Sie sind auf Seite 1von 13

Stoch Environ Res Risk Assess (2013) 27:209–221

DOI 10.1007/s00477-012-0577-x

ORIGINAL PAPER

Evaluating DEM source and resolution uncertainties in the Soil


and Water Assessment Tool
Shengpan Lin • Changwei Jing • Neil A. Coles •

Vincent Chaplot • Nathan J. Moore •


Jiaping Wu

Published online: 1 March 2012


Ó Springer-Verlag 2012

Abstract DEMs as important input parameters of envi- simulations even it was provided with a smaller grid size and
ronmental risk assessment models are notable sources of higher vertical accuracy. The predicted outputs based on
uncertainties. To illustrate the effect of DEM grid size and ASTER GDEM and SRTM were similar, and much lower
source on model outputs, a widely used watershed manage- than the ones based on DLG. This study presents potential
ment model, the Soil and Water Assessment Tool (SWAT), uncertainties introduced by DEM resolutions and data
was applied with two newly available DEMs as inputs (i.e. sources, and recommends strategies choosing DEMs based
ASTER GDEM Version 1, and SRTM Version 4.1). A DEM on research objects and maximum acceptable errors.
derived from 1:10,000 high resolution digital line graph
(DLG) was used as a baseline for comparisons. Eleven Keywords DEM resolution  SWAT 
resample resolutions, from 5 to 140 m, were considered to Model uncertainty  ASTER GDEM  SRTM
evaluate the impact of DEM resolution on SWAT outputs.
Results from a case study in South-eastern China indicate
that the SWAT predictions of total phosphorus and total 1 Introduction
nitrogen decreased substantially with coarser resample
resolution. A slightly decreasing trend was found in the Quantitative models have proven effective for managing
SWAT predicted sediment when DEMs were resampled to risks for over half a century, and have been and are being
coarser resolutions. The SWAT predicted runoff was not applied in environmental risk management contexts world-
sensitive to resample resolution. For different data sources, wide. To prepare for better management in environment, we
ASTER GDEM did not perform better than SRTM in SWAT need fully understand the model uncertainty. Considerable
progress has been seen in the quantification of environmental
model uncertainty in last few decades, and the uncertainty in
S. Lin  C. Jing  N. J. Moore  J. Wu (&)
College of Environment and Natural Resources, model parameters has been emphasised (e.g. Maddalena
Zhejiang University, Hangzhou 310058, China et al. 2001; Tarantola et al. 2002; Wagener and Gupta 2005;
e-mail: jw67@zju.edu.cn Skeffington 2006). DEM-based modelling of watershed
processes has become increasingly popular in environmental
N. A. Coles
Centre of Excellence for Ecohydrology, University of Western research, mainly due to the advances in availability and
Australia, Crawley, WA 6009, Australia efficiency of DEMs in acquiring topographic properties of
watersheds, such as slope, field slope length, and channel
V. Chaplot
network. In recent years, many new methods for generating
IRD-BIOEMCO School of Bioresources Engineering
and Environmental Hydrology, University of KwaZulu-Natal, DEMs of the Earth’s surface have become available, with the
Scottsville 3209, South Africa release of the space-borne shuttle radar topography mission
(SRTM) and advanced space-borne thermal emission and
N. J. Moore
reflection radiometer (ASTER) elevation data sets. The
Department of Geography, Center for Global Change and Earth
Observations, Michigan State University, East Lansing, DEM data from these two space missions cover most of the
MI 48823, USA populated regions of the world and are publicly and freely

123
210 Stoch Environ Res Risk Assess (2013) 27:209–221

available at a spatial resolution of 3 arc-seconds for SRTM DEMs were resampled to a wide range of resolutions
and 1 arc-second for ASTER global DEM (ASTER GDEM). (5, 10, 20, 30, 40, 60, 80, 90, 100, 120, and 140 m) to
With an increasing availability of DEM data from remote evaluate the resolution impact on model outputs.
sensing sources, the problem of data availability appears to
be less important nowadays than a decade ago. However, it 1.1 Study area
remains difficult for model users to select appropriate DEM
for a specific application. Moreover, the risk of data misuse The study area locates in the Xiekengxi River watershed,
increases with more data sources and resample tools. an 81.7 km2 watershed in Lin’an City, Zhejiang Province,
Therefore, the objective of this paper is to assess the impacts South-eastern China (Fig. 1). Mountains dominate most
of different data sources and different resample resolutions parts of the watershed. Based on DLG 5 m data, the ele-
of input DEMs on environment models. vation of the watershed is 713.4 ± 235.5 m. The major
The comparison DEMs were from three different data land use within the watershed is mixed forest, accounting
sets: digital line graph (DLG, 1:10,000 DLG based on local for 96.0% of total area. Within the watershed, there are six
field survey, 5 m interval, named as DLG 5 m hereafter); soil types: 61.1% is Mountain yellowish clayey soil with
ASTER GDEM (1 arc-second ASTER GDEM Version 1, moderate infiltration rate; 36.6% is yellowish red soil with
approximately 30 m resolution, named as ASTER 30 m high infiltration rate; the remaining four soil types account
hereafter); and SRTM (3 arc-second SRTM Version 4.1, for only 2.2% of the total watershed area (Wei 1993). The
approximately 90 m resolution, named as SRTM 90 m Xiekengxi River watershed is a subbasin of the Tiaoxi
hereafter). In China, high resolution DLG is used as stan- River watershed. The Tiaoxi River flows to the Taihu Lake,
dard data and was therefore used as a reference data set in which is the third largest freshwater lake in China and was
this study. We focused on ASTER 30 m and SRTM 90 m, noted world-wide for the eutrophication incident in sum-
because they were newly available in 2009, easily acces- mer 2007, when millions of people, whose drinking water
sible, free of charge, and probably the most useful global was supplied from the Taihu Lake, had no potable water
data sources for regional and global model applications. All for several weeks.

Fig. 1 a Location of the


Xiekengxi River watershed
(81.7 km2), b estimated stream
networks derived from DEMs of
different sources (all of which
are in 5 m resample resolution
for comparisons), c stream
networks zooming in on the
rectangle of (b), topographic
maps derived from original
resolution DEMs: d DLG 5 m,
e ASTER 30 m, and f SRTM
90 m

123
Stoch Environ Res Risk Assess (2013) 27:209–221 211

This paper presents a preliminary study of the Tiaoxi methods for estimating surface runoff volume: the SCS
River watershed simulation that seeks to tackle the ques- curve number procedure (USDA-SCS 1972) and the Green
tion of which DEM is suitable for multiple temporal and and Ampt (1911). The erosion caused by rainfall and surface
spatial process analyses. The Xiekengxi River watershed as runoff is computed with the modified universal soil loss
a small subbasin was selected as it had regionally repre- equation (MUSLE) (Williams 1975). The maximum amount
sentative landscapes of bamboo forestry and nurseries, of sediment that can be transported from a ‘‘reach segment’’
which have high cash returns, cover about 20% area of the is a function of peak channel velocity. The soluble N and P in
Tiaoxi River watershed. Farmers may apply up to five runoff are estimated as products of the volume of water and
times as much fertilizers in the bamboo forest areas as used average concentrations of N and P in soil layers. Sediment
for regular paddy rice fields. Unlike the paddy fields, which transport of N and P is calculated with a loading function
are flat and bunded along the field boundary, these two land developed by McElroy et al. (1976) and modified by Wil-
use/cover types are distributed in relatively steep, hilly liams and Hann (1978) for applications to individual runoff
areas and subject to soil erosion. Therefore, they are con- events based on the sediment yield in the HRUs. A complete
sidered as potentially major sources of non-point source description of the SWAT model can be found in Arnold et al.
pollutions–namely nutrients and sediments. (1998) and Neitsch et al. (2005).
Several studies have been performed to analyse the sen-
1.2 Soil and Water Assessment Tool (SWAT) model sitivity of the SWAT model outputs on DEM resolutions
(Cotter et al. 2003; Chaplot 2005; Di Luzio et al. 2005; Dixon
There have been many studies on the effect of grid reso- and Earls 2009). Cotter et al. (2003) and Chaubey et al.
lution on terrain derivatives and environmental models, (2005)evaluated the impact of resample resolution of DEM
especially TOPMODEL (e.g. Zhang and Montgomery (at 30, 100, 150, 200, 300, 500, and 1,000 m) on the uncer-
1994; Sørensen and Seibert 2007; Lin et al. 2010). To tainties of the predicted flow, sediment, NO3–N, and total
illustrate the effect of DEM data on model outputs, a phosphorus (TP) transport in Moores Creek watershed
widely used watershed scale model the SWAT (Arnold (18.9 km2) in Washington County, Arkansas, USA. Their
et al. 1998) was assessed in this study. studies showed that DEM resolution affected the watershed
The objective in developing the SWAT model was to delineation, stream network and subbasin classification in
predict the impact of management on water, sediment and the model. A coarser DEM resulted in decreased slopes, and
agricultural chemical yields in large ungauged basins increased slope lengths. A decrease in DEM resolution
(Arnold et al. 1998). SWAT has been proven to be an resulted in decreased stream flow, sediment, NO3–N, and TP
effective tool for assessing water resources and nonpoint load predictions with short-term fluctuations. In the Lower
pollution problems (Cheng et al. 2007) for a wide range of Walnut Creek watershed (21.8 km2, central Iowa, USA),
scales and environmental conditions across the globe Chaplot (2005) evaluated the impact of the mesh size of
(Gassman et al. 2007). DEM (from 20 to 500 m) within the SWAT model to sim-
DEM is an important input source for the SWAT model. ulate runoff, sediment, and NO3–N at the outlet of an agri-
The topographic attributes of the subbasins, including area, cultural watershed, and found that computed sediment, and
slope, and field slope length, are all derived from the DEM. NO3–N were more sensitive to DEM resolution than runoff.
So are the channel length, slope, width, and depth, if Dixon and Earls (2009) compared the SWAT modelled
channels are automatically generated based on DEM but predicted stream flow at the original and resampled DEMs in
not previously defined. the Charlie Creek drainage basin (855 km2), located in the
In the SWAT model, the watershed is divided into mul- Peace River drainage basin of central Florida, USA. The
tiple subbasins based on topographic features of the water- results indicated that model was indeed sensitive to the data
shed. Topographic attributes (slope, field slope length, sources of DEMs: ‘‘original’’ 90 m DEM and 90 m DEM
channel length, channel width, etc.) are calculated at the ‘‘resampled’’ from 30 m DEM did not achieve the same
subbasin level and then assigned for all hydrologic response accuracy in stream flow modelling.
units (HRUs) within the subbasin. HRUs that consist of The uncertainty of DEM could be caused by the grid
homogeneous land uses, managements, and soil character- sizes or the data sources, previous studies barely discuss
istics are the basic calculation units. Major model compo- both of them together. Furthermore The DEMs used in the
nents include climate, hydrology, soil temperature, plant previous SWAT model sensitivity studies are mostly con-
growth, nutrients, pesticides, and land management. The tour-based DEMs, few studies focus on DEMs newly
flow, sediment, and other water contaminants generated derived from remote sensing. Therefore, there is a need to
from each HRU in subbasins are integrated and the resulting evaluate the sensitivity of the SWAT model to the newly
loads are routed through channels, ponds and reservoirs to available DEMs (i.e. ASTER 30 m and SRTM 90 m),
the watershed outlet. The SWAT model provides two systematically considering the effects of both grid size and

123
212 Stoch Environ Res Risk Assess (2013) 27:209–221

Table 1 The effect of DEM resolution on the root mean square error (RMSE, m) of the elevation
5m 10 m 20 m 30 m 40 m 60 m 80 m 90 m 100 m 120 m 140 m

DLG 5 m 3.6 3.7 4.4 5.6 7.4 11.3 15.5 17.7 19.8 24.3 28.0
ASTER 30 m 33.3 33.4 33.5 33.1 34.1 34.5 35.8 36.2 37.2 38.4 40.4
SRTM 90 m 57.4 57.5 57.5 57.8 58.0 58.2 59.2 58.9 60.3 60.4 60.6

data source. In this paper, we analyse sensitivity of the 90 m (6 m). In Greece, the accuracy of ASTER 30 m is more
SWAT model annual simulation to three DEM data sources than 16.01 m (RMSE) (Chrysoulakis et al. 2011). Reuter
and eleven resample resolutions. et al. (2009) evaluated five scenes of ASTER 30 m
(N05W02, N40E023, N14W087, N45E008 and S02W079)
1.3 DEM dataset using elevation calculated from laser altimeter signal from
NASA’s GLAS/ICESAT platform as the reference. They
The SRTM 90 m data (Jarvis et al. 2008) is the result of a found that the RMSE was between 18 and 29 m, whereas
collaborative effort by National Aeronautics and Space SRTM 90 m had a much lower RMSE (10–15 m). They
Administration (NASA), USA, the National Imagery and strongly suggested that ASTER 30 m product should not be
Mapping Agency (NIMA), USA, the German space agency, used for any terrain analysis at 30 m resolution, and one
and Italian space agency. The mission was launched on 11 should refer back to SRTM 90 m data wherever possible.
February 2000 aboard the Space Shuttle Endeavour. Data Vertical accuracies of the input DEM datasets including
was collected using two interferometers, C-band and DLG 5 m was evaluated with 1,106 geographic control
X-band systems, which cover most land regions between points which were derived from 1:2,000 scale aerophoto-
60N and 56S at 3 arc-second resolution (about 90 m). grammetry DLG. As shown in Table 1, the DEM accuracy
Following the first release of a research-grade SRTM data generally decreased on resampled grid sizes. The RMSE
set in 2004, a variety of post-processed releases were increased from 3.6 to 28.0 m when the DLG5 m grid size
available. The latest post-processed SRTM is SRTM Ver- was resampled from 5 to 140 m. For ASTER 30 m and
sion 4.1 (SRTM 90 m), released by the Consortium for SRTM 90 m, it increased from 33.3 to 40.4 m and 57.4 to
Spatial Information of the Consultative Group of Interna- 60.6 m, respectively. In the resolution range of 5–140 m, the
tional Agricultural Research in 2009. The accuracy repre- vertical accuracy of DLG was more accurate than ASTER
sented as root mean square errors (RMSEs) for early version 30 m, which was more accurate than SRTM 90 m (Table 1).
of SRTM 90 m varies between 4 and 6 m in low-elevation The accuracies of ASTER 30 m and SRTM 90 m were lower
terrain and deteriorates to 11–14 m in rugged terrain than the global evaluation reports, because the check points
(Rodrı́guez et al. 2005). There should be no deterioration in were not random GPS points but geographic control points
accuracy for improved version SRTM 4.1. The accuracy of which usually were extreme value points located on peaks
the new version of SRTM reported by Hirt et al. (2010) is and valleys. The accuracy of remote-sensing-based DEMs
about 6 m over Western Australia. Up to now, only limited varies with landform and land cover, and differs from place
evaluations have been performed in the new version dataset. to place. Usually, more pits are found in ASTER 30 m than
The ASTER 30 m data is a new global 1 arc-second SRTM 90 m, but in the study area of this paper obvious pits
(about 30 m) elevation data set that was firstly released in were found in SRTM 90 m, and ASTER 30 m was more
June 2009 by METI (Ministry of Economy, Trade and consistent with DLG 5 m data (Fig. 1d–f). The accuracy of
Industry), Japan, and NASA, USA. ASTER 30 m covers ASTER 30 m was higher than SRTM 90 m; that was dif-
land surfaces between 83N and 83S, which is larger than the ferent from some of the previous reports (Hirt et al. 2010;
SRTM 90 m coverage. The data used to generate ASTER Reuter et al. 2009).
30 m were from Visible and near infrared radiometer
(VNIR) Band 3N and 3B (0.76–0.86 l wavelengths). The
overall accuracy of the ASTER 30 m on a global basis is
approximately 9.35 m (average RMSE) (ASTER Validation 2 Methodology
Team 2009). However, the accuracy varies in different pla-
ces. ASTER 30 m is considered to be a research-grade 2.1 Input data
because of a number of artifacts (systematic errors) remain in
the elevation data (ASTER Validation Team 2009). Over The following data were selected as SWAT model inputs:
Western Australia, Hirt et al. (2010) reported that the vertical (1) DEM data from three sources: DLG 5 m, ASTER
accuracy of ASTER 30 m (15 m) was worse than SRTM 30 m, and SRTM 90 m. DLG 5 m was of vector

123
Stoch Environ Res Risk Assess (2013) 27:209–221 213

format of DLG at 1: 10,000 scale provided by the which auto fertilization would be applied when soil N
Bureau of Survey and Mapping, Zhejiang Province or P concentrations were below the thresholds.
with 5 m interval contour lines and 3 m maximum
vertical error, and was different from grid format
STER 30 m and SRTM 90 m. A triangular irregular 2.2 Model run
networks format DEM was generated from the
original contour lines before resampling to different SWAT2005 (available on website http://swatmodel.tamu.
resolution DEMs using ArcGIS 9.3 Desktop (ESRI edu, Accessed on 29 January, 2010) was used in this study. In
Inc., Redlands, CA). ASTER 30 m and SRTM 90 m the SWAT model, the watershed is divided into subbasins
were downloaded from International Scientific & based on the topographic features of the watershed derived
Technical Data Mirror Site, Computer Network from DEMs to preserve the natural flow paths, boundaries,
Information Centre, Chinese Academy of Sciences and channels required for the realistic routing of water,
(http://datamirror.csdb.cn, Accessed on 10 January, sediment and chemicals. When evaluating the uncertainties
2010). All DEMs were resampled to different reso- of the SWAT model outputs on DEM resolution, it is nec-
lutions, including 5, 10, 20, 30, 40, 60, 80, 90, 100, essary to examine the uncertainties of watershed topographic
120, and 140 m, where 5, 30, and 90 m was (or was parameters as model inputs. Therefore, the following topo-
close to) the original resolution for DLG 5 m, ASTER graphic attributions were compared in this study: mean slope
30 m, and SRTM 90 m, respectively. Except for the gradient of subbasins, subbasin area, mean field slope length
original resolutions, the resample resolution series is of subbasins, reach length of subbasins, mean reach slope of
based on a ratio 2 progression in the small scale and subbasins, mean reach width of subbasins, and mean reach
20 m interval in the large scale to obtain 2–3 sample depth of subbasins. Because more than one subbasin was in
points between or after the original resolutions. The the watershed, all the watershed topographic attributes were
resample method was bilinear interpolation using the averaged and weighted by the subbasin areas to simplify
Resample tool in ArcGIS 9.3 Desktop (ESRI Inc., the comparison. Runoff, sediment, total nitrogen (TN),
Redlands, CA). The edge of each DEM was 1 km and TP were the model outputs of interest. Runoff and
outside of the watershed boundary, so that the sediment could be read from the model output file
watershed boundary definition for each DEM would directly. TP consisted of organic phosphorus and min-
not be affected by edge grids. Figure 1 shows the eral phosphorus. TN consisted of organic nitrogen, nitrate,
shaded terrains of original resolution DEMs of DLG nitrite, and ammonium.
5 m, ASTER 30 m and SRTM 90 m. Note that it is It should be noted that model output is affected by both
not common to interpolate ASTER 30 m or SRTM the change of DEM grid size and the subsequently related
90 m to fit a small grid DEM such as 5 m DEM. The changes of topographic variables. Changes in DEM grid
purpose of the down-scaled resampling was to com- size also alter other variables such as soil and land cover,
pare with the finest DEM (i.e. DLG 5 m). whose grid sizes are resampled at the same resolution as
(2) Land use/cover in 2008 was interpreted from AVNIR- the DEM by the model GIS interface. This means that
2 (advanced visible and near infrared radiometer type processes like interception, infiltration, and consequently
2) images of Advanced Land Observing Satellite with runoff generation are altered by changed coarser/different
10 m resolution. soil and land cover parameters. Therefore, the uncertainty
(3) The soil information was digitized from 1:50,000 soil measured here is an overall uncertainty in a complicated
maps of the Second National Soil Survey in China. model.
Soil physical and chemical characteristics referred to ArcSWAT2.3.4 (available on website http://swatmodel.
Wei (1993). tamu.edu, Accessed on 29 January, 2010) was used as a
(4) Ten years (1998–2008) of daily weather gauge data GIS interface for SWAT. ArcSWAT is a pre-processor
(precipitation, temperature, solar radiation, wind interface and post-processor of SWAT model outputs,
speed, and relative humidity) from one station were providing a complete set of tools for watershed delineation,
provided by Weather Service of Zhejiang Province, definition and editing of the hydrological and agricultural
China. management inputs, and running and calibrating the model
(5) The Xiekengxi River watershed was primarily a (Di Luzio et al. 2004). ArcSWAT2.3.4 was used to derive
forested watershed (accounting for 96.0% of area) model parameters from GIS layers for the SWAT model.
without significant point source pollutions, so there The uncertainties in the model outputs due to different
was no input data of point source pollutions. Only DEM resolutions were assessed by running the model at 11
0.9% of the area was covered by farm land, and the different DEM resolutions (i.e. 5, 10, 20, 30, 40, 60, 80, 90,
crop practices were set as model default values, by 100, 120, and 140 m) for each DEM from different data

123
214 Stoch Environ Res Risk Assess (2013) 27:209–221

sources (DLG 5 m, ASTER 30 m and SRTM 90 m), performance in runoff simulation (y = 1.08x ? 14.17,
keeping other simulation conditions constant. These con- r2 = 0.82, where x was the simulated monthly runoff, and
ditions included: (1) the same other input data including y was the observed monthly runoff during January 2000–
land uses, soils, climate, ponds, reservoirs and land man- December 2008). We do not address issues concerning the
agements; (2) minimum upslope drainage area of 800 ha SWAT model calibration and validation in this paper.
for stream network extraction; and (3) the same HRUs
definition thresholds of land use/cover (20%), soil (10%)
and slope (20%), recommended by Winchell et al. (2005) 3 Result
to eliminate tiny polygons of land use/cover, soil and slope
in HRUs. 3.1 Terrain characteristics and estimated stream
The SWAT model outputs and the topography parameter network characteristics
tables derived from watershed delineations were saved in the
SWAT project databases for each run. They covered a period To understand the effect of DEM resolution on SWAT
from 2006 to 2008, but only the annual outputs of 2008 were outputs, DEM-based watershed terrain and reach charac-
analysed to avoid errors in the model’s ‘‘warm-up’’ period. teristics were studied as well as the elevation itself (Figs. 2,
There were 33 runs and 396 variables involved. Visual Basic 3). Those characters are input parameters of SWAT. For all
for Applications (Microsoft Corp.) programming was used DEM sources, results showed that the mean altitudes and
to assist data retrieval from the model databases. watershed areas were not sensitive to resample resolutions
and data sources (Fig. 2a, c), but the mean slopes were
2.3 Quantification of model uncertainties sensitive (Fig. 2b). The mean slope RD variation of finer
due to different DEM resolutions DEM due to resampling (i.e. 40.0% for DLG 5 m) was
higher than the coarser DEMs (i.e. 12.8% for ASTER
The uncertainties of watershed delineations and model out- 30 m, and 13.2% for SRTM 90 m). A change of slope
puts due to different DEM resolutions and data sources were could cause a substantial variation in field slope length in
quantified as relative difference (RD). This was defined as an ArcSWAT watershed delineation. When DEMs were
resampled to coarser resolutions, the field slope lengths
RDð%Þ ¼ 100ðPx  P5 Þ=P5 ð1Þ
derived from ASTER 30 m and SRTM 90 m remained
where Px was any of the watershed delineations or model constant. On the other hand, the one derived from DLG
outputs based on different resolution and source DEMs, 5 m increased with plateaus, and reached maximum (RD =
and P5 was the corresponding base value derived from the 66.7%) at 90 m resolution (Fig. 2d).
finest DEM (DLG 5 m) with a 5 m resolution, and was Reach lengths and reach slopes varied substantially on
assumed to be the ‘‘best’’. Specifically, for the model resample resolution, but no obvious trend could be found
outputs, annual values of runoff (m3 s-1), sediment (ton), (Fig. 3a, b). Because of the different DEM vertical accu-
TP (ton), and TN (ton), derived from the 5 m DLG DEM racies, the reach circuits derived from different data sour-
was used as the baseline scenario. ces had obvious differences, even when they were
Calibration for each run was not necessary in this study, resampled to a same resolution (Fig. 1c). This resulted in
because even with the best possible model calibration for major differences in the topographic features (i.e. reach
each run at different resolution, model errors solely due to lengths and reach slopes) of the estimated reaches. The
DEM inputs could not have been isolated. Instead of using reach widths and reach depths were the function of the
measured watershed response data, the best possible model rainfall accumulation areas and not sensitive to resample
predicted values, P5, were used as baselines to calculate resolution (Fig. 3c, d). The fact that the values of reach
RD. This approach could not only isolate the uncertainty in lengths, widths and depths derived from SRTM 90 m were
the model due to DEM resolution and source, but also much higher than those derived from DLG 5 m and
allow the results of watershed delineations or model out- ASTER 30 m (Fig. 3a–d), indicated that the subbasin
puts to be compared in default parameter conditions, con- boundaries derived from SRTM 90 m were obviously
sidering parameter calibrations varied with different different from the ones derived from ASTER 30 m or DLG
catchments. Another reason for ‘‘no calibration’’ was that 5 m. That could be attributed by the obvious artifacts (pits)
the objective in SWAT’s development was to predict the in the SRTM 90 m data as shown in Fig. 1f.
impact of management on water, sediment and agricultural
chemical yields in large ‘‘ungauged’’ basins where usually 3.2 SWAT predicted outputs
no data available for calibrations (Arnold et al. 1998).
Moreover, in Tiaoxi River watershed which covered the Runoff (m3 s-1), sediment (ton), TP (ton), and TN (ton) are
study area, the uncalibrated model already had good usually the 4 most interested outputs for water resource and

123
Stoch Environ Res Risk Assess (2013) 27:209–221 215

Fig. 2 Watershed topographic


characteristics derived from
DEMs vary with DEM
resolution. The benchmarks of
relative differences (RDs, see
the text) are the values derived
from the finest DEM (DLG
5 m). The lines between data
points are cubic interpolated
trend lines. These notes apply to
the following figures

Fig. 3 Estimated stream


network characteristics derived
from DEMs vary with DEM
resolution

non-point source pollution studies with SWAT. Compari- The predicted sediment values derived from DLG 5 m,
sons of the RDs in the modelled outputs indicated that ASTER 30 m and SRTM 90 m also could be considered as
resample DEM resolution had more substantial effects on close at any resample resolution (RD difference \5%),
TPs (Fig. 4c), followed by TNs (Fig. 4d), both of which whilst decreased slightly with coarser resolutions (Fig. 4b).
decreased substantially with coarser resample resolutions; At any resample resolution, the TP and TN relative pre-
sediments (Fig. 4b) slightly decreased with short-term dicted accuracies derived from DLG 5 m were always
fluctuations; runoffs (Fig. 4a) were not sensitive to ‘‘better’’ (relatively) than the ones derived from SRTM
resample resolution. More specifically, the runoff predic- 90 m, followed by the ones derived from ASTER 30 m
tions derived from DLG 5 m, ASTER 30 m and SRTM (Fig. 4c, d). For nutrient predictions (TP and TN), which
90 m were close at any resample resolution (Fig. 4a), so were very sensitive to DEM resolutions, the output relative
the predicted accuracies of runoffs would not be affected accuracies of the SWAT model derived from ASTER 30 m
by resampling or alternative DEM sources in this study. and SRTM 90 m could be ‘‘improved’’ (relatively) by

123
216 Stoch Environ Res Risk Assess (2013) 27:209–221

Fig. 4 SWAT predicted runoff


(m3 s-1), sediment (ton), TP
(ton), and TN (ton) (measured at
watershed outlet) vary with
DEM inputs of different
resolutions

down-scaled resampling, but they were unable to reach the 2 


Qsurf ¼ Rday  Ia = Rday  Ia þ S ð2Þ
accuracies of the finer DEM (DLG 5 m) at the same res-
olutions (Fig. 4c, d). At original resolutions (i.e. 30 for where Rday is the rainfall for the day (mm), Ia is the initial
ASTER 30 m and 90 m for SRTM 90 m), the predicted abstractions which include surface storage, interception
runoff, sediment, TP, and TN based on ASTER 30 m and and infiltration prior to runoff (mm), and S is the retention
SRTM 90 m were similar (less than 2% of RD difference). parameter (mm), which is a function of the curve number
In a smaller scale of SWAT model outputs, subbasin (CN) for the day. SWAT2005 does not adjust the CN for
simulations with DLG 5 m as input were studied. Figure 5 slope. It explains that, in this study, the mean slopes
illustrates that the subbasin outputs varied with the resam- apparently decreased with coarser resolutions (whether
ple DEM resolution in similar trends as the outputs via resampled or original), but runoff was not sensitive to the
the whole watershed outlet, except that the sediment resolutions. We had observed a consistent trend of runoff
output via the up-stream subbasin outlet decreased sub- and area variations on resampling resolutions. The pre-
stantially on coarser DEM resolution, while the values via dicted runoffs based on DLG 5 m, ASTER 30 m and
the mid-stream and watershed outlet were relatively stable SRTM 90 m all slightly fluctuated with the resampled
(Fig. 5b). The stable watershed sediment could be attrib- resolution, following the same trends of the areas, indi-
uted to depositions and degradations in mid- and down- cating that the changes in runoffs were mainly attributed
streams. More details will be discussed in the following to the changes in water accumulation areas. Previous
Discussion section. Other up- and mid-steam subbasins studies show inconsistent results concerning the effects of
(i.e. subbasin 2, 3, 4, and 6) showed same variation trends the original resolutions on modelled runoff. Some (Wo-
as subbasin 1 and 5, which were used as examples in lock and Price 1994; Cho and Lee 2001; Di Luzio et al.
Fig. 5. 2005) found that runoff decreased with coarser original
DEM resolutions, while others (Bosch et al. 2004; Dixon
and Earls 2009) did not hold the same opinion. Those
4 Discussion comparisons were within a few DEMs with different
resolutions, and were similar with the fluctuation trends in
4.1 Model mechanism a small variation scale of DEM resolution of this study.
Nonetheless, they all found that the mean slope decreased
4.1.1 Runoff with coarser resolutions, which also illustrated that the
mean slope played a minor role in generating runoff
The SCS runoff equation (USDA-SCS 1972) is used in output using the SWAT model. Similar results were found
SWAT2005 surface runoff (Qsurf, mm) calculation: by Chaplot (2005), who resampled DEM from 20 to

123
Stoch Environ Res Risk Assess (2013) 27:209–221 217

Fig. 5 SWAT predicted runoff


(m3 s-1), sediment (ton), TP
(ton), and TN (ton) (at subbasin
outlets) vary with DEM
resolutions. The positions of the
subbasins are shown as (e)

500 m and found no consistent variation for both area and area (Wu et al. 2008). Areas are affected by resolution
runoff. In the study of Dixon and Earls (2009), original more obviously in smaller watersheds. In summary, if the
and resample DEMs with resolutions of 30, 90, and predicted runoff is the focus of the study, area bias should
300 m were used as inputs to assess the uncertainty of the be considered, especially in small area watersheds.
SWAT. Across their nine runs, larger areas resulted in
higher runoffs. Their findings also showed that the vari-
4.1.2 Sediment
ations of areas were irregular on resolution, corroborating
the irregular variations of areas found in this study.
In the SWAT model, erosion (sedsurf) caused by rainfall
However, Cotter et al. (2003) found that the decreases in
and surface runoff is computed using the MUSLE
DEM data resolution (resampled to 30, 100, 150, 200,
(Williams 1975), in which a decreased sediment yield from
300, 500, and 1,000 m) resulted in decreased areas and
HRU is expected with a decrease in mean slope and field
decreased predicted flows, when studying an 18.9 km2
slope length:
watershed, which was inconsistent with the results of this
0:56
study. The reason for the difference might be a smaller sedsurf ¼ 11:8 Qsurf qpeak areahru K C P L S CFRG
watershed area in their study. The DEM aggregation can
ð3Þ
influence the calculated area, but the direction and extent
of the effect will vary upon its base grid size, level of where qpeak is the peak runoff rate (m3 s-1), areahru is the
increase, and edge pattern of the DEM (Vieux and area of the HRU (ha), K is the USLE soil erodibility factor
Needham 1993). A greater variation in area is attributable (0.013 ton m2 h (m3 ton cm)-1), C is the USLE cover and
to a more irregular shape associated with rapid changes in management factor, P is the USLE support practice factor,
topography (due to height), as well as a smaller watershed L is the USLE slope length factor, S is the USLE slope

123
218 Stoch Environ Res Risk Assess (2013) 27:209–221

factor and CFRG is the coarse fragment factor. In this sediment yields in HRUs, which decrease with coarser
study, the sediments did not decrease as much as the mean resolutions. The same variation trend shared by slopes and
slopes and field slope lengths, because sediment channel nutrients (TP and TN) in this study indicates that the nutrient
routing was also simulated in the SWAT. Sediment outputs are affected by DEM resolution via slope. Because
transport in the channel network is a function of two most of P was transported by the sediment, the decreased
processes, deposition and degradation. The maximum gradients of TPs were more than TNs, which were mainly
amount of sediment that can be transported from a reach transported by water (Fig. 4c, d). This result is consistent
segment is calculated by (USDA-SCS 1986): with those reported by Cotter et al. (2003) and Chaubey et al.
(2005), who found that the modelled TP and NO3–N
concsed;ch;mx ¼ csp vspexp
ch;pk ð4Þ
decreased with coarser resample DEM resolution from 30 to
where concsed,ch,mx is the maximum concentration of sed- 1,000 m. The same trends shared by nutrients (TP and TN)
iment that can be transported by the water (ton m-3 or and mean slopes indicate that it is possible to reduce the slope
kg l-1), csp is a coefficient, vch,pk is the peak channel effect caused by DEM resolutions by adding a parameter to
velocity (m s-1), and spexp is an exponent defined by the the SWAT. By doing that, errors caused by resampling could
user. The maximum concentration of sediment is compared possibly be reduced by model calibrations.
to the concentration of sediment in the reach at the
beginning of the time step, concsed,ch,i. If concsed,ch,i [ 4.2 DEM resolution
concsed,ch,mx, deposition is the dominant process in the
reach segment. If concsed,ch,i \ concsed,ch,mx, degradation is Resampling is widely used by GIS-based model workers to
the dominant process in the reach segment. ensure all raster data layers have a same cell size to
Obviously, the peak channel velocity (vch,pk) plays a facilitate spatial analyses. The resolution of GIS data is
critical role in the sediment transported by the water (Eq. 4). closely associated with the computation speed in spatial
The sediment concentrations in the watershed outlet are analysis. When the computing time is too long to run a
influenced by the channel degradations and depositions in model in high resolution, model users may want to
mid- and down-streams. As shown in Fig. 5b, even when the resample the DEM data to a coarser resolution to speed up
sediment outputs from the up-stream subbasins increased, the computation procedure. The model users should be
the sediments in mid-streams and down-streams were still aware that the resample resolution is not a ‘‘real’’ resolu-
stable since they depended on peak channel velocities tion, but a ‘‘nominal’’ resolution. Decreasing the size
(Eq. 4). Moreover, the peak channel velocities could change beyond the resolution of the original data does not increase
by the stream network topographic attributes, which were not the accuracy of the land surface representation of the DEM
sensitive to the resample resolution in this study (Fig. 3). and potentially introduced interpolation errors (Zhang and
Those explains only very slight changes were observed in the Montgomery 1994; Straumann and Purves 2004).
sediment via watershed outlet even the up-stream sediment This study shows that the accuracy of the predicted
outputs increased (Fig. 5b). In the findings of Chaplot SWAT model outputs of finer DEM (DLG 5 m) decreased
(2005), sediment loads predicted by the SWAT model also more rapidly than those of coarser DEMs (ASTER 30 m
varied irregularly and slightly on resample resolutions from and SRTM 90 m) on resample resolutions. Resampling
20 to 200 m. Where soil erosions are not severe and channel coarser DEM input data (ASTER 30 m or SRTM 90 m) to
velocities are high, the sediment contents in the water could higher resolutions could increase the relative accuracy of
be below the water transportation capabilities. In those cases, the annual predicted TPs and TNs, but the predicted out-
coarser the DEM resolutions could cause less sediment loads puts still were lower than values derived from the finer
for lower slopes and less sediments calculated by MUSLE resolution (DLG 5 m), indicating that resampling to finer
(Eq. 3). That explains slight decrease trends in sediment load resolutions could possibly provide DEMs with better rep-
simulations when DEM resolution decrease in some study resentative of the real terrain, but their accuracies were still
areas (Wolock and Price 1994; Cotter et al. 2003; Chaubey limited by the resolution of the original data. For example,
et al. 2005; Di Luzio et al. 2005). as shown in Fig. 4c, d, the predicted TP and TN with 5 m
resampled SRTM 90 m as input are almost equal to the
outputs with 90 m resampled DLG 5 m as input.
4.1.3 TP and TN Whether to resample or not depends on objectives of
studies. For example, in this study, if annual outputs of
In the SWAT model, soluble N and P in runoff, as products of runoff and sediment were the foci of the study and ±10.0%
the volume of water and average concentrations of N and P in of the maximum RD was allowed, all the predicted outputs
soil layers, are not sensitive to watershed topographic attri- were acceptable with any of resample resolutions and any
butions. But sediment transport of N and P are based on the source of the DEM data as input.

123
Stoch Environ Res Risk Assess (2013) 27:209–221 219

4.3 DEM data quality summarized the vertical accuracies of ASTER 30 m and
SRTM 90 m (resampled to 30 m resolution), and indicated
Environmental models often require detailed terrain data. that SRTM 90 m vertical accuracies were the highest for
Nowadays, remote-sensing-derived DEMs form an increas- built-up areas and bare soils, and the lowest for forests,
ingly important source of information for these applications. whereas ASTER 30 m accuracies were the highest for
Yet, it is often not easy to decide what the most suitable source forests and the lowest for agriculture and grassland land
of information is and how different input data affect modelled covers. Therefore, ASTER 30 m could be a better DEM
outcomes. dataset in some circumstances (e.g. mixed forest area of
This study compared qualities and performances of this study area) and should be used with cautions.
remote sensing-derived data (ASTER 30 m, SRTM 90 m)
for the SWAT Model developed for a forest watershed in 4.4 Limitations
South-eastern China. These results indicate that the SWAT
modelled outputs, which were based on ASTER 30 m and This study was conducted under specific conditions for
SRTM 90 m, were similar, even though their vertical accu- specific area, climate, and time scale (yearly). We believe
racies were different (33.1 m RMSE for ASTER 30 m and that our results have indicated DEM uncertainties in the
58.9 m RMSE for SRTM 90 m) and the grid size of SRTM uncalibrated SWAT model. In other circumstances, such as
90 m was larger than ASTER 30 m by about 3 times. For a low gradient (flat) watershed where slope is not sensitive
example, simulated TN based on SRTM 90 m provided to DEM grid size, scales of uncertainties can be different.
similar performance with the one based on ASTER 30 m Especially, when the model is calibrated and weights of the
(RD was -8.2% for ASTER 30 m and -8.0% for SRTM model parameters are changed, the extents of model out-
90 m). It would not be a problem to replace ASTER 30 m puts reacting to DEM might be different. We only present
with the coarser SRTM 90 m in this case, if the annual out- possible uncertainties derived from DEM resolutions and
puts were the foci of the study and ±10.0% of the maximum data sources by a case study.
RD was allowed. In another word, ASTER 30 m with a
higher accuracy did not assure a better performance in the
model simulation. That was because DEM as one of many
5 Conclusions
input parameters, its errors might have been diluted by the
model itself, especially like SWAT which was an integrated
DEM grid size and data source are sources of model
model with hundreds of equations. In this study, the errors in
uncertainties, their effects on the SWAT model predictive
SRTM 90 m coincidently changed the stream network
capabilities are important (particularly for TN and TP
parameters (Fig. 3) resulting in a performance improvement
simulations as shown in this study). This study has high-
in SRTM 90 m. For example, the mean slopes of ASTER
lighted the potential to introduce uncertainties into the
30 m were higher than SRTM 90 m at any resample reso-
SWAT model by using alternatively sourced and resam-
lution, but the predicted TNs and TPs derived from ASTER
pled DEMs and what the likely impacts that these data will
30 m were not higher than those derived from SRTM 90 m
have on the model’s ability to successful simulate the
as expected (Fig. 3). This was explained by the longer reach
chemical and physical processes in watersheds.
length, wider reach width, and deeper reach depth of SRTM
90 m (Fig. 2), all of which resulted in a lower peak channel (1) The SWAT model estimation of runoff was not
velocity, more subsequent sediment deposit in the subbasin sensitive (RD B1.0%) to both DEM resample reso-
reaches and hence less nutrients in the sediment discharged lutions and the data sources in this study. The extent
from the watershed. Notwithstanding, a DEM with better of the watershed areas affected by DEMs indicated
accuracy should be recommended to avoid possible errors in the extent of the runoffs affected by DEMs. There-
model simulations. fore, in smaller watersheds whose areas could be
Comparing to SRTM 90 m, the quality of the new sensitive to DEM grid sizes, coarser DEMs should be
ASTER 30 m dataset has been debating, and significant used with cautions.
anomalies and artifacts in ASTER 30 m has been noticed (2) If the sediment was the focus of the study and a 6.0%
even it has better nominal resolution than SRTM 90 m RD was permitted, we could choose DEMs from any
(Hirt et al. 2010; Reuter et al. 2009). Further investigation source (DLG 5 m, ASTER 30 m, or SRTM 90 m)
indicates that ASTER 30 m spatial resolution is around and resampled resolution (5–140 m) as model input in
100 m which is worse than SRTM 90 m (ASTER Valida- this study. The results suggested that any of the data
tion ASTER Validation Team 2009). However, in some sources are comparable for sediment simulations in
places (e.g. the area of this study) accuracy of ASTER the study area. However, in watersheds where soil
30 m is better than SRTM 90 m. Blanchard et al. (2010) erosions are not severe and channel velocities are

123
220 Stoch Environ Res Risk Assess (2013) 27:209–221

high, the sediment contents in water could be below Cho SM, Lee M (2001) Sensitivity considerations when modeling
the water transportation capabilities. In those cases, hydrologic processes with digital elevation model. J Am Water
Resour Assoc 37(4):931–934
coarser DEMs could cause lower simulated Chrysoulakis N, Abrams M, Kamarianakis Y, Stanislawski M (2011)
sediments. Validation of ASTER GDEM for the area of Greece. Photo-
(3) The modeled TP and TN were sensitive to DEM gramm Eng Remote Sens 77(2):157–165
resolutions and data sources as much as the DEM Cotter AS, Chaubey I, Costello TA, Soerens TS, Nelson MA (2003)
Water quality model output uncertainty as affected by spatial
derived slopes did. The TP and TN decreased resolution of input data. J Am Water Resour Assoc 39(4):
significantly on the resample grid sizes, and this 977–986
should be considered when replacing the original Di Luzio M, Srinivasan R, Arnold JG (2004) A GIS-Coupled
resolution DEMs with resample ones. For finer DEM hydrological model system for the watershed assessment of
agricultural nonpoint and point sources of pollution. Trans GIS
(DLG5 m), the predicted values of TP and TN were 8(1):113–136
much higher than those in coarser DEMs (ASTER Di Luzio M, Arnold JG, Srinivasan R (2005) Effect of GIS data
30 m and SRTM 90 m) (RD difference [8.0%). quality on small watershed stream flow and sediment simula-
tions. Hydrol Process 19(3):629–650. doi:10.1002/hyp.5612
To decide which DEM source and grid size is appro- Dixon B, Earls J (2009) Resample or not?! Effects of resolution of
priate for the SWAT model, interested outputs and the DEMs in watershed modeling. Hydrol Process 23(12):1714–1724.
doi:10.1002/hyp.7306
maximum acceptable errors should be predetermined for a
Gassman PW, Reyes MR, Green CH, Arnold JG (2007) The soil and
specific watershed. For sensitive research objects such as water assessment tool: historical development, applications, and
TP and TN, resampling DEMs to finer resolutions could future research directions. Trans ASAE 50(4):1211–1250
improve the model performances, but extents of the Green WH, Ampt GA (1911) Studies on soil physics, part I—the flow
of air and water through soils. J Agric Sci 4:1–24
improvements were limited by the original DEM accura-
Hirt C, Filmer MS, Featherstone WE (2010) Comparison and
cies. Global modeled results of finer DEM (DLG 5 m) validation of the recent freely available ASTER-GDEM ver1,
showed substantial differences from those of coarser DEMs SRTM ver4.1 and GEODATA DEM-9S ver3 digital elevation
(ASTER 30 m and SRTM 90 m) even at a same resample models over Australia. Aust J Earth Sci 57(3):337–347. doi:
10.1080/08120091003677553
resolution. We recommend higher accuracy DEMs as
Jarvis A, Reuter HI, Nelson A, Guevara E (2008) Hole-filled SRTM
model inputs to avoid more uncertainties. for the globe Version 4, available from the CGIAR-CSI SRTM
90 m Database (http://srtm.csi.cgiar.org)
Acknowledgments The funding of this study was provided partly Lin K, Zhang Q, Chen X (2010) An evaluation of impacts of DEM
by National Key Project of Water Research (Grant No. 2011ZX07). resolution and parameter correlation on TOPMODEL modeling
We sincerely appreciate Stefan Fritsch and GBP Huddlestone for their uncertainty. J Hydrol 394(3–4):370–383. doi:10.1016/j.jhydrol.
helps on language improvement of this paper. 2010.09.012
Maddalena RL, McKone TE, Hsieh DPH, Geng S (2001) Influential
input classification in probabilistic multimedia models. Stoch
Environ Res Risk Assess 15(1):1–17. doi:10.1007/pl00009786
McElroy AD, S.Y. Chiu JW, Nebgen AA, Bennett FW (1976)
References Loading functions for assessment of water pollution from
nonpoint sources. EPA document EPA 600/2-76-151. USEPA,
Arnold JG, Srinivasan R, Muttiah RS, Williams JR (1998) Large area Athens, GA
hydrologic modeling and assessment—part 1: model develop- Neitsch SL, Arnold JG, Kiniry JR, Srinivasan R, Williams JR (2005)
ment. J Am Water Resour Assoc 34(1):73–89 Soil and water assessment tool input/output file documentation
ASTER GDEM Validation Team (2009) ASTER global DEM version 2005. Temple, Texas
validation summary report Reuter HI, Nelson A, Strobl P, Mehl W, Jarvis A, IEEE (2009) A first
Blanchard SD, Rogan J, Woodcock DW (2010) Geomorphic change assessment of ASTER GDEM tiles for absolute accuracy,
analysis using ASTER and SRTM digital elevation models in relative accuracy and terrain parameters. In: 2009 IEEE inter-
central Massachusetts, USA. GISci Remote Sens 47(1):1–24. national geoscience and remote sensing symposium, vols 1–5.
doi:10.2747/1548-1603.47.1.1 IEEE International Symposium on Geoscience and Remote
Bosch DD, Sheridan JM, Batten HL, Arnold JG (2004) Evaluation of Sensing IGARSS. IEEE, New York, pp 3665–3668
the SWAT model on a coastal plain agricultural watershed. Rodrı́guez E, Morris CS, Belz JE, Chapin EC, J.M. M, W. D, Hensley
Trans ASAE 47(5):1493–1506 S (2005) An assessment of the SRTM topographic products,
Chaplot V (2005) Impact of DEM mesh size and soil map scale on Technical Report JPL D-31639
SWAT runoff, sediment, and NO3–N loads predictions. J Hydrol Skeffington R (2006) Quantifying uncertainty in critical loads:
312(1–4):207–222. doi:10.1016/j.jhydrol.2005.02.017 (A) literature review. Water Air Soil Pollut 169(1–4):3–24. doi:
Chaubey I, Cotter AS, Costello TA, Soerens TS (2005) Effect of 10.1007/s11270-006-0382-6
DEM data resolution on SWAT output uncertainty. Hydrol Sørensen R, Seibert J (2007) Effects of DEM resolution on the
Process 19(3):621–628. doi:10.1002/hyp.5607 calculation of topographical indices: TWI and its components.
Cheng HG, Ouyang W, Hao FH, Ren XY, Yang ST (2007) The non- J Hydrol 347:79–89
point source pollution in livestock-breeding areas of the Heihe Straumann RK, Purves RS (2004) Resolution sensitivity of a
River basin in Yellow River. Stoch Environ Res Risk Assess compound terrain derivative as computed from LiDAR-based
21(3):213–222. doi:10.1007/s00477-006-0057-2 elevation data. In: Fabrikant SI, Wachowicz M (eds) Proceedings

123
Stoch Environ Res Risk Assess (2013) 27:209–221 221

of AGILE, Aalborg, Denmark (lecture notes in Geoinformation Williams JR (1975) Sediment-yield prediction with universal equation
and Cartography) using runoff energy factor. In: Lab US (ed) Present and prospective
Tarantola S, Giglioli N, Jesinghaus J, Saltelli A (2002) Can global technology for predicting sediment yield and sources: Proceedings
sensitivity analysis steer the implementation of models for of the sediment-yield workshop, Oxford, MS, November 28–30
environmental assessments and decision-making? Stoch Environ 1972. USDA Sedimentation Lab., Oxford, pp 244–252
Res Risk Assess 16(1):63–76. doi:10.1007/s00477-001-0085-x Williams JR, Hann RW (1978) Optimal operation of large agricultural
USDA-SCS (1972) National engineering handbook, hydrology sec- watersheds with water quality constraints. Texas Water
tion 4, chap. 4–10. U.S. Department of Agriculture, Soil Resources Institute, Texas A&M Univ., Tech. Rept. No. 96
Conservation Service, Washington, DC, USA Winchell M, Srinivasan R, Di Luzio M, Arnod J (2005) ArcSWAT
USDA-SCS (1986) Urban hydrology for small watersheds. Tech. 2.3 interface for SWAT2005 user’s guide. Temple, Texas
Release 55. US Dept. of Agriculture, Soil Conservation Service, Wolock DM, Price CV (1994) Effects of digital elevation model map
Washington, DC, USA scale and data resolution on a topography-based watershed
Vieux BE, Needham S (1993) Nopoint-pollution model sensitivity to model. Water Resour Res 30(11):3041–3052
gride-cell size. J Water Resour Plan Manag-ASCE 119(2): Wu S, Li J, Huang GH (2008) A study on DEM-derived primary
141–157 topographic attributes for hydrologic applications: sensitivity to
Wagener T, Gupta HV (2005) Model identification for hydrological elevation data resolution. Appl Geogr 28(3):210–223. doi:
forecasting under uncertainty. Stoch Environ Res Risk Assess 10.1016/j.apgeog.2008.02.006
19(6):378–387. doi:10.1007/s00477-005-0006-5 Zhang WH, Montgomery DR (1994) Digital elevation model grid
Wei X (1993) Records of Soil Species in Zhejiang. Zhejiang Science size, landscape representation, and hydrologic simulations.
and Technology Press, Hangzhou Water Resour Res 30(4):1019–1102

123

Das könnte Ihnen auch gefallen