You are on page 1of 14

Available online at www.sciencedirect.

com

ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283 – 296


www.elsevier.com/locate/isprsjprs

Comparison of remotely sensed water stages from LiDAR,


topographic contours and SRTM
G. Schumann a,b,⁎, P. Matgen a , M.E.J. Cutler b , A. Black b , L. Hoffmann a , L. Pfister a
a
Department of Environment and Agro-biotechnologies, Public Research Centre-Gabriel Lippmann, Belvaux, L-4422, Luxembourg
b
Environmental Systems Research Group, School of Social Sciences, Dundee University, Dundee, DD14HN, UK
Received 27 February 2007; received in revised form 7 August 2007; accepted 24 September 2007
Available online 8 November 2007

Abstract

Digital elevation models (DEMs) are at the core of most environmental process modelling and disaster management. In flood
inundation modelling, surface elevation constitutes one of the most important model boundary conditions. With the availability of
high-precision DEMs (e.g. LiDAR) and globally available DEMs (e.g. SRTM InSAR) a big step seems to have been taken in terms
of hydraulic modelling application or hydraulic information retrieval from such DEMs, with high potential in particular for
ungauged basins. Comparative studies exist that report on both the validation of different remotely sensed elevation sources and
their use for both hydrologic and hydraulic studies. To contribute to the existing literature on DEMs and hydraulic information, this
study aims at comparing water stages derived from LiDAR, topographic contours and SRTM. A flood inundation model calibrated
with distributed ground-surveyed high water marks is used to evaluate the remotely sensed water stages. The results show that, as
expected, LiDAR derived water stages exhibit the lowest RMSE (0.35 m), followed by the contour DEM (0.7 m). A relatively good
performance of the SRTM (1.07 m), which is possibly linked to the low-lying floodplain, suggests that the SRTM is a valuable
source for initial vital flood information extraction in large, homogeneous floodplains. Subsequent 3D flood mapping from
remotely sensed water stages confirms this but also indicates that flood mapping with low-resolution, low-precision surface
elevation data is hardly possible on the small scale, as the accuracy of the resulting map depends too much on DEM uncertainties
and errors both in the horizontal and vertical directions.
© 2007 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

Keywords: LiDAR; SRTM; Topographic contour DEM; Water stage; Flood inundation model

1. Introduction and background der, 2006). Hydrologic and flood inundation modelling, in
particular, are highly dependent on accurate (vertical
Topographic data, in form of digital elevation models precision in the cm range), high-resolution (cell size
(DEMs), are a fundamental requirement for many resolution ≤ area of an individual building, i.e. b10 m),
environmental modelling problems (Ludwig and Schnei- spatially distributed elevation data (e.g. Walker and
Willgoose, 1999; Yin and Wang, 1999; French, 2003;
⁎ Corresponding author. Department of Environment and Agro-
Charlton et al., 2003; Webster et al., 2004; Haile and
biotechnologies, Public Research Centre-Gabriel Lippmann, Belvaux, Rientjes, 2005; Wang et al., 2005; Hollaus et al., 2005; Liu
L-4422, Luxembourg. Tel.: +352 470261 417; fax: +352 470264. et al., 2005; LeFavour and Alsdorf, 2005; Ludwig and
E-mail address: schumann@lippmann.lu (G. Schumann). Schneider, 2006). This is also reflected in extensive reviews
0924-2716/$ - see front matter © 2007 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V.
All rights reserved.
doi:10.1016/j.isprsjprs.2007.09.004
284 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

(e.g. Wechsler, 2007) and comparative studies (e.g. resolution, high-precision LiDAR data using multiple
Kenward et al., 2000; Wilson and Atkinson, 2003; Ling linear regression analysis to allow continuous derivation
et al., 2005; Hancock et al., 2006; Casas et al., 2006) of of remotely sensed water stages.
available DEM datasets. However, an important limitation In this study, an integrated and improved modelling
of three-dimensional topographic data is that of varying approach to that adopted by Matgen et al. (2007) is used.
precision. It is sensible to argue that the precision of Based on the assumption of steady-state (as satellite image
available DEMs depends largely on the method used to acquisition is instantaneous and not continuous over time),
derive the datasets. Elevation data can be obtained by the Regression and Elevation-based Flood Information
topographic survey, photogrammetry, or remote sensing. It eXtraction (REFIX) model described in more detail by
may be argued that these systems are complementary. For Schumann et al. (2007), uses linear, non-linear or piecewise
instance, remote sensing in the form of Interferometric linear regression to derive spatially continuous (flood) water
Synthetic Aperture Radar (InSAR) can provide large area stages at the contact zone of remotely sensed flood
DEMs but has inherent speckle noise and requires ground boundaries and a DEM, with respect to localised flow
control points (GCPs) to establish height accuracy (Sun et behaviour. Apart from deriving spatially continuous water
al., 2003). Arguably, the InSAR system used on the Shuttle flow from remote sensing, REFIX data can be used to
Radar Topography mission (SRTM) in February 2000 produce a triangular irregular network mesh able to reliably
currently provides the most complete and robust source of map flood area, extent and depth. Schumann et al. (2007)
elevation data at the global scale. Airborne Light Detecting have shown that flood depth can be mapped with an
and Ranging (LiDAR) instruments provide high precision accuracy of below 20 cm, evaluated with spatially distribu-
topographic mapping tools via discrete surface height ted high water marks in the field, but highlight the impact of
samples instead of a continuous coverage across small the horizontal resolution of a DEM on the performance of the
areas (Sun et al., 2003). Such high-resolution, high- REFIX model as a major error source. They conclude that
precision data may be required for field or local scale for small scale modelling a DEM of a grid size of ≤25 m
studies or to obtain highly accurate height references. should be used. They do not, however, investigate the impact
However, DEM availability, which is often determined of varying vertical precision of DEMs but merely comment
by economic constraints and condition of surveying that vertical accuracy of below 0.5 m seems necessary to
environment (e.g. terrain accessibility, topography and achieve good modelling results at scales ranging from a few
geometry, vegetation cover) may quickly turn the comple- hundred metres to a few hundred kilometres.
mentarity of DEM datasets of varying scale and quality Hence, this study builds upon previous work by the
(e.g. airborne LiDAR-, topographic contour-, or SRTM- authors in applying the REFIX model to estimate water
DEM) into competitiveness. Therefore, it is important to stages from an airborne LiDAR-, a topographic contour-,
assess the robustness and potential of different DEMs for and the globally available SRTM DEM. SAR-derived
use in hydrologic or hydraulic modelling and information flood boundaries are integrated with the different DEMs
extraction at the meso-scale. Also, for ungauged basins, to retrieve water stage data. These data will be evaluated
remotely sensed height data are often some of the only with water stage outputs from a 1D hydrodynamic model
available data for environmental process modelling and (HEC-RAS) calibrated with field data.
disaster management. It is to be expected that the LiDAR DEM will provide
In this context, it is the aim of this paper to provide a the most accurate estimate of water stages and thus also
comparative study on the potential use of different the most consistent flood map, given its reported higher
DEMs to provide useful hydraulic information, such as precision for hydraulic modelling (Wilson and Atkin-
water stage (for reviews, see Smith, 1997; Raclot, son, 2003; Casas et al., 2006) and the fact that HEC-
2006). Typically, previous approaches have been used RAS is conditioned on the same DEM.
that consist of marking the water–land contact zone Due to the coarse vertical resolution of SRTM
from satellite image data, the spatial position of which (specified at around 16 m linear vertical absolute height
can be transferred to an elevation source to derive water error prior to the mission and assessed at around 6 to 7 m
volume, width, length, area and stage (Gupta and globally by Rodriguez et al. (2005)) and its 3-arc seconds
Bodechtel, cited in Gupta and Banerji, 1985). Cracknell spatial resolution, it is likely that the SRTM DEM may not
et al. (1987) used a similar method to find heights above be of optimal spatial and/or vertical resolution for flood
mean sea level of shorelines from multiple remotely studies. However, the large variation in average global
sensed images and termed it the waterline method. As a height precision found in the SRTM literature (Table 1)
refinement to this approach, Matgen et al. (2007) seems to indicate that the vertical precision of the SRTM
combine flood boundary information with high spatial dataset depends considerably on location, terrain
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 285

Table 1 river and, as a result, the low-lying villages within its


Varying reported height accuracies for the SRTM DEM floodplain have experienced frequent flooding in the last
Average height precision Source few decades. On the 2nd January 2003 a high magnitude
over land (m) event with a specific peak discharge of 0.63 mm h− 1
16 SRTM mission specification (70.5 m3 s− 1) and a return period of five years occurred.
(Rodriguez et al., 2005) Fortunately, this flood was recorded by the ASAR
11.25 Sun et al. (2003)
instrument on board the ENVISAT satellite at the time of
10 Rabus et al. (2003)
6 Haase and Frotscher (2005) flood peak. The remotely sensed flood image as well as
5.5 LeFavour and Alsdorf (2005) a dense measuring network (recording rainfall, temper-
7.6 Gorokhovich and Voustianiouk (2006) ature, discharge, piezometric levels, soil moisture, and
4.1 Gorokhovich and Voustianiouk (2006) water stages) operating over a decade, supplemented by
3.6 Berry et al. (2007)
an extensive field campaign after the flood, provided
detailed flood inundation information.
characteristics and surface feature properties, and so
deserves further investigation. 2.2. Remotely sensed data
It is worth noting that some other studies on large
areas have used SRTM data to derive (i) water slope and For remotely sensed water stage estimation, flood
discharge with an absolute percentage error of 4.03 boundaries and a DEM are combined. A VH-polarised
(max: 27.35) for the Amazon River (LeFavour and C-band ASAR image acquired with an incidence angle
Alsdorf, 2005) and (ii) water volume and inundated of 35° was used to produce the flood extent map. It is
water surface area of the Three Gorges Reservoir (Wang worth noting that determining the spatial position of
et al., 2005). Although similar hydraulic variables are flood boundaries from radar imagery is not as
derived to those generated by the REFIX model, it straightforward as could be expected given the specular
should be noted that the methods applied to the SRTM reflection of smooth water surfaces away from the radar
data in these two studies differ significantly from the antenna. This is largely due to the positional uncertainty
REFIX modelling approach in that the water stages used of the flooded zone resulting from image geo-coding
in this study are remotely sensed and not ground-based: errors (Matgen et al., 2007), and due to wind roughening
and the effects of protruding vegetation, both of which
(i) LeFavour and Alsdorf (2005) estimate river may produce significant radar returns (Horritt et al.,
discharge using the Manning equation (Eq. (1)) 2001).
with the slope, S and channel width estimated from The following three DEMs of varying precision,
the SRTM DEM but use ground-surveyed water coverage and availability were obtained for the study
depth from nautical maps. Also, water slope (S) area (Fig. 1b):
should not be confused with the REFIX generated
water surface slope (a) of the regression model. S (i) A high spatial resolution, high-precision LiDAR
denotes the ‘real’ slope of the river whereas a DEM of the natural floodplain at a spatial
represents the slope of the regression line based on resolution of 2 m and a reported height accuracy
remotely sensed water heights at the flood extent. of around 15 cm (RMSE). This DEM is the result
(ii) Wang et al. (2005) use a flood mapping procedure of an ‘eroded’ digital surface model generated
similar to that of the REFIX model. However, from discrete height sampling.
instead of deriving water stages from the SRTM (ii) A 50 m DEM generated from vectorised topo-
DEM at the maximum flood extent for regression graphic contours.
modelling of the water surface, they use only (iii) The globally available SRTM DEM at a spatial
uppermost-and lowermost-gauged stages within resolution of 3″ and a height precision of over 6 m
the modelled reach. (vertical absolute height error; Rodriguez et al.,
2005). The SRTM DEM has been pre-processed
2. Study site and available data and is the result of substantial editing (MapMart,
2006).
2.1. Study site description
For validation of surface elevation from the different
The study site is a flood prone area drained by the DEMs, well-distributed accurate elevation data from the
Alzette River north of Luxembourg City (Fig. 1a). The national geodesy network are available (Fig. 1a).
286 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

Fig. 1. (a) Map showing study site location and description and (b) the different DEM datasets used.

2.3. Hydraulic modelling parameters to be calibrated. Manning's equation for


gradually varied flow in open channels is given by:
Hydraulic modelling was carried out on a 10 km reach 1 2 1
of the Alzette River between the towns of Steinsel and V ¼ R3 S 2 ð1Þ
n
Mersch (Fig. 1a) with the widely used 1D hydrodynamic
HEC-RAS model (www.hec.usace.army.mil). Data from For model calibration, high water marks measured in
76 evenly spaced river cross-sections collected through the field at several river cross-sections during the 2003
channel geometry surveying and LiDAR scanning of the flood event were used (Fig. 1a). The HEC-RAS simulated
floodplain, along with the geometry of two bridges, waterline was used to evaluate the remotely sensed
provided the necessary geometric data input. As an waterline modelled from the estimated water stages.
upstream and downstream boundary condition to the
model, a flow hydrograph of the 2003 flood event and a 3. Methodology
friction slope set at 0.0005 were used. In this study the
model was set to compute unsteady flow regime based on The methodology consisted of four different steps:
the continuity and momentum equations. The friction
force that acts on the system is given by the well-known (i) Assess the quality of the different DEMs.
Manning equation (Eq. (1)), which relates flow velocity V (ii) Estimate water stages from the different DEM
(m s− 1) to the hydraulic radius R (cross-sectional area in datasets integrated with the ASAR image using
m2/wetted parameter in m) according to the given friction the REFIX model.
slope, S (m/m). The friction or roughness coefficient, n, in (iii) Calibrate the flood inundation model with field-
Eq. (1) for both channel and floodplain are the model based high water marks and evaluate the REFIX-
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 287

estimated flood waterline with that simulated by to generate a triangular irregular network (TIN) mesh of
the model. the water surface (Matgen et al., 2007). The difference
(iv) Conduct 3D flood mapping from the remotely between the water surface TIN and the DEM provides
sensed waterline. an actual water or flood depth map.
For waterline modelling based on remotely sensed
3.1. Assessing DEM quality water stages of the Alzette River reach, Schumann et al.
(2007) propose a linear regression model of the form:
The different DEMs were validated with reference
H ¼adþb ð3Þ
elevation data distributed across the low-lying flood-
prone area. The height precision or quality of each DEM where H denotes the estimated water stage (m a.s.l.), a
is expressed by the RMSEDEM given by: represents the slope of the regression (m/m) which of
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi course changes with stream (bed) characteristics, d is the
uP
un downstream river distance (m) and b is the intercept (m).
u ðERi  EDEMi Þ2
ti¼1 It is, however, necessary to verify whether or not
RMSEDEM ¼ ð2Þ uncertainties resulting from the integration of an
n
‘uncertain’ remotely sensed flood map with coarser
where ER denotes the reference elevation data, EDEM is DEMs in terms of both height and spatial resolution are
the elevation data as provided by the different DEMs, too large. If so, then the effect may be that the slope of
and n corresponds to the total number of reference data the water flow is hydraulically unrepresentative for the
points used. river reach under study. In such a case, an algorithm
proposed by Raclot (2006) is adopted that adjusts or
3.2. Water stage estimation harmonises water height fluctuations in such a way as to
reflect flow coherence. For flow this means that water
Water stage was estimated with the REFIX model. height has to decrease with flow direction. Fig. 2
Flood boundaries extracted from the ASAR flood image illustrates the automated procedure for flow coherence
were overlain on a DEM to extract elevation data on the adjustment that is applied in this study if initial flow-
left and right extent at a given river cross-section thereby incoherent water stages generate an unrepresentative
assuming, like most hydrodynamic models, a horizontal waterline slope for the Alzette River reach. After
water level across the floodplain perpendicular to the successful application of this procedure, the regression
channel. The mean water height data of the left and right model given by Eq. (3) is used to generate a more
extent at each cross-section are the forcing data for the reliable waterline.
regression analysis. Prior to flood model calibration, the field-based high
The regression estimated water stages at river cross- water marks are used to evaluate the REFIX-estimated
sections and interpolated between sections can be used water stages with the flood inundation model simulated

Fig. 2. Flow coherence algorithm adopted from Raclot (2006).


288 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

water stages according to the RMSEREFIX measure directly or indirectly related to location on the globe, it is
given by: reasonable to suggest that the SRTM dataset should be
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi confined as much as possible to the specific area of
uP
un interest. Therefore, in this study, all DEM datasets have
u ðHMi  HEi Þ2
ti¼1 been constrained to the flood-prone area, within which
RMSEREFIX ¼ ð4Þ well-distributed in-situ reference elevation data have been
n
selected from the national geodesy network. Using these
where HM denotes the HEC-RAS modelled stage, HE marks, elevation from each remotely sensed DEM is
stands for the REFIX estimated water stage, and n evaluated with little variation between the reference and
corresponds to the total number of water stages assessed. DEMs observed (Fig. 3). Even the SRTM does extremely
well in comparison to the other DEM datasets.
3.3. Flood inundation model calibration DEM precision according to Eq. (2) is as follows:

Height precision is of prime importance for hydraulic (i) RMSELiDAR: 0.23 m


modelling, as it constitutes a major boundary condition in (ii) RMSEcontours: 0.95 m
flood inundation models. Therefore, in this study the (iii) RMSESRTM: 1.55 m
LiDAR dataset along with channel surveying provides the
surface elevation for the model. Moreover, it is not the aim The fact that these error values are so low is related to
to test the modelling capabilities of HEC-RAS with the strict limitation of each DEM to the area of interest,
different remotely sensed elevation data, as this has been which is a relatively flat, low-lying and small floodplain.
done before (e.g. Wilson and Atkinson, 2003; Casas et al.,
2006), but rather to investigate the use of different DEMs 4.1.1. SRTM DEM
to derive remotely sensed water stages. Although the SRTM DEM has the lowest precision
The field-based high water marks are used to calibrate and also the maximum error values (up to 3.28 m) and
the model. During a large number of simulations highest variations (Fig. 4), its overall height accuracy of
(N 20,000) performed in a Monte Carlo-type framework 1.55 m indicates that it is applicable to large,
with randomly generated model parameter sets, the homogeneous floodplains to derive useful hydrologic
distance between each observed water level and the one and hydraulic information. This confirms the observa-
simulated gives a fuzzy rules-based model performance tions of previous work by LeFavour and Alsdorf (2005)
between 0 and 1 according to a trapezoidal membership and Wang et al. (2005), for example.
function (see Matgen et al. (2004) for a more detailed
description of the calibration procedure). The simulation 4.1.2. Topographic contour DEM
obtained with the median parameter set values from the The performance of the topographic contour DEM
range of sets of the best performing model runs is taken to with a height precision of only 0.95 m is surprisingly
evaluate the remotely sensed water stages.

4. Results and discussion

This section is subdivided into three parts. First, the


quality and suitability of each DEM is determined and
commented upon. Then, the results of water stage
estimation and waterline modelling based on the
integration of the ASAR-derived flood extents and the
different DEMs are reported. Finally, after each remotely
sensed waterline has been evaluated with that simulated
by the HEC-RAS model, flood maps are generated from
the remotely sensed waterlines and assessed.

4.1. DEM quality assessment

Due to the reported large uncertainties in elevation of Fig. 3. Graph showing the relative performance of the different
the SRTM DEM (Table 1), which depend on many factors remotely sensed DEMs used.
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 289

be used in PUB research. When integrating the SRTM (or


LiDAR) DEM and a remotely sensed flood image, using
the REFIX waterline method, it is possible to get a general
idea on the magnitude of relevant flood inundation
forcing data, though the accuracy and reliability of such
data needs to be verified somehow.

4.2. Remotely sensed water stages and waterline modelling

For water stage and water surface line modelling of the


2003 flood event on the Alzette River, a flood map was
first extracted from the ASAR-VH image. Then, the mean
values between left and right flood boundary elevation
data extracted from each DEM denote the remotely sensed
Fig. 4. Box plots showing the distribution of the residuals of the
water stages and are used to model the flood waterline by
different DEMs. applying the regression model in Eq. (3).
The ENVISAT ASAR image has an initial spatial
resolution of 25–30 m, which was received with a
poor and related to the fact that there is not enough sampling pixel size of 12.5 m produced from multi-look
variation in height within the area of interest to be filtering. After successful geo-referencing to a positional
captured with contours at 10 m intervals. As a accuracy of approximately 25 m (∼ 1 raw pixel) and
consequence, there are regions in the resulting contour image speckle removal with a 5 × 5 Frost filter window,
DEM with very poor interpolated estimations of the flooded area was extracted with the simple but
elevation. Although for this study this DEM turns out nonetheless effective thresholding method (Fig. 1a)
to be of relatively poor quality, better performance is to (Deshmukh and Shinde, 2005). Ninety-one field-based
be expected on larger floodplains with more height GPS marks of the maximum flood boundary showed a
variation. satisfactory fit with the ASAR-derived flood extent
(within 1–2 pixels location error). Water stage data
4.1.3. LiDAR DEM extracted from each DEM at the left and right ASAR-
As expected, LiDAR gives the best performance. Its derived boundaries are plotted against downstream river
height precision compared to the geodesic reference reach distance (Fig. 5). Remotely sensed water stage
data of around 20 cm is within proximity to the product extraction is only possible at river cross-sections, for
specification of 15 cm. This superiority is confirmed by which flooding is detectable on the ASAR image. If
many studies that use LiDAR for either hydrologic (e.g. water stages at other cross-sections are required for more
Charlton et al., 2003; Hollaus et al., 2005; Liu et al.,
2005) or hydraulic modelling (e.g. French, 2003;
Wilson and Atkinson, 2003; Lane et al., 2003; Webster
et al., 2004; Haile and Rientjes, 2005; Casas et al.,
2006). However, two important limitations of LiDAR
are that (i) it provides only discrete surface height
samples and not a continuous coverage (Sun et al.,
2003) and (ii) its availability is very much limited by
economic constraints. The latter limitation poses a major
constraint to flood studies and other investigations that
require high-resolution and high-precision data over
remote or rarely visited areas.
Another point worth noting is that in the context of
Prediction in Ungauged Basins (PUB) studies, field-based
forcing or evaluation data are rarely available. As DEMs
from topographic contours require some sort of ground-
based data to generate contour lines, DEMs from remote
sensing present the only source of elevation data that can Fig. 5. Initial remotely sensed water stages.
290 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

accurate 3D flood mapping, they can be estimated using line is lowered in each case, which will of course impact
the derived waterline regression model. on subsequent 3D flood mapping where the equations are
In Fig. 5, the effects of the data uncertainties inherent applied to estimate a reach-representative water stage at
in the SRTM DEM (see Fig. 4) on remotely sensed each river cross-section for the generation of a water
water stages become apparent. It is clear that there are surface TIN. The lowering of the regression line results
several zones of flow incoherence, meaning that SRTM- from a reduction of the initial over-estimation in water
derived water stages exhibit a lot of incoherent ‘jumps’ stages at several locations due to image position errors
in the direction of the flow, far more than the other two near higher sloping terrain. Eqs. (5)–(10) in Table 2 give
datasets. Prior to waterline modelling, erroneous water the regression model for reach-representative remotely
stages are adjusted using the flow coherence algorithm sensed waterline modelling, before (Eqs. (5)–(7)) and
(Fig. 2). The results are shown in Fig. 6. after (Eqs. (8)–(10)) flow adjustment.
The linear regression equations for waterline model- It can be established for each DEM whether flow-
ling in Table 2 show that although the flow adjustment coherent waterline modelling is required. In the case of
algorithm improves the initial water stage data in the SRTM-derived water stages this largely improves
hydraulic terms, the intercept position of the regression regression modelling, as is reflected by the change in the

Fig. 6. Remotely sensed waterlines compared to that simulated by the 1D hydrodynamic HEC-RAS model. Flow-adjusted remotely sensed water
stages as well as the level of the field-based high water marks are also indicated on each graph.
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 291

Table 2 Fig. 6 shows that indeed the HEC-RAS simulated


LiDAR, topographic contours and SRTM regression models for reach- waterline agrees well with the level of the field-based high
representative remotely sensed waterline modelling, before and after
flow adjustment
water marks. As expected, when assessing the REFIX
waterlines with that of HEC-RAS, the LiDAR gives the
Elevation source Regression model R2
best performance followed by the topographic contours
Initial waterline modelling (no coherence algorithm applied) and the SRTM. The accuracy of remote sensing-based
LiDAR H ¼  0:0009  d þ 225:74 ð5Þ 0.96
waterline modelling according to Eq. (4) is as follows:

Topographic contours H ¼  0:0009  d þ 225:96 ð6Þ 0.92 (i) LiDAR — RMSEREFIX: 0.35 m
(ii) Topographic contours — RMSEREFIX: 0.7 m
(iii) SRTM — RMSEREFIX: 1.07 m
SRTM H ¼  0:0004  d þ 225:16 ð7Þ 0.06

It is worth noting that the initial remotely sensed water


Flow-coherent waterline modelling stages have been modelled to represent a continuous (high
LiDAR H ¼  0:0009  d þ 225:54 ð8Þ 0.98
magnitude) flow for all 76 river cross-sections, which is
assessed with a hydrodynamic model calibrated for the
Topographic contours H ¼  0:001  d þ 225:74 ð9Þ 0.96 Alzette.
For both the topographic contours and the SRTM,
performance values of the remotely sensed waterline are
SRTM H ¼  0:0008  d þ 224:3 ð10Þ 0.89
much higher than the height precision values given in
Section 4.1. This increase is clearly the result of
regression modelling, which has the inherent ability to
coefficient of determination (R2) value. Before coherent smooth out most remote sensing uncertainties (Schu-
modelling, regression explained only 6% of the mann et al., 2007), and the integrated flow coherence
variation of the remotely sensed SRTM water stages. algorithm, which was introduced to adjust incoherent
Increasing model credibility by applying the flow remotely sensed water stages, thereby increasing the
coherence algorithm means the regression now explains credibility of remote sensing-based waterline modelling.
89% of the remotely sensed water stage variation. However, the LiDAR derived DEM shows a slight
Moreover, the average slope of the field-surveyed decrease in performance as a consequence of waterline
channel elevations corresponds to − 0.00086. Initial lowering through the application of the adjustment
waterline modelling shows that a nearly identical algorithm. Initial localised fluctuations due to geo-
remotely sensed waterline slope can be derived from coding errors near higher sloping terrain (see Fig. 5)
both LiDAR and topographic contours but not from have the effect of raising the regression line thereby
SRTM (a = − 0.0004). However, applying coherent getting closer to the field-based high water marks. The
waterline modelling generates a SRTM-derived slope initial remotely sensed water stages from LiDAR
of − 0.0008, which agrees well with the observed (Fig. 5) would not have required any adjustment though,
riverbed slope. as there were hardly any considerable ‘jumps’ (i.e.
Although flow-coherent modelling is not actually fluctuations) in water stages moving downstream. The
required for all three datasets and also lowers the algorithm was only applied for the sake of comparabil-
regression line, it is applied nonetheless to each dataset ity. If waterline modelling with LiDAR is performed
to allow comparison of results. without forcing flow coherence on the data, then the
RMSE of the REFIX model output is 0.25 m and 0.18 m
4.3. Evaluation of remote sensing-based waterline when evaluated with the HEC-RAS model and the field-
modelling based high water marks, respectively (cf. Schumann
et al., 2007).
The calibrated HEC-RAS model has an overall mean These findings further confirm the superiority of the
vertical error of 13 cm when compared to the high water LiDAR instrument to derive useful hydrologic or hydraulic
marks. As the hydraulic model is dynamic and adapts parameter information and suggest its use for high-
more easily to hydraulic structures that change flow precision 3D flood information extraction. For assessment
behaviour than the remote sensing-based steady-state of floods across large areas where lower precisions may be
model, the former is thought a valuable source of sufficient or in areas where high-precision DEMs, such as
information to evaluate the REFIX modelled waterline. LiDAR, are not readily available, the results demonstrate
292 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

that the globally available SRTM DEM coupled with the Due to spatial constraining (i.e. omission of easily
flow coherence algorithm could indeed represent a predictable dry areas outside the area of interest and
valuable source of information to initiate flood disaster areas of flood detection uncertainties on the SAR
management. To investigate this in more detail, a last step image), there is a higher confidence in the goodness of
consisted of mapping actual flood depths and extent from fit between the observed and predicted flood map.
the REFIX modelled water stages. An improved version of the well-used but sometimes
criticised (see Foody, 2002) contingency table-based
4.4. 3D flood mapping Accuracy measure, termed High Probability Fit (HPF),
is proposed that summarises the assessment procedure
The flood mapping procedure was briefly outlined described above:
in Section 3.2. Using the three regression models given
in Eqs. (8)–(10), a water stage for each river cross- 2P 3
section is estimated and used as TIN breakpoints,
n P
n
PIiR1 P1 þ PIiR0 P0
where linear interpolation is applied between cross- 6i¼1 7
HPF ¼ 6
4 P
n
i¼1 7
5 ð11Þ
sections, to generate an event representative remotely
PIi
sensed 3D water surface. An actual flood depth and i¼1 AhpG
extent map is obtained when subtracting the DEM from
the water surface TIN. Individual pixels or areas that
exhibit a negative value, i.e. where the modelled water where i is the pixel index and n represents the total
surface is lower than the terrain height, are excluded. number of pixels inside the area of interest. PIiR 1 P1 and
3D flood maps for each DEM dataset were produced PIiR 0 P0 are contingency table variables denoting whether
(Fig. 7). a binary classified pixel in the predicted map, P, agrees
The validity, reliability and accuracy of each map are with the pixel classification in the reference map, R. The
rather difficult to assess. The level of accuracy can be subscripts 1 and 0 are the binary classification codes for
given according to a contingency table-based measure ‘flooded’ and ‘dry’ pixels, respectively, in the case of
of fit that relates the ASAR-observed flooded area flood mapping. The subscript AhpG is a newly proposed
(Fig. 1a) to that based on remotely sensed water stages spatial constraining variable and denotes areas of high
(Fig. 7). The following flood map assessment procedure probability of agreement with ground observations (i.e.
which compares the DEM-derived flood maps with that areas of substantial remote sensing uncertainties have
observed by the ASAR instrument was used and been omitted).
consists of: According to the HPF measure, the LiDAR-derived
flood map agrees to 75% with that observed by the ASAR
(i) Constraining the area of interest as much as instrument. Both the flood map from topographic
possible to reduce the easily predictable dry areas contours and the SRTM flood map perform similarly
outside the flood extent zone. (62% and 59%, respectively). Although the topographic
(ii) Identifying zones of flood detection uncertainties contour mapping slightly outperforms the SRTM in terms
on the radar image. Uncertainties may be in the of flood area mapping on average, it is worth noting that
form of geo-referencing errors, high-sloping there seems to be more consistency in area mapping in the
terrain that results in geometric distortions case of the SRTM (Fig. 7). This higher consistency of the
(Schumann et al., 2006), wind roughening of the SRTM is directly linked to the way heights are sampled.
water surface and areas of trees and other sig- Contour DEMs are based on topographic contour vectors
nificant vegetation cover that hinder easy flood usually drawn at 10 m height intervals, whereas the
detection (Horritt et al., 2001; Horritt et al., 2003). SRTM, although of only ∼90 m spatial resolution,
Roads and other smooth surfaces that have similar samples heights continuously and thus SRTM DEMs do
specular effects can also be confused with flooded not require interpolation. From this it can be concluded
areas. Zones of uncertain or no flood detection that the SRTM presents a valuable source of important
due to land cover can be identified through aerial hydraulic or hydrologic information for large floodplains,
photography investigation (Horritt et al., 2003). when high-precision datasets are not available.
(iii) Determining accuracy only in areas where classi- However, for adequate 3D flood mapping based on a
fied flood pixels correspond to actual flooding with remotely sensed waterline, two major aspects are of
a very high probability by omitting image zones of importance, namely (i) river gradient and (ii) micro-
flood detection uncertainties. topography obstructing flooding (e.g. protection walls,
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 293

Fig. 7. Flood depth and extent maps based on modelled remotely sensed water stages from (a) LiDAR, (b) topographic contours and (c) SRTM.
294 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

Table 3 were obtained by integrating SAR image flood bound-


Flood mapping capabilities of the different DEMs compared with the aries with elevation data from LiDAR, topographic
HEC-RAS model
contours and SRTM. Prior to regression analysis of
Flood information HEC-RAS LiDAR Topographic SRTM water stages, an algorithm was applied to the data that
contours
ensured decreasing water heights in the direction of the
Area flooded (km2) 2.97 2.51 1.97 2.44 flow. This enabled waterline modelling with an accuracy
Flood mapping 0.73 0.75 0.62 0.59
of 0.35 m for the LiDAR, 0.7 m for topographic contours
performance (HPF)
Maximum flood 5.89 5.69 3.43 4.85 and 1.07 m for the SRTM. The estimated waterline also
depth (m) allowed 3D mapping of important flood parameters, such
Mean flood depth (m) 0.77 0.65 0.90 1.19 as flood area, extent and depth.
N.B. It is worth noting that in contrast to a hydrodynamic model based As expected, LiDAR provided the best results. A
on water volume input and physical laws such as mass conservation surprisingly good performance by the SRTM suggests
and continuity, a static and steady-state remote sensing model based on that this is a potentially valuable source for initial flood
(low resolution) DEMs cannot adequately derive dynamic 3D flood
information extraction in large, topographically homoge-
information such as mean flood depth or, even more so, volume.
neous floodplains, which exhibit a gently sloping river
gradient. Topographic contours did not perform as well as
dykes, levees), and are, therefore, worth considering with expected, as there was insufficient variation in height
respect to the performance of the DEMs reviewed here: within the investigated flood-prone area for the generation
of a reliable DEM.
(i) Low-precision DEMs should only be applied to Subsequent 3D flood mapping from the remotely
areas exhibiting a river gradient large enough for sensed water stages confirms this but also indicates that
coarse DEMs to capture variations in slope. On the flood mapping with lower precision surface elevation data
one hand, if gradients are too low, low-precision is hardly possible at the local scale, as the accuracy of the
DEMs cannot capture the minimal variations in resulting map depends too much on DEM uncertainties
height difference. On the other hand, if gradients and errors in both horizontal (i.e. spatial) and vertical
vary too much, the same limitation is likely to occur. dimensions.
Consequently, erroneous height estimates deleteri- The use of SAR imagery for flood monitoring and
ously affect not only the derived waterline but also assessment is well known and has been successfully
the quality of the subsequently generated 3D flood demonstrated a number of times (e.g. Aplin et al., 1999;
maps (see Table 3 for a quantitative assessment). Horritt et al., 2001; Matgen et al., 2007; Schumann et al.,
(ii) Micro-topography in the form of bridges, dams, 2007). Being able to extract reliable water stages from
dykes, levees or protection walls obstructs flood radar imagery at different scales and thereafter generating
water and impacts considerably on the accuracy of 3D flood maps through the integration of remotely sensed
the 3D flood map product. Given its centimetre flood boundaries with scale-adapted DEMs using the
height precision, it is expected that such features REFIX method presents an inviting complement to the
can only be captured and adequately represented by direct use of radar altimetry and InSAR technologies for
a LiDAR-generated digital surface model (DSM), water stage sensing (Alsdorf et al., 2000). It is clear that in
however this needs to be verified prior to flood contrast to altimetry, which measures water stages directly
depth and area mapping. Another possibility is to and more frequently (the revisit time is several days only
supplement less accurate DEMs with ground- for the JASON missions for example), SAR or other
surveyed geometries of relevant flood obstructing remotely sensed flood boundaries do not provide direct
features. Ultimately, if these geometries are known, measurements of water stages and their revisit time is
they can be used as breakpoints to prevent flooding usually longer (may take up to 35 days in the case of ERS-
to be mapped irrespective of such features. 2 SAR). However, their integration with a DEM through
the application of the REFIX model provides reliable and
5. Concluding remarks accurate spatially continuous water stages of centimetre
accuracy to be used either for the evaluation of flood
Rapid flood information retrieval and 3D flood inundation models or the generation of 2D as well as 3D
mapping is vital for effective flood hazard and risk flood information. It is also worth noting that the current
management. Spatially distributed remote sensing obser- SAR revisit issue may be solved with more timely data
vations can provide such relevant information. Using the acquisition that is achieved within hours of receiving a
REFIX model, water stages from remotely sensed data request in the case of an emergency (Aplin et al., 1999).
G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296 295

It is shown that identifying areas of flow incon- in the accuracy of hydraulic flood modelling. Earth Surfaces
sistencies and addressing these appropriately increases Processes and Landforms 31 (4), 444–456.
Charlton, M.E., Large, A.R.G., Fuller, I.C., 2003. Application of airborne
credibility and confidence in results. It can be said that LiDAR in river environments: the river Coquet, Northumberland,
LiDAR data are at present the most reliable source of UK. Earth Surfaces Processes and Landforms 28 (3), 299–306.
topographic data for precision and small scale remote Cracknell, A.P., Hayes, L.W.B., Keltie, G.F., 1987. Remote sensing of
sensing based waterline modelling. Topographic contour the Tay Estuary using visible and near-infrared data: Mapping of
the inter-tidal zone. Proceedings Royal Society of Edinburgh, 92B,
vectors are of reasonable accuracy when sufficient
pp. 223–236.
elevation heterogeneity exists to allow the generation of Deshmukh, K.S., Shinde, G.N., 2005. An adaptive color image
an area-representative and ‘fit for purpose’ DEM. The segmentation. Electronic Letters on Computer Vision and Image
relatively reliable performance of the SRTM-derived Analysis 5 (4), 12–23.
DEM indicates that it provides a promising new means to Foody, G.M., 2002. Status of land cover classification accuracy
obtain important hydrological or hydraulic parameters for assessment. Remote Sensing of Environment 80 (1), 185–201.
French, J.R., 2003. Airborne LiDAR in support of geomorphological
broader-scale assessment of large catchments. Similar and hydraulic modelling. Earth Surfaces Processes and Landforms
conclusions on SRTM data are drawn by e.g. Ling et al. 28 (3), 321–335.
(2005), LeFavour and Alsdorf (2005), Wang et al. (2005) Gorokhovich, Y., Voustianiouk, A., 2006. Accuracy assessment of the
and Hancock et al. (2006). processed SRTM-based elevation data by CGIAR using field data
The undeniable advantage of the SRTM dataset over from USA and Thailand and its relation to the terrain character-
istics. Remote Sensing of Environment 104 (4), 409–415.
any other source of elevation data is that for the first time Gupta, R.P., Banerji, S., 1985. Monitoring of reservoir volume using
continuously sampled height information is freely Landsat data. Journal of Hydrology 77 (1–4), 159–170.
available at the global scale. This is particularly Haase, D., Frotscher, K., 2005. Topography data harmonisation and
important for PUB studies. However, to derive useful uncertainties applying SRTM, laser scanning and cartographic
information on important hydraulic (model) parameters, elevation models. Advances in Geosciences 5, 65–73.
Haile, A.T., Rientjes, T.H.M., 2005. Effects of LiDAR DEM resolution
such as water stages, remote sensing data uncertainties in flood modelling: a model sensitivity study for the city of
and scaling issues need to be well understood and Tegucigalpa, Honduras. International Archives of Photogramme-
adequately accounted for. In this study, the promising try. Remote Sensing and Spatial Information Sciences 36 (Part 3/
results of the SRTM became only possible after careful W19), 168–173.
adjustment (i.e. confinement to the area of interest and Hancock, G.R., Martinez, C., Evans, K.G., Moliere, D.R., 2006. A
comparison of SRTM and high-resolution digital elevation models
application of the flow coherence algorithm). and their use in catchment geomorphology and hydrology:
Australian examples. Earth Surface Processes and Landforms 31
Acknowledgments (11), 1394–1412.
Hollaus, M., Wagner, W., Kraus, K., 2005. Airborne laser scanning
and usefulness for hydrological models. Advances in Geosciences
This study is supported by the ‘Ministère de la
5, 57–63.
Culture, de l’Enseignement Supérieur et de la Recher- Horritt, M.S., Mason, D.C., Luckman, A.J., 2001. Flood boundary
che’ of Luxembourg. The LiDAR DEM was made delineation from Synthetic Aperture Radar imagery using a
available by the ‘Ministère de l'Intérieur et de l'Amén- statistical active contour model. International Journal of Remote
agement du Territoire’ of Luxembourg and the ‘Admin- Sensing 22 (13), 2489–2507.
istration du Cadastre et de la Topographie (ACT)’ of Horritt, M.S., Mason, D.C., Cobby, D.M., Davenport, I.J., Bates, P.D.,
2003. Waterline mapping in flooded vegetation from airborne SAR
Luxembourg provided the geodesy data used in this imagery. Remote Sensing of Environment 85 (3), 271–281.
study. Kenward, T., Lettenmaier, D., Wood, E., Fielding, E., 2000. Effects of
digital elevation model accuracy on hydrologic predictions.
References Remote Sensing of Environment 74 (3), 432–444.
Lane, S.N., James, T.D., Pritchard, H., Saunders, M., 2003. Photogram-
Alsdorf, D.E., Melack, J.M., Dunne, T., Mertes, L.A.K., Hess, L.L., metric and laser altimetric reconstruction of water levels for extreme
Smith, L.C., 2000. Interferometric radar measurements of water flood event analysis. Photogrammetric Record 18 (104), 293–307.
level changes on the Amazon flood plain. Nature 404, 174–177. LeFavour, G., Alsdorf, D., 2005. Water slope and discharge in the
Aplin, P., Atkinson, P.M., Tatnall, A.R., Cutler, M.E., Sargent, I., 1999. Amazon River estimated using the shuttle radar topography
SAR imagery for flood monitoring and assessment. Proceedings mission digital elevation model. Geophysical Research Letters 32,
RSS99 (Remote Sensing Society) Earth Observation— from Data L17404. doi:10.1029/2005GL023836.
to Information, Cardiff, UK, pp. 557–563. 8–10 September. Ling, F., Zhang, Q., Wang, C., 2005. Comparison of SRTM data with
Berry, P.A.M., Garlick, J.D., Smith, R.G., 2007. Near-global validation other DEM sources in hydrological researches. Proceedings 31st
of the SRTM DEM using satellite radar altimetry. Remote Sensing International Symposium on Remote Sensing of Environment,
of Environment 106 (1), 17–27. June 20–24, St Petersburg, Russia (on CD-ROM).
Casas, A., Benito, G., Thorndycraft, V.R., Rico, M., 2006. The Liu, X., Peterson, J., Zhang, Z., 2005. High-Resolution DEM
topographic data source of digital terrain models as a key element Generated from LiDAR Data for Water Resource Management.
296 G. Schumann et al. / ISPRS Journal of Photogrammetry & Remote Sensing 63 (2008) 283–296

In: Zerger, A., Argent, R.M. (Eds.), International Congress on 7th International Symposium on Spatial Accuracy Assessment in
Modelling and Simulation ʽMODSIM05'. Australia, Melbourne, Natural Resources and Environmental Sciences, July 5–7, Lisbon,
pp. 1402–1408. 12–15 December 2005. Portugal, pp. 633–642.
Ludwig, R., Schneider, P., 2006. Validation of digital elevation models Schumann, G., Hostache, R., Puech, C., Hoffmann, L., Matgen, P.,
from SRTM X-SAR for applications in hydrologic modelling. Pappenberger, F., Pfister, L., 2007. High-resolution 3D flood
Photogrammetry and Remote Sensing 60 (5), 339–358. information from radar imagery for flood hazard management.
MapMart, 2006. World digital elevation model (SRTM). Available at: IEEE Transactions on Geoscience and Remote Sensing 45 (6),
http://www.mapmart.com/DEM/DEM.htm#International (Accessed 1715–1725 Disaster Special Issue.
November 28 2006). Smith, L.C., 1997. Satellite remote sensing of river inundation area, stage,
Matgen, P., Henry, J.-B., Pappenberger, F., de Fraipont, P., Hoffmann, L., and discharge: a review. Hydrological Processes 11 (10), 1427–1439.
Pfister, L., 2004. Uncertainty in calibrating flood propagation models Sun, G., Ranson, K.J., Kharuk, V.I., Kovacs, K., 2003. Validation of
with flood boundaries observed from Synthetic Aperture Radar surface height from shuttle radar topography mission using shuttle
imagery. International Archives of Photogrammetry. Remote Sensing laser altimetry. Remote Sensing of Environment 88 (4), 401–411.
and Spatial Information Sciences 35 (Part 7), 352–357. Walker, J.P., Willgoose, G.R., 1999. On the effect of digital elevation
Matgen, P., Schumann, G., Henry, J.B., Hoffmann, L., Pfister, L., model accuracy on hydrology and geomorphology. Water Resources
2007. Integration of SAR-derived inundation areas, high precision Research 35 (7), 2259–2268.
topographic data and a river flow model toward real-time flood Wang, Y., Liao, M., Sun, G., Gong, J., 2005. Analysis of water
management. Journal of Applied Earth Observation and Geoin- volume, length, total area and inundation area of the Three Gorges
formation 9 (3), 247–263. Reservoir, China using SRTM DEM data. International Journal of
Rabus, B., Eineder, M., Roth, A., Bamler, R., 2003. The shuttle radar Remote Sensing 26 (18), 4001–4012.
topography mission — a new class of digital elevation models Webster, T.L., Forbes, D.L., Dickie, S., Shreenan, R., 2004. Using
acquired by spaceborne radar. Photogrammetry and Remote topographic lidar to map flood risk from storm-surge events for
Sensing 57 (4), 241–262. Charlottetown, Prince Edward Island, Canada. Canadian Journal of
Raclot, D., 2006. Remote sensing of water levels on floodplains: a Remote Sensing 30 (1), 64–76.
spatial approach guided by hydraulic functioning. International Wechsler, S., 2007. Uncertainties associated with digital elevation
Journal of Remote Sensing 27 (12), 2553–2574. models for hydrologic applications: a review. Hydrology and Earth
Rodriguez, E., Morris, C.S., Belz, J.E., Chapin, E.C., Martin, J.M., System Sciences 11, 1481–1500.
Daffer, W., Hensley, S., 2005. An assessment of the SRTM Wilson, M.D., Atkinson, P.M., 2003. A comparison of remotely sensed
topographic products. Technical Report JPL D-31639. Jet Propul- elevation data sets for flood inundation modelling. Proceedings 7th
sion Laboratory, Pasadena, California. 143 pp., http://www2.jpl. International Conference on GeoComputation, September 8–10,
nasa.gov/srtm/srtmBibliography.html (Accessed October 2, 2007). University of Southampton, Southampton, UK (on CD-ROM).
Schumann, G., Black, A., Cutler, M., Henry, J.B., Hoffmann, L., Yin, Z.-Y., Wang, X., 1999. Technical communication. A cross-scale
Matgen, P., Pfister, L., 2006. Hydraulic and event knowledge to comparison of drainage basin characteristics derived from digital
reduce the positional uncertainty in SAR flood images for elevation models. Earth Surfaces Processes and Landforms 24 (6),
improved flood model calibration and development. Proceedings 557–562.

You might also like