts have been made to understand the effect of DEM resolution on flow simulation though results remain mixed.

Cotter et al. (2003) and Chaubey et al. (2005) examined the resolution effect of resampled DEMs (30 to 1000 m) on the simulated flow for the Moores Creek watershed, Arkansas, with an average slope of 3.43 degrees. They found that coarsening the DEM resolution not only affected the delineation process in SWAT model but also led to the reduction in simulated flow. Based on flow simulation results from Hydrological Simulation Program-FORTRAN (HSPF) model for Yixun River basin in China, Wang et al. (2015) reported that simulated flows were substantially affected by the DEM resolution. Kienzle (2004) assessed DEM resolution effects on terrain derivatives for four study areas with various slopes in Alberta, Canada. He inferred that various resolutions of DEM were required based on the complexity of the intended derivative.

Chaplot (2005) indicated that the flow simulated using SWAT was not sensitive to DEM resolutions. Similar results were achieved by Zhang et al. (2014) for mountainous watersheds in China, and by Yang et al. (2014) for 3 watersheds in Idaho. Despite the significant progress and efforts put into modelling and understanding DEM resolution effects on hydrologic simulation, the results from previous studies are not consistent, calling for additional studies to understand the inconsistency. More importantly, the parameter uncertainties, which were introduced in the calibration process of HSPF model parameters, have been rarely analyzed and have not yet been evaluated with respect to DEM resolution uncertainties. Therefore, more studies are needed particularly on comparisons of calibrated models for multiple watersheds with different topographic conditions. Addressing the modelling need and particularly the uncertainties associated with the DEM resolution and HSPF model parameters under differing watershed conditions motivated this study.

The primary objective of this study was to identify and reduce the uncertainty caused by the DEM resolution by examining the DEM resolution effects on watershed-based flow simulation and comparing them with the effects of HSPF model parameter uncertainty. To that end, streamflow was simulated for three 10-digit hydrologic unit watersheds with different slopes using the HSPF model within the BASINS (Better Assessment Science Integrating Point and Nonpoint Sources) system. Effects of DEM resolution on the properties of delineated watersheds and the simulated flow were quantified and analyzed with regard to parameter uncertainties. Suggestions were provided for reducing the uncertainties based on watershed characteristics and data availability.

Materials and Methods

Study Area

Three watersheds of 10-digit hydrologic units were selected, including (1) Little North Santiam River watershed in Oregon (HUC 1709000505), (2) Wolf Creek watershed in Iowa (HUC 0708020508), and (3) Bayou Des Cannes watershed in Louisiana (HUC 0808020103), the United States (U.S.), as shown in Figure S1. These watersheds are respectively draining 285, 862, and 1049Footnote 1 km2 with average slopes of 22.2 degrees (41%), 2.27 degrees (4%), and 0.61 degrees (1%), respectively. Details of the data layers for the study areas, including land cover, 3.5 m DEM, subwatersheds, and STATSGO soil, are provided in Figures S2 to S4. Study areas were selected based on the availability of meteorological, streamflow and DEM data and more importantly the overall slope. Another important parameter that was taken into consideration in selection of the study areas is the size of the watershed. Specifically, 10-digit hydrologic unit watersheds with above mentioned criteria were chosen to facilitate breaking them into 12-digit hydrologic units in the manual delineation process. HSPF software is more likely to crash when larger study areas awith very fine DEMs (extremely large files) are used.

Data Collection

In order to identify effects of DEM resolution on simulated flows, DEMs with original resolutions of 1/9, 1/3 and 1 arc-second along with 100 m DEMs were downloaded from the United States Geological Survey (USGS) National Elevation Dataset (NED), https://nationalmap.gov/elevation.html, where the original DEMs refer to those datasets which were obtained at the mentioned resolutions without any resampling by the authors. The original DEMs were utilized from a single source to prevent the uncertainties from DEM resampling (Dixon and Earls 2009) and DEM sources (Roostaee and Deng 2018). DEMs with 1/9, 1/3 and 1 arc-second, which have the approximate horizontal resolutions of 3.5, 10.3 and 30.9 m at the equator, are, respectively, defined as 3.5, 10 and 30 m DEMs, hereafter.

Other datasets include: (1) STATSGO (State Soil Geographic) data layers mapped at the scale of 1:250000 (https://water.usgs.gov/lookup/getspatial?ussoils); (2) USGS National Land Cover Dataset 2006 represented at a spatial resolution of 30 m available at https://www.mrlc.gov/nlcd2006.php; (3) the National Hydrography Dataset (NHD) and Watershed Boundary Dataset (WBD) from https://nhd.usgs.gov/; (4) high temporal resolution (hourly) precipitation data obtained via BASINS for representative rain gauges within the watersheds: OR352292 (Detroit Dam) in Little North Santiam watershed, IA138296 (Toledo) in Wolf Creek watershed, and LA162981 (Eunice) in Bayou Des Cannes watershed. Required hourly meteorological data, such as temperature, solar radiation, cloud, potential evapotranspiration and wind, were obtained from BASINS for the closest weather stations with full sets of data: OR357500 (Salem AP McNary Field), IA725461 (Marshalltown Muni) and LA165021 (Lafayette), respectively; and (5) measured daily streamflow data for mainstream reaches in the selected watersheds which were obtained from USGS stations 14,182,500 (Little North Santiam River near Mehama), 05464220 (Wolf Creek near Dysart), and 8,010,000 (Bayou Des Cannes near Eunice).

BASINS and HSPF Model

BASINS is a multipurpose environmental analysis system integrating GIS, data analysis and modelling system designed to support watershed and water quality-based studies (https://www.epa.gov/exposure-assessment-models/basins). HSPF is one of the watershed modelling tools included in BASINS. It is a spatially distributed and temporally continuous watershed model. HSPF model has been widely used for Total Maximum Daily Load (TMDL) development for impaired waterbodies by local, regional, and state agencies due to its unique capabilities in simulation of in-stream hydraulic and sediment-chemical interactions along with overland flow and contamination processes, making it an ideal choice for hydrologic simulation and water quality assessment.

Watershed Delineation Methods

DEMs are commonly employed to delineate a watershed by creating a stream network either automatically or manually, deriving critical topographical characteristics of the watershed and channel geometry attributes. Two methods, including automatic and manual ones, were employed for watershed delineation. The eight directional flow algorithm (D-8) (Jenson and Domingue 1988), which is the primary algorithm available in BASINS/HSPF, was used in the automatic delineation process for determining flow directions. To be able to divide each of 10-digit hydrologic unit watershed into multiple subwatersheds at the scale of 12-digit hydrologic unit, yet maintaining the details, the threshold area in automatic delineation was set to 13 km2 in HSPF to meet the Federal Standards and Procedures for the National Watershed Boundary Dataset (WBD). Based on this standard, the typical size of a watershed at scale of 12-digit hydrologic unit was never less than 12.1 km2 (USGS 2013). The automatic delineation has become more and more practical due to the availability of high-resolution DEMs (Oksanen and Sarjakoski 2005) and has frequently been employed in watershed modelling. This method also is considerably sensitive to uncertainties originated from DEMs. The manual delineation, although it is less affected by DEMs, requires a more advanced skill in watershed delineation and thus it could be challenging to a watershed modeller due to complex watershed topography (Oksanen and Sarjakoski 2005), making more and more watershed modellers choose the automatic delineation. In the manual delineation, the size and boundary of a subwatershed and streams are manually defined by the user while other attributes are automatically extracted from the DEM. Therefore, the elevation, watershed and stream slopes are the only inputs which need to be assigned to the user-defined drainage network via manual delineation of any DEM. Based on the accuracy and resolution of the applied DEM, a user-defined network of streams may not be in compliance with the lowest path of the DEM, producing another source of errors in the stream slope definition. Hence, the manual delineation of a watershed with mild slope or complex network of streams may contribute high human errors. To avoid user errors in the manual delineation the 12-digit hydrologic unit layer and NHD flowlines layer were incorporated into the watershed model as pre-defined watershed boundaries and stream network layers in this study, as shown in Figures S2S4, for study watersheds. Although these datasets may not capture adequate details for specific applications, they serve the purpose of this study well. In this study, results from both the manual and automatic delineations were compared to test the impacts of two delineation methods on mitigating or propagating the DEM resolution-induced uncertainty.

DEM Resolution Effect on Flow Simulation

Various flow simulation scenarios were defined in order to determine the effects of DEM resolution on flow simulation and identify the methods for mitigating these effects. Specifically, four DEMs (including 3.5, 10, 30 and 100 m) were used in both automatic and manual delineations in BASINS/HSPF, creating 8 scenarios for each study watershed. To isolate the uncertainty in flow simulation produced by DEM resolution coarseness from other sources of uncertainties, such as parameter uncertainty, other simulation conditions must be kept constant. If each scenario is calibrated and validated separately, the differences between scenarios would be the results of a mixture of two different types of uncertainties including DEM resolution uncertainty and parameter uncertainty. Therefore, we calibrated the base scenario (the model with automatic delineation of 3.5 m DEM) for each study area and used the calibrated parameter sets in all other scenarios for the same watershed (10, 30 and 100 m DEMs under automatic calibration and 3.5, 10, 30 and 100 m DEMs under manual delineation).

Model Calibration and Validation

Flow calibration is an iterative process of adjusting hydrology-associated parameters in HSPF model. When certain level of agreement is achieved between trends and values of observed and simulated series of data, the model is considered calibrated. To test this agreement, various metrics have been used in previous studies (Moriasi et al. 2007).

In this study, three different statistical metrics, including coefficient of determination (R2), Percent BIAS (PBIAS), and Nash-Sutcliffe model Efficiency coefficient (NSE) (Nash and Sutcliffe 1970) (Eqs. 1, 2 and 3), were selected based on the guideline by Moriasi et al. (2015). The performance of HSPF model with NSE > 0.5, PBIAS<15% and R2 > 0.6 was considered satisfactory (Moriasi et al. 2015).

$$ {R}^2=\frac{\sum \limits_{i=1}^n\left(\left({y}_{sim}^i-{\overline{y}}_{sim}\right).\Big({y}_{obs}^i-{\overline{y}}_{obs}\right)}{\sum \limits_{i=1}^n{\left({y}_{sim}^i-{\overline{y}}_{sim}\right)}^2.\sum \limits_{i=1}^n{\left({y}_{obs}^i-{\overline{y}}_{obs}\right)}^2} $$
(1)
$$ PBIAS=\frac{\sum_{i=1}^n\left({y}_{sim}^i-{y}_{obs}^i\right)}{\sum_{i=1}^n\left({y}_{obs}^i\right)}\times 100 $$
(2)
$$ NSE=1-\frac{\sum \limits_{i=1}^n{\left({y}_{sim}^i-{y}_{obs}^i\right)}^2}{\sum \limits_{i=1}^n{\left({y}_{obs}^i-{\overline{y}}_{obs}\right)}^2} $$
(3)

where \( {y}_{obs}^i \) and \( {y}_{sim}^i \) are the observed monthly flow and simulated monthly flow for the ith month; the parameters, \( {\overline{y}}_{obs} \) and \( {\overline{y}}_{sim} \), respectively, denote the observed and simulated average streamflow in the simulation periods.

HSPF model input parameters, associated with runoff simulation, were calibrated and validated using the automatically delineated 3.5 m DEM for all the three selected watersheds. Daily streamflow data from previously mentioned USGS Stations were employed for the flow calibration (2002–2007, 5 years) and validation (2008–2009, 2 years), respectively. This time span was selected since this is the only period in which observed data were available for all the three study watersheds.

An automated method was used for the flow calibration of HSPF model to avoid subjectivity. Specifically, 5000 samples were taken using Monte Carlo sampling method from a feasible range (Bicknell 2000) for each of the most sensitive water balance parameters of HSPF model reported in Fonseca et al. (2014). Then, 5000 HSPF models for each study watershed were generated and run using the randomly sampled parameter sets. The parameter set, which led to the model meeting the criteria suggested by Moriasi et al. (2015), was considered as the calibrated parameter set. The scenario with 3.5 m resolution DEM was used for calibration since this DEM has the highest available resolution.

The effects of DEM resolution on watershed hydrology were quantified by comparing observed flow in each study watershed with the HSPF model simulated flows (8 scenarios) for that watershed using the statistical metrics including mean, PBIAS and Mean Absolute Error (MAE) (Moriasi et al. 2015).

$$ MAE=\frac{1}{n}\sum \limits_{i=1}^n\mid {y}_{sim}^i-{y}_{obs}^i\operatorname{}\mid $$
(4)

Model Recalibration and Parameter Uncertainty Analysis

Recalibration of the HSPF model parameters can be considered as a means to alleviate the adverse impact of DEM coarseness. However, equally acceptable simulations from several distinct parameter sets are likely to be produced by hydrological models (Hope et al. 2004) and it is highly unlikely that a single set of “best” parameters can be identified (Fonseca et al. 2014). Therefore, the estimation of the parameter uncertainty and evaluating its effect (offsetting or aggravating) regarding the DEM resolution-induced uncertainty is necessary. To that end, the Generalized Likelihood Uncertainty Estimation (GLUE) method (Beven and Binely 1992) was used for uncertainty assessment of water balance parameters in this study. Uncertainties originated from parameters were then compared to those from DEM resolution.

Several steps were included in the application of GLUE method: (1) Determination of a feasible parameter range: Fonseca et al. (2014) listed and ranked the most sensitive water balance parameters of HSPF model, which were reported in the literature. The top five sensitive water balance parameters (INFILT, LZSN, AGWRC, UZSN and DEEPFR)Footnote 2 were employed in the uncertainty estimation using GLUE framework in this study. Possible value ranges of the parameters were obtained from BASINS Technical Note 6 (Bicknell 2000); (2) Determination of prior distributions of parameters: A uniform distribution was employed for all parameters due to lack of sufficient information about the prior distributions (Gong et al. 2011); (3) Monte Carlo sampling from a possible parameter range: 5000 sets of parameters were randomly selected from the specified parameter ranges and then HSPF model was run using each of these sets; (4) Specification of the likelihood measure and threshold: NSE was used in this study as the likelihood measure. Specification of the likelihood measure and threshold to divide the parameter sets into behavioral and non-behavioral is subjective (Freer et al. 1996); the lower the threshold is set, the wider are the ranges of parameters (Arabi et al. 2007). In this study, the threshold value for NSE was set to 0.6; (5) Calculation of likelihood measures for all randomly selected parameter sets; and (6) Incorporation of all behavioral parameter sets in order to derive uncertainty bands of 97.5% and 2.5% (95% confidence interval) for annual and monthly flow prediction using HSPF model.

Results from this method for parameter uncertainty analysis can also be utilized to test the impacts that re-calibration of the model may have on alleviating the negative effects of DEM coarseness since various parameters are implemented in model each time (for 5000 times) to cover almost every possible combination of parameter values. In this study, the parameter uncertainties for the HSPF models developed with 3.5 m and 100 m DEMs (the finest and the coarsest resolutions, respectively) were estimated and compared. HSPF model with 100 m DEM was also re-calibrated for each study watershed to test the re-calibration benefits.

Results and Discussion

Effects of DEM resolution on watershed-scale flow simulation are introduced in two major modelling steps including watershed delineation and flow simulation. Specifically, DEM resolution influences derived topographic attributes of watershed through watershed delineation and then the topographic attributes affect simulated flow through watershed-scale flow simulation.

Effects of DEM Resolution on Watershed Delineation in Terms of Derived Topographic Attributes

Automatic Delineation

Effects of DEM resolution on watershed delineation are described using the changes, caused by DEM resolution variations, in stream network properties (particularly channel length, depth, width and slope), drainage areas, average slopes and average elevations of automatically delineated watersheds, as shown in Table S1. Descending trends in drainage area are detected as the DEM resolution coarsens, although not all coarse DEMs lead to smaller watershed areas. The drainage area of the Little North Santiam watershed delineated with 3.5 m DEM is slightly smaller than that delineated with 10 m DEM. The results on descending trends in the drainage area are consistent with those of previous studies (Cotter et al. 2003; Wang et al. 2015). The average slopes of overland flow path (known as SLSUR) of automatically delineated watersheds also display similar descending trends with coarsening DEM resolutions while the elevations tend to be higher, as seen in Table S1. Greater reductions are seen in flat watersheds compared to other study areas. The maximum reduction in the delineated contributing drainage area in Bayou Des Cannes watershed (slope = 1%) is 55% while the maximum drainage area reduction in other two watersheds with steeper gradients is less than 5%. The reason behind the reduction in the delineated drainage area could be attributed to the inaccurate delineation of watershed due to the inaccurate elevation data derived from coarser DEMs, leading to the failure of flow accumulation algorithms and thereby the definition of disconnected streams and non-contributing areas. These algorithms were employed in the automatic delineation of the network of streams and subsequently their associated drainage area based on the difference in elevations of DEM cells/pixels.

There is a strong correlation between SLSUR reduction and DEM resolution for Little North Santiam, Wolf Creek and Bayou Des Cannes with the corresponding R2 values of 0.97, 0.88 and 0.71, respectively. There is also a strong correlation between the DEM resolution and the length of a defined stream, as indicated by the high R2 values of 0.94, 0.99, and 0.95 for Little North Santiam, Wolf Creek and Bayou Des Cannes, respectively. The maximum reductions in the channel length are 10.2, 23.9 and 52.1 km, respectively, when coarser DEMs are applied in the delineation of the three watersheds (Table S1). The total lengths of delineated streams are generally shorter when lower resolution DEMs are used in watershed delineation (up to 58% reduction in length). In addition to the DEM resolution, the channel length defined in automatic delineation of a DEM is also affected by the threshold used for channel/watershed delineation. In this study the effect of this threshold is eliminated by setting a fixed threshold value for all scenarios.

The average depths of stream channels are negligibly affected by the DEM resolution while fluctuations can be seen in the average stream width due to changes in stream lengths and elevations, as shown in Table S1. In BASINS, a simplified stream geometry is employed through FTABLEs (Hydraulic Function Tables) which are independent of the shape of the waterbodies (USEPA 2007). Therefore, the depth and width of stream are not major determinants of the flow simulated by HSPF model. It is clear from the above-mentioned results that the extracted attributes of the watersheds in flat areas are sensitive to the DEM resolution. However, the DEM resolution hardly exerts noticeable effects on attributes of delineated watersheds in mountainous areas with steep slopes (Table S1). It should be pointed out that these results may be applicable to watersheds with similar characteristics and are not necessarily applicable to substantially different watersheds.

Manual Delineation

The effects of DEM resolution on topographic attributes of the manually delineated watersheds are also shown in Table S1. In spite of the marked change in the DEM resolution from 3.5 to 100 m, the average slope of channel only experiences a slight change of 0.09%, 0.009% and 0.005% in Little North Santiam watershed, Wolf Creek watershed and Bayou Des Cannes watershed, respectively. It is clear from the table that the effects of DEM resolution on attributes of the manually delineated watersheds are negligible as compared to those on automatically delineated watersheds. However, the attributes of the watersheds delineated manually and automatically are significantly different particularly in the watershed slope and area,that affect flow simulation. There are significant differences between the SLSUR values derived by HSPF model with manual and automatic delineation methods. Calculated SLSUR values (%) vary from 48.1 to 41.73, 4.19 to 2.28 and 1.05 to 0.37 with changing DEM resolutions according to the automatic delineation of Little North Santiam, Wolf Creek and Bayou Des Cannes watersheds, respectively. Corresponding SLSUR ranges for manual delineation are 20.75 to 21.36%, 3.78 to 2.47% and 0.94 to 0.2%, respectively (Table S1). The significance of this finding is that the calibrated parameter values for a watershed model delineated with manual and automatic methods could be significantly different.

Effects of DEM Resolution on Monthly Flow Simulation

Effects of DEM Resolution on Flow Calibration and Validation

The time series of simulated monthly flow for calibration and validation periods are displayed and compared to measured flow for each study watershed (Fig. 1). As mentioned earlier, more than one sets of parameters, which result in acceptable (calibrated) hydrological simulations, can be found. Therefore, the final model presented here, called “calibrated” hereafter, is the modeller’s choice.

Fig. 1
figure1

Comparisons between the monthly flows observed and simulated with HSPF model in Little North Santiam River (a), Wolf Creek (b), and Bayou Des Cannes (c) watersheds for the calibration and validation periods

For calibration of the Little North Santiam River watershed in Oregon, the values of R2, NSE and PBIAS were 0.93, 0.92 and − 6.1%, respectively, while the corresponding values of these metrics were 0.77, 0.76 and − 3.5% for Wolf Creek watershed and 0.97, 0.9 and 13.9% for Bayou Des Cannes watershed, respectively (Table S2). The statistical metrics for the validation period are presented in Table S2. Although magnitudes of these metrics differ from the calibration period, model performance is still satisfactory. Overall, the model-simulated flows (particularly high flows) fit observed ones well in terms of timing and variation trends, suggesting the acceptable performance of the watershed models in flow simulation.

Effects of DEM Resolution on Performance Metrics of Watershed Models

To differentiate the parameter-induced uncertainties in flow simulation from those produced by DEM resolution coarseness, we applied the parameter sets of the calibrated model from the 3.5 m DEM and automatic delineation to all other scenarios (10, 30 and 100 m DEMs under automatic calibration and 3.5, 10, 30 and 100 m DEMs under manual delineation). Table S3 presents the performance metrics calculated for various scenarios.

Table S3 illustrates that the R2 value barely responds to any changes in the DEM resolution (maximum R2 change is 0.01 when the resolution of DEM is lowered from 3.5 to 100 m) since this parameter describes the agreement between the trends and disregards the magnitudes. It appears that NSE and PBIAS are better metrics for identifying the effects of the DEM resolution on flow simulation as they are sensitive not only to DEM resolutions but also to other characteristics of individual scenarios. Table S3 indicates that (1) the changes in DEM resolution exert stronger effects on the automatically delineated watersheds than manually delineated watersheds in terms of the model performance metrics (NSE, R2 and PBIAS); (2) the DEM resolution effects are more important to watersheds with a mild slope, as indicated by over 40% change in PBIAS and the change of 0.31 in NSE when the DEM resolution is lowered from 3.5 to 100 m for the Bayou Des Cannes watershed with the smallest slope of 1%. The corresponding changes in the Little North Santiam River watershed with the steepest slope of 41% are smallest.

The results clearly demonstrate that the performance of watershed models in simulation of streamflow is highly affected by the DEM resolution even if calibrated parameters are employed. Manual delineation can improve model performance in most of the scenarios, as confirmed by the lower PBIAS values. The greatest improvement is seen in the results for the Bayou Des Cannes watershed, where the average NSE value increases from 0.61 to 0.74 and the PBIAS decreases from −50.48% to 10.19% for calibration period. More detailed effects of the DEM resolution on flow simulation are discussed in following subsections.

Effects of DEM Resolution on Mean Flow Simulation

Effects of DEM resolution on simulated monthly average streamflow at or near the outlets of the three watersheds, delineated using both the automatic and the manual methods, are described using the statistical measures and summarized in Table S4. These results are provided for the 8-year period of 2002 to 2009 (including both calibration and validation) in order to cover both high and low flows (particularly for Wolf Creek watershed). Therefore, this table contains magnitudes different from Table S3.

Automatic Delineation

Monthly average flow values, simulated with HSPF model for Little North Santiam watershed using the DEM data of four different resolutions, are in the range of 17.46–17.05 m3/s, showing a strong correlation of 0.98 with the DEM resolution (Fig. 2). Overall, a slightly decreasing trend in the simulated flow is detected as the DEM resolution becomes coarser, as expressed by decreasing mean values of simulated flow and declining PBIAS values from −8.78% (3.5 m DEM) to −10.92% (100 m DEM) (Table S4 and Fig. 2). The simulated mean flow for the Wolf Creek watershed decreases from 7.17 m3/s to 6.91 m3/s when the DEM resolution is lowered from 3.5 to 100 m (Fig. 2a). The PBIAS value also varies in the range of −0.6% (for 3.5 m DEM) to −4.14% (for 100 m DEM). It can be seen from Table S4 that the simulated flow in the Bayou Des Cannes watershed is strongly affected by the 100 m DEM since PBIAS value sharply drops from 2.71 to −48.47% as the DEM resolution becomes coarser from 3.5 to 100 m.

Fig. 2
figure2

DEM resolution effects on monthly flow simulated with automatically delineated watersheds. The vertical axis denotes monthly mean flow

The results from Table S4 and Fig. 2 suggest that the flow simulated with watershed models is sensitive to the resolution of DEM data. Specifically, the statistical metrics (mean and PBIAS) describing the performance of watershed models in flow simulation are strongly dependent on the resolution of DEM data particularly for flat watersheds. Strong correlations (R2 = 0.95–0.99) are found between PBIAS values and DEM resolutions for all automatically delineated scenarios, confirming the significance of the DEM resolution to watershed-based flow simulation. Negligible differences are seen among the results from 3.5, 10 and 30 m DEMs. Therefore, due to the limited availability of 3.5 m DEM and the intensive computation resource required for 10 m DEM, 30 m DEM could be considered as the optimum DEM resolution for watershed-scale flow simulation.

Manual Delineation

The manual delineation method reduces the difference in simulated flows to less than 0.01 m3/s among different scenarios for a watershed and the change of PBIAS values to less than 0.1%, as shown in Table S4. The manual delineation method is indispensable for flat areas to maintain the extent of the watershed and to prevent the loss of accuracy in watershed modelling. Although the DEM resolution has a negligible effect on average flow simulation, its effects on flow hydrograph and the simulation of slope-associated water quality parameters, such as sediment, may be significant because of the impact of DEM resolution on derived stream slope and should be further investigated.

A comparison of the PBIAS values (Table S4) for individual watersheds constructed using both the manual and the automatic delineation methods demonstrates that the dominant factor responsible for the flow reduction with coarsening DEM resolution is the reduction in the contributing drainage area caused by the automatic delineation (Table S1). This finding is supported by the near perfect correlations (greater than 0.99) between the PBIAS values calculated using HSPF model outputs and the watershed areas delineated using different DEM resolutions for all the three watersheds with diverse topographic characteristics, as shown in Fig. 3.

Fig. 3
figure3

Correlation between automatically delineated areas and calculated PBIAS in Little North Santiam (a), Wolf Creek (b), and Bayou Des Cannes (c)

While the underestimation of simulated flow due to the use of low resolution DEM data has long been reported (Cotter et al. 2003), the flow reduction was generally attributed to the smoothing effect of decreasing the DEM resolution (Chaubey et al. 2005), errors in the estimation of mean slope at coarser DEMs (Kalin et al. 2003), and the loss of topographic details (Wang et al. 2015). Results from this study indicate that a smoother slope is not always associated with a lower simulated flow (Tables S1, S3 and S4), as evidenced by the slopes calculated from 3.5 and 10 m DEMs for Little North Santiam and Bayou Des Cannes, and their corresponding simulated flows. Therefore, the reason behind the reduction in the drainage area and the simulated flow could be attributed to the inaccurate elevation data represented by coarser DEMs that led to the definition of disconnected streams and non-contributing areas.

The new finding from this study clarifies the long-standing inconsistency in the sensitivity of the hydrologic simulation to DEM resolution. That is, watershed gradient is the key factor in how the watershed responds to lowering DEM resolution which was not reported in previous studies. Based on the results presented in previous sub-sections, the flow simulated for the watershed with steep slope is not sensitive to DEM resolution (in agreement with results from previous studies (Zhang et al. 2014)), while for flat watersheds, moderate to high sensitivity is expected as evidenced by results from this study and reported by Wang et al. (2015). Therefore, these results provide a theoretical basis for future watershed modelling to choose the proper DEM resolution based on the gradient of study watershed in order to prevent the high uncertainty induced by the adoption of low resolution DEM data.

Effects of Model Recalibration and Parameter Uncertainty on Resolution-Induced Uncertainty

The effects of DEM resolution on flow simulation presented in previous sections suggested small differences among the results of the scenarios with the automatic delineation of 3.5, 10 and 30 m DEMs while the highest effect of DEM resolution was seen for 100 m DEM under the automatic delineation. DEM resolution was found to have negligible effects on simulated flow when the manual watershed delineation was conducted. Therefore, the greatest discrepancy was observed between the results from the two scenarios corresponding to 3.5 m and 100 m DEMs. In this section results from the best and worst case scenarios (i.e. automatically delineated 3.5 and 100 m DEMs) are used to analyze the parameter uncertainty and benefits of model recalibration.

Figure S6 illustrates the uncertainties, induced by both the model input parameters and the DEM resolutions, in simulated flow for the automatic delineation method. The widths of the uncertainty bands represent the parameter-induced uncertainties while the overall distance (spacing) between the red and green dash lines representing simulated flows (using 3.5 and 100 m DEMs, respectively, the finest and the coarsest DEMs used in this study) denotes the uncertainty caused by the DEM resolution.

It is clear from Figure S6 that the variations in water balance parameters within their feasible ranges affected the flow simulation and produced a range of behavioral results. For Little North Santiam and Wolf Creek watersheds, uncertainty bands for 3.5 and 100 m DEMs overlap each other with similar band widths. Therefore, similar results can be achieved from any of these models by making adjustments to the water-balance parameters. Results from the model recalibrations are shown in Figure S7. Average widths of 95% confidence intervals are 4.18 and 4.05 m3/s for models with 3.5 and 100 m DEM, respectively, for Little North Santiam watershed. The corresponding values for Wolf Creek watershed are 3.12 and 3.98 m3/s, respectively. However, for Bayou Des Cannes watershed the 95% uncertainty band, estimated from 100 m DEM, has no overlap with the band from 3.5 m DEM, as shown in Figure S6(c). Therefore, none of the 5000 tested parameter sets could eliminate the uncertainty induced by the DEM resolution no matter how the model is calibrated or recalibrated (Figures S6 and S7). In other words, the uncertainty caused by the DEM resolution is much higher than the HSPF model parameter-induced uncertainty and the adjustment of water balance parameters would not make up for that even with extremely unreasonable parameter values (such as an infiltration rate of zero). Additionally, it is shown in Figure S6 that the uncertainty band for 100 m DEM is narrower than the band for 3.5 m DEM (average width of 1.3 versus 2.42 m3/s, respectively), indicating lower flexibility of 100 m DEM-based model for producing behavioral results. In case of the 3.5 m DEM the maximum bandwidth/uncertainty caused by the model input parameters (maximum thickness) in the simulated mean annual flow is 3.28 m3/s that is higher than the uncertainty caused by the 3.5 m DEM resolution (distance between 3.5 m simulated flow and the observed flow). In case of the 100 m DEM, however, the maximum bandwidth/uncertainty caused by the model input parameters in the simulated mean annual flow is much smaller than the uncertainty caused by the 100 m DEM resolution. Specifically, the average simulated flow of the band representing the 100 m DEM is 4.93 m3/s that is about 40% lower than the observed flow. While the simulated average flow of 4.93 m3/s could be increased to 5.59 m3/s by the model recalibration, the recalibrated flow of 5.59 m3/s is still 30% lower than the observed flow (Figure S7). Therefore, the uncertainty caused by low resolution DEMs in simulated flow could be significantly higher than the uncertainty caused by model input parameters even if model parameters are recalibrated. In order to reduce the uncertainty caused by low DEM resolutions, it is recommended that DEMs of the 30 m or finer resolution be employed in watershed-scale flow simulation particularly for watersheds in flat areas following the procedure, shown in Figure S5, for reducing DEM resolution-induced uncertainties. For watersheds with higher gradients, adverse impacts of coarser DEMs can be offset by adjusting the water balance parameters (Figure S7).

Conclusions

The following conclusions can be drawn from this study:

  1. (1)

    Watershed delineation methods may significantly affect flow simulation by changing the derived watershed attributes (particularly average slope (SLSUR) and drainage area) controlling the flow generation. Watershed-based flow simulation is sensitive to the DEM resolution if the watershed is constructed using the automatic delineation method. Watershed attributes derived from an automatically delineated watershed depend heavily on the DEM resolution. The coarser is the DEM resolution, the greater is the reduction in the delineated drainage area (up to 55%) and the channel length. Therefore, the flow reduction with decreasing DEM resolution is primarily due to the reduction in the delineated drainage area. The effect of DEM resolution on flow simulation is negligible when the manual delineation method is used.

  2. (2)

    The watershed gradient determines whether and how the DEM resolution affects the simulated flow. The milder is the overall slope of the watershed, the greater is the reduction in the delineated drainage area (higher possibility of producing non-contributing areas by DEMs) and thus in the simulated flow with coarsening DEM resolution. Simulated flow in watersheds with flat to mild slopes is sensitive to DEM resolution. Specifically, the drainage area derived from a flat watershed decreases markedly when the DEM resolution changes from 3.5 to 100 m, causing a significant reduction (up to 50% change in RE) in simulated average monthly flow. Simulated flow in a mountainous area with a steep slope displays only a slightly decreasing trend characterized by up to 4% change in RE, which is negligible as compared to the 50% change in the flat watershed, when the DEM resolution changes from 3.5 to 100 m.

  3. (3)

    The DEM resolution-induced uncertainty in simulated flow for the flat watershed delineated with a low resolution DEM is greater than the parameter-induced uncertainty even if the model input parameters are recalibrated. For watersheds with steep slopes, however, parameter uncertainty is substantially greater than the resolution-induced uncertainty. Therefore, the investigation of parameter uncertainty is more important than the assessment of DEM resolution impact for watersheds with steep slopes.

  4. (4)

    The effects of DEM resolution on watershed-scale flow simulation can be minimized by using DEMs of 30 m or finer resolution for flat watershed. For watersheds with moderate to steep slopes the calibration of water-balance parameters can offset the adverse impact of DEM coarseness.

  5. (5)

    The findings from this study clarify the effects of DEM resolutions on watershed-scale flow simulation in terms of how watershed delineation methods, watershed gradients, and DEM resolutions affect derived watershed attributes and thereby simulated flow, providing guidelines for watershed modelling and particularly reducing the uncertainties in simulated flow.

Notes

  1. 1.

    The drainage area to the gage station (Bayou Des Cannes near Eunice) is 339 km2.

  2. 2.

    INFILT = infiltration, LZSN = lower zone storage nominal, AGWRC = groundwater recession rate, UZSN = upper zone storage nominal, DEEPER = fraction of groundwater inflow to deep recharge

References

  1. Arabi M, Govindaraju RS, Hantush MM (2007) A probabilistic approach for analysis of uncertainty in the evaluation of watershed management practices. J Hydrol 333(2–4):459–471

    Article  Google Scholar 

  2. Beven K, Binely M (1992) The future of distributed model: model calibration and uncertainty prediction. Hydrol Process 6:279–298

    Article  Google Scholar 

  3. Bicknell BR (2000) Basins technical note 6: estimating hydrology and hydraulic parameters for HSPF. Environmental Protection Agency, US

    Google Scholar 

  4. Chaplot V (2005) Impact of DEM mesh size and soil map scale on SWAT runoff, sediment, and NO 3–N loads predictions. J Hydrol 312:207–222

    Article  Google Scholar 

  5. Chaubey I, Cotter AS, Costello TA, Soerens TS (2005) Effect of DEM data resolution on SWAT output uncertainty. Hydrol Process 19(3):621–628

    Article  Google Scholar 

  6. Cotter AS, Chaubey I, Costello TA, Soerens TS, Nelson MA (2003) Water quality model output uncertainty as affected by spatial resolution of input data. J Am Water Resour Assoc 39:977–986

    Article  Google Scholar 

  7. Dixon B, Earls J (2009) Resample or not?! Effects of resolution of DEMs in watershed modeling. Hydrol Process 23(12):1714–1724

    Article  Google Scholar 

  8. Fonseca A, Ames DP, Yang P, Botelho C, Boaventura R, Vilar V (2014) Watershed model parameter estimation and uncertainty in data-limited environments. Environ Model Softw 51:84–93

    Article  Google Scholar 

  9. Freer J, Beven K, Ambroise B (1996) Bayesian estimation of uncertainty in runoff prediction and the value of data: an application of the GLUE approach. Water Resour Res 32:2161–2173

    Article  Google Scholar 

  10. Gallagher M, Doherty J (2007) Parameter estimation and uncertainty analysis for a watershed model. Environ Model Softw 22:1000–1020

    Article  Google Scholar 

  11. Gong Y, Shen Z, Hong Q, Liu R, Liao Q (2011) Parameter uncertainty analysis in watershed total phosphorus modeling using the GLUE methodology. Agric Ecosyst Environ 142:246–255

    Article  Google Scholar 

  12. Hope A, Stein A, McMichael C (2004) Uncertainty in monthly river discharge predictions in a semi-arid Shrubland catchment. BHS 284–290

  13. Jenson SK, Domingue JO (1988) Extracting topographic structure from digital elevation data for geographic information system analysis. Photogramm Eng Remote Sens 54:1593–1600

    Google Scholar 

  14. Kalin L, Govindaraju RS, Hantush MM (2003) Effect of geomorphologic resolution on modeling of runoff hydrograph and sedimentograph over small watersheds. J Hydrol 276:89–111

    Article  Google Scholar 

  15. Kienzle S (2004) The effect of DEM raster resolution on first order, second order and compound terrain derivatives. Trans GIS 8:83–111

    Article  Google Scholar 

  16. Moriasi D et al (2007) Model evaluation guidelines for systematic quantification of accuracy in watershed simulations. Trans ASABE 50:885–900

    Article  Google Scholar 

  17. Moriasi D et al (2015) Hydrologic and water quality models: performance measures and evaluation criteria. Trans ASABE 58:1763–1785

    Article  Google Scholar 

  18. Nash JE, Sutcliffe JV (1970) River flow forecasting through conceptual models part I — a discussion of principles. J Hydrol 10:282–290

    Article  Google Scholar 

  19. Oksanen J, Sarjakoski T (2005) Error propagation analysis of DEM-based drainage basin delineation. Int J Remote Sens 26:3085–3102

    Article  Google Scholar 

  20. Patil A, Deng Z, Malone RF (2011) Input data resolution-induced uncertainty in watershed modelling. Hydrol Process 25:2302–2312

    Article  Google Scholar 

  21. Roostaee M, Deng Z (2018) Uncertainty analysis of watershed-based flow and water quality modelling with different DEM data sources. Ed HIC 2018. 13th international conference on hydroinformatics, 1778–1786

  22. USEPA 2007. BASINS technical note 1: creating hydraulic function tables (FTABLES) for reservoirs in BASINS

    Google Scholar 

  23. USGS (2013) Federal standards and procedures for the National Watershed Boundary Dataset (WBD). Ed 2013

  24. Wang H, Wu Z, Hu C (2015) A comprehensive study of the effect of input data on hydrology and non-point source pollution modeling. Water Resour Manag 29:1505–1521

    Article  Google Scholar 

  25. Yang P, Ames DP, Fonseca A, Anderson D, Shrestha R, Glenn NF, Cao Y (2014) What is the effect of LiDAR-derived DEM resolution on large-scale watershed model results? Environ Model Softw 58:48–57

    Article  Google Scholar 

  26. Zhang P, Liu R, Bao Y, Wang J, Yu W, Shen Z (2014) Uncertainty of SWAT model at different DEM resolutions in a large mountainous watershed. Water Res 53:132–144

    Article  Google Scholar 

Download references

Funding

This research was supported by the United States Geological Survey through Louisiana Water Resources Research Institute.

Author information

Affiliations

Authors

Corresponding author

Correspondence to Zhiqiang Deng.

Ethics declarations

Conflict of Interest

There are no known conflicts of interest associated with this publication.

Availability of Data and Codes

Data and models will be available upon request.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

ESM 1

(DOCX 2.82 kb)

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Roostaee, M., Deng, Z. Effects of Digital Elevation Model Resolution on Watershed-Based Hydrologic Simulation. Water Resour Manage 34, 2433–2447 (2020). https://doi.org/10.1007/s11269-020-02561-0

Download citation

Keywords

  • Digital elevation model resolution
  • Flow simulation
  • Watershed modelling
  • Uncertainty analysis