1. Introduction
Floods account for 43% of all natural disasters [
1] that cause death, adverse health effects, property damage, and socio-economic imbalance [
2,
3,
4]. By 2001, flooding became seven times more frequent around the world compared to 1975. From 1975 to 2001, the number of people at risk and fatalities due to floods increased by nine times [
5]. The increasing trend of flood disasters continued into the 21st century as well. Between 1995 and 2015, floods affected 2.3 billion people globally and killed 157,000 [
1]. Increasing flood frequency, along with unplanned urbanization, population growth, and resource limitation is making flood risk management more challenging [
6,
7]. However, two-thirds of flood-related economic losses remain unreported [
8]. This data gap is more prevalent in developing countries than in developed ones. Therefore, it is a global concern because sustainable flood risk management (FRM) attempts to combine the knowledge from past events and the existing situation to better prepare for the future. Flood impact assessment (FIA) is an important component of FRM that estimates the direct and indirect consequences of a flood event. Necessary flood parameters (i.e., depth, extent, velocity, duration) for consequence assessment are commonly derived from hydrodynamic modeling.
The accuracy of both hydrodynamic modeling and FIA are subject to accuracy and availability of data (i.e., hydrologic, terrain, land use, infrastructure). For instance, Bhuyian and [
9] showed that flood consequence estimates could be affected by the elevation error (hereafter called as “error”) and spatial resolution of the Digital Elevation Model (DEM). Root mean square error (RMSE) for some of the commonly used DEMs such as National Elevation Dataset (NED), Shuttle Radar Topography Mission (SRTM), and Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) version 2 are 1.84, 4.01, and 8.68 m respectively at one-arc second spatial resolution [
10]. Global DEMs (i.e., SRTM and ASTER) often have a higher error but are suitable for remote areas for their wider coverage. However, SRTM is a better choice for flood modeling in such areas, considering the low RMSE and recently (the year 2015) published one-arc second products for areas outside of the USA. Therefore, being able to reduce errors in SRTM DEM would help to improve FIA in remote and data-poor areas.
The primary sources of error in a spaceborne DEM (i.e., SRTM) are speckle noise, stripe noise, absolute bias, and tree height bias, in addition to the inability of mapping river bathymetry [
11,
12,
13,
14]. The first four error sources primarily affect hydraulic connectivity within the floodplain, and inaccurate river bathymetry affects the river conveyance. Spaceborne, regional and local DEMs all suffer from the limited representation of river bathymetry. This is often supplemented by overlaying surveyed river cross sections on a reach scale. However, this approach is unfeasible on a regional to global scale [
13]. There have been multiple methods in practice to correct DEM for better representation of river conveyance in the absence of surveyed cross sections. The existing methods are either based on a prismatic channel assumption or limited to only water surface elevation (WSE) prediction [
14,
15,
16,
17]. However, river planform (i.e., channel alignment, width, curvature, etc.) itself is dynamic [
18], which may dictate the flow parameters (i.e., stage, flow, inundation, velocity) and, hence, influence flood consequences. Considering the existing approaches, [
19] proposed a conveyance-based DEM correction technique combining river planform, geometry, and historical stage-discharge relation to generating modified bathymetry along defined cross-sections. This technique incorporates channel conveyance along with the location of thalweg in parabolic shaped cross-sections. They were able to implement the DEM correction technique on a 1/3 arc-second NED DEM to perform hydrodynamic simulations of historic flood events. However, extra effort may be required if DEM correction is intended on a global DEM such as SRTM, because of high error (channels and floodplains) and moderate resolutions of global DEMs. Moderate-resolution DEMs are often preferred for ease of data handling or lack of fine resolution data for FIA. Therefore, it is important to estimate the extent to which conveyance-based DEM correction technique can perform using moderate resolution regional or global DEM.
Another motivation for exploring errors in global DEMs is to aid the future satellite remote sensing missions expected to map indirectly floodplain elevations on a global scale. One such mission is the proposed Surface Water Ocean Topography (SWOT) mission [
20]. SWOT aims to provide an understanding of the behavior of river water elevation that is 100 m wide against the backdrop of human impacts on the water cycle (i.e., global warming, river regulation, etc.). Although SWOT’s signal-to-noise ratio for land topography is expected to be poor, it is expected that many years of repeatedly using SWOT to make elevation observations of water inundation extent in dynamic rivers can help create a globally consistent “floodplain DEM.” This can be used to aid the community’s existing efforts on flood inundation already being done through satellite imagery [
21,
22]. Thus looking for possible solutions via linking remotely sensed river water elevation (i.e., SWOT) with existing DEM correction techniques would help reducing dependence on in-situ stage measurements.
The research questions presented in this study were (i) what are the limitations of implementing a conveyance-based DEM correction on a moderate resolution (one-arc second) regional (i.e., NED) or global DEM (i.e., SRTM)? (ii) What is the corresponding impact in FIA after applying a conveyance-based DEM correction technique to the aforementioned DEMs? These questions were addressed through the following objectives:
2. Methodology
The scope of this study was limited to correcting one moderate resolution regional DEM and one moderate resolution global DEM dataset for FIA of the 2010 Nashville flood. The uncorrected DEMs are hereafter called “base” DEMs. This study applied the DEM correction technique by [
4] on the base DEMs to quantify the relative improvement in FIA. DEMs are called “modified” DEMs after applying the aforementioned DEM correction. The base and modified DEM datasets are collectively called “test” terrain datasets. A “control” terrain dataset was produced from the best available terrain dataset at disposal. Models based on control and test terrain datasets are hereafter named “control scenario” and “test scenarios.” The relative accuracies of FIAs derived from test scenarios were compared to the control scenario. A summary of the methodology is presented in the next sub-sections along with a flowchart (
Figure 1).
2.1. Control and Test Terrain Datasets
First, the control terrain dataset was produced by merging field-surveyed river bathymetry and LiDAR floodplain topography. Second, base DEMs for the study area were collected from the U.S. Geological Survey (USGS) Earth Explorer. Third, the DEM correction technique as described by [
19] was employed to correct the base DEMs. This correction technique works on predefined cross-section alignments and requires prior knowledge of the longitudinal slope, stage-discharge relation and flow roughness of the study reach. Four major steps of implementing this DEM correction technique are briefly described below.
2.1.1. Raster Processing
The river centerline and bank lines were traced over a Landsat satellite image. The cross-section alignments for DEM correction were kept same as the surveyed ones. The point with the steepest gradient (slope) on each side of the bank was considered as high bank points (hbp) for respective sides.
2.1.2. Hydrologic Processing
The flow corresponding to the DEM elevation at a nearby stream gage is called the “reference flow.” Reference flows corresponding to each DEM source were estimated from a stage-discharge plot of the nearby stream gage.
2.1.3. Three-Point Cross-Section Generation
The longitudinal slope of the river was estimated for each DEM source (NED and SRTM) and was calculated via linear correlation of center point elevations (from base DEMs) with respect to their chainage. A reasonable Manning’s roughness factor for the entire study reach was selected as an input for the DEM correction technique. Thalweg locations were estimated by channel width and side slope. Thalweg elevation for each cross-section was calculated via solving Manning’s equation utilizing the reference flow, longitudinal slope, and the Manning’s roughness factor.
2.1.4. Raster Modification
Modified elevation for points along the cross-sections were calculated assuming parabolic cross-sections. The point elevations were then interpolated to produce bathymetry corrected (along with the predefined river cross-sections) raster. The values of base DEM along the river were then replaced by the values from the corrected raster to produce modified DEM.
2.2. Flood Event
A historic flood event was selected for which inundation extent was readily available.
2.3. Hydrodynamic Simulation
To perform the comparative FIA, a one-dimensional hydrodynamic model was set up using the Hydrologic Engineering Center’s River Analysis System (HEC-RAS). Input hydrologic data for the hydrodynamic model was collected at available stream gages within the study area. The hydrodynamic model was calibrated for the control scenario. Manning’s roughness factor was selected as the calibration parameter while keeping the downstream flowing at a normal depth. The simulated flood extent was also compared to an observed flood inundation map to validate the performance of the control scenario. The calibrated hydrodynamic model parameters (i.e., Manning’s roughness factor) were used to simulate the test scenarios. Corresponding flood parameters (i.e., stage, flow, inundation extent) were used for FIA.
2.4. Flood Impact Assessment
The Hydrologic Engineering Center’s Flood Impact Assessment (HEC-FIA) was selected to estimate flood consequences for the study event. The ease of coupling HEC-RAS model outputs with HEC-FIA played a key role in selecting these platforms for this research. For each scenario, the geospatial information (river alignment, impact area) and simulation period were kept the same as the corresponding hydrodynamic simulation. The warning times for each scenario were also assigned, corresponding to the flood stage at a flood-warning gage within the study reach. Simulated flood parameters (maximum inundation extent, maximum depth and arrival time) for individual scenarios were furnished as the primary input for FIA. Local inventories such as infrastructures, administrative areas, and agricultural use were generated from the Hazus-MH 2.1 database. Flood consequences estimated via a control scenario were used as the baseline for comparison with test scenario-derived estimates. It was assumed that the control terrain data would best agree with the field condition because that was derived from the best available data (i.e., field-surveyed data and LiDAR).
5. Discussion
A conveyance-based DEM correction technique was applied on one arc-second NED and SRTM DEMs. The absence of adequate conveyance in base DEMs signifies that they are only useful in simulating flows for a very narrow window in a hydrologic year irrespective of the DEM quality. This is because in most of the days in a hydrologic year, the actual river WSE remains below the elevation present in an uncorrected (base) DEM. Therefore, the simulated stages (using uncorrected DEM as terrain) would have high error for instances when the flow remains mostly within the channel. Thus, the uncorrected DEM based models would be only applicable for narrow spans in a hydrologic year (very high flow instances).
DEM errors not only affect the channel conveyance (in DEM) but also affect other hydraulic properties (i.e., longitudinal slope, bank features). As shown in
Figure 1, the DEMs failed to show the presence of levees accurately. It could be because of the coarseness of the spatial resolution of DEM, or inadequate data acquisition mechanism or timing of DEM production. DEM only represents the terrain, as it was at the time of data acquisition; hence any change on the ground surface made afterward would be absent in a DEM. Additionally, river longitudinal channel slope is often calculated from DEM elevations along the river [
25,
26]. It was observed that longitudinal channel slope (over entire study reach) calculated from base NED was more reliable than that of base SRTM. This affected implementation of the DEM correction technique on SRTM because longitudinal channel slope is a primary input in the selected DEM correction technique. The accuracy of SRTM DEMs immediately upstream and downstream of a dam within the entire length (about 1110 km from headwaters to the Ohio River confluence) of the Cumberland River was found to be limited. Non-void filled SRTM (V1) also exhibited similar pattern.
Figure 8 shows the SRTM elevations and average WSE on streamflow gages for (a) the entire river length (1110 km) of the Cumberland River, and (b) a stretch from OHD to the Cheatham Navigation Lock (CNL). Abrupt elevation changes similar to the U.S. stretch (refer to
Figure 4b) were also observed downstream of Cordell Hull and Wolf Creek Dams. This indicates the phenomena in U.S. stretch could be due to its location relative to OHD. Moreover, reservoirs also affected the accuracy of SRTM in reaches upstream of a dam. Therefore, the representative segment was selected such that it had least structural influence but long enough to exhibit the downward slope. For the SRTM-derived channel slope in this study, a 22.5 km stretch (D.S. stretch) was considered adequate because the distance from OHD to CNL was 107 km where structural presence had a significant influence on a more than 60 km river length. However, channel slope could also be estimated from streamflow gages, but in a data-poor area, that scope could be limited, leaving global DEM (i.e., SRTM) as the only viable option. Nevertheless, the agreement of the estimated channel slope from SRTM DEM and surveyed (control) bathymetric data (0.000077 and 0.000083) showed that longitudinal channel slope could be calculated via a representative segment.
Interpolation of corrected bathymetric elevation along predefined cross-sections at intervals equal to the DEM spatial resolutions (as prescribed in [
19] for 1/3 arc-second NED) seemed inadequate for one arc-second DEMs. Estimating modified channel bathymetry via interpolating corrected bed elevation at intervals (~3 m) smaller than the DEM spatial resolution (one-arc second or ~30 m) was useful in the given river reach (average width 180m). It was also challenging to find the exact transition points from floodplains to channel (referred to as “high bank points” or “hbp”) using one arc-second DEMs. The assistance of corresponding satellite images may help in more accurately finding the hbp if the ratio of average river width to DEM cell size (spatial resolution) is small. Nevertheless, implementation of the DEM correction technique with additional measures (i.e., estimating longitudinal channel slope from the representative segment, updating the interpolation method) successfully reduced the errors in base DEM elevations over the channel. After correcting the DEM, the RMSE in thalweg elevations (in 68 cross-sections) dropped to 1.59 m for NED, and 1.67 m for SRTM from 9.66 m to 11.46 m respectively. The resulting error in thalweg elevation was also lower than the RMSE of NED and SRTM (1.84 m and 4.01 m), estimated over multiple land cover types [
10]. Additionally, the longitudinal channel slope in SRTM DEM increased from 0.000012 m/m to 0.0007 m/m where the longitudinal channel slope in control terrain dataset was 0.0008 m/m. On the contrary, the longitudinal slope in NED DEM has slightly decreased from 0.0007 m/m to 0.0006 m/m. This may have happened because the base NED-derived longitudinal profile demonstrated a step-like drop instead of decreasing gradually. The change was very small along the river length, except for sudden drops of about 0.5 m at each 10 to 15 km. This phenomenon in the base NED DEM would affect the accuracy of simulating low stages in a flood event. Nevertheless, the longitudinal slopes from both modified DEM types were close to their respective representative slopes and the one derived from the control dataset, indicating the sub-reach longitudinal channel slope can be effectively used in the discussed DEM correction technique.
The performance of the simulated WSE from control and test scenarios with respect to observed data (
Figure 6b) shows that the mean errors (in simulated WSE) were 4.9 m and −2.3 m for base NED and base SRTM DEM-based test scenarios that dropped to 1.5 m and −0.7 m respectively, upon DEM correction. Similarly, the RMSE in simulated WSE dropped to 1.7 m and 1.1 m from 5.3 m and 4.1 m respectively. This shows that the DEM correction was able to cut the errors in simulated WSE up to 65%.
The errors in simulated WSE were reflected in the corresponding simulated flood extents, as shown in
Figure 7b. High biases in the simulated WSE from the base NED-derived scenario produced an inundation extent 33% larger than the observed one. However, the simulated inundation extent from the base SRTM-derived scenario was scattered and resulted in a very large inundation extent. It was also noticeable that the error was high above the DS stretch where simulated flood extents spread over wide swaths of the floodplain. Inconsistencies in longitudinal channel slope along with lack of channel conveyance may have attributed to this localized error. Nevertheless, DEM correction was able to reduce the error in simulating flood extent. The simulated inundation extent from the modified NED derived test scenario resulted in an overestimation by 7%, and the modified SRTM-derived test scenario underestimated by 19%.
Figure 9a shows a bar plot inundation extent from different scenarios, and
Figure 9b plots the inundation extents with respect to their peak WSE above the NWS flood stage at Nashville (122.22 m). The dotted line in
Figure 9b shows a correlation of flood extents with peak flood stages, excluding the base-SRTM derived scenario. The SRTM-based scenario fits into the pattern (with control, observed and NED) only after DEM modification. It indicates that for a particular area if a hydrodynamic simulation is stable, then the resulting stage-inundation extent should fit in a specific pattern data irrespective of the DEM source. However, this study only included one arc-second DEMs; therefore, the sensitivity of the correlation to change in spatial resolution of DEM was not investigated.
The flood extents derived from HEC-RAS one-dimensional modeling was considered adequate for this study because the comparative FIA (for four test scenarios) was done on similar model (hydrodynamic and FIA) setups. The accuracy of inundation extent obtained from control scenario (
Figure 7a) also acted as a benchmark for subsequent FIAs (in test scenarios). HEC-FIA only considers the maximum inundation extent and maximum flood depth; thus, detailed assessment of spatial distribution of flow and velocity (as in a two-dimensional model) was out of scope for this research. Additionally, a one-dimensional model is often preferred when data and computation capacity are limited [
27,
28].
The errors and uncertainties in hydrodynamic simulations were carried over to the consequence simulations (FIA) accordingly.
Figure 10 shows a number of structures affected, and total monetary loss estimated from control and test scenarios. It is interesting to find that both of these loss indexes fit very well with the corresponding inundation extent even though the base SRTM-based scenario was unstable. This finding is similar to the findings of [
9]. However, the instability of the base SRTM-based model indicates that the flood consequence estimates are not entirely dependent upon the accuracy of the hydrodynamic simulation, but rather, it is dependent upon the accuracy of inundation extent and its occupancy. Nevertheless, the DEM correction successfully reduced the error in estimated flood consequences in both NED and SRTM-based scenarios. Underestimation in peak stage (refer to
Figure 9b) by the modified SRTM-based scenario resulted in an underestimation of inundation extent and consequences. This could be partially caused by the errors in floodplain elevation (in DEM) that produced scattered inundation extents in addition to the lack of hydraulic properties (i.e., conveyance, slope).
6. Conclusions
The first objective of this research was to implement a conveyance-based DEM correction technique [
19] on moderate resolution NED and SRTM DEMs. One arc-second spatial resolution was selected for both DEM sources because of data availability, convenience in data handling, and optimum simulation accuracy. The intention was to investigate challenges to implement a conveyance-based DEM correction technique on moderate resolution regional and global DEMs. The DEM correction technique by [
19] showed promising results in incorporating channel conveyance into DEM-derived cross-sections for both NED and SRTM DEMs. This technique depends on some hydraulic properties (i.e., overbank elevations, side slopes, longitudinal channel slope) to characterize river cross-section shape and predict the depth of the channel thalweg that is often difficult to extract from global DEMs such as SRTM DEM. Therefore, the efficacy of the DEM correction depends upon the accuracy of DEM itself. This research has demonstrated that representative hydraulic properties (i.e., longitudinal channel slope) from a smaller segment may be used to implement DEM correction over the entire reach if the quality of DEM is compromised. Additionally, the DEM correction technique used a second-order polynomial equation to generate elevation for locations between thalweg and high-bank points. Due to a limitation in interpolation methods, coupled with the coarser spatial resolution of DEM (one-arc second), the entire conveyance was not possible to be captured, which may eventually have introduced error in the modified DEM-based scenarios. Therefore, estimating modified bathymetric elevation at a spacing smaller than the DEM spatial resolution proved helpful for the selected DEMs. The above-mentioned extensions (i.e., longitudinal slope estimation and spatial interpolation) to the selected DEM correction technique were especially helpful in correcting the SRTM DEM, highlighting the extra effort needed for correcting global DEMs. Nevertheless, the estimated river elevations via the technique in both one arc-second NED and SRTM DEM were comparable (
Figure 8).
The second objective of the research was to estimate the impact on FIA after applying the conveyance-based DEM correction technique on selected NED and SRTM DEMs. NED was found to better represent the terrain compared to SRTM, and hence produced less error in FIA. It was found that NED DEM had errors mostly along channels, but SRTM DEM had errors both along the channel and over floodplains. Therefore, simulated WSE amplified at variable degrees as they were used for inundation mapping and subsequent FIA. Flood loss is calculated as a function of flood depth and flood wave [
29,
30]. Therefore, having a DEM that not only aids in estimating WSE properly, but also produces accurate depths on floodplains is necessary for FIA. However, it was found that DEM correction could improve the FIA by both stabilizing model performance and reducing errors. For a set of stable simulations, the consequences have been found to follow a specific pattern similar to the flood loss curve defined by [
19]. It indicates that a hydraulic-based DEM correction technique can be applied on SRTM (in the case of data scarcity) to produce reasonable flood consequence estimates. However, it should be noted that the conveyance correction of DEMs over the channel does not improve the floodplain errors. Additionally, reducing reliance on the DEM-derived hydraulic property or field data is necessary for correcting DEMs over data-limited areas. Hence, an extended study on improving channel conveyance with minimum use of DEM-derived or field-surveyed input can be a potential future field of research.
Finally, the data gap not only pertains to limited access to ground surveyed data, but also includes the unavailability of up-to-date data. Many of the global DEMs are ageing, and may soon become irrelevant in featuring actual ground conditions. Hence, a hydrologically well-gauged area may become data-poor in terms of limited bathymetric data and floodplain terrain data due to rapid urbanization, modification of water routes, change in river planform, aggradation-degradation of riverbed levels etc. However, undertaking extensive field survey on a regular basis is often unfeasible. Therefore, a combination of conveyance-based (for rivers and streams) DEM correction with floodplain-based DEM corrections would be especially helpful for FIA in remote, data-poor and rapidly growing areas. Additionally, strong correlation of flood damage with inundated area indicates FIA can also be made via remotely sensed flood extent if flood extent versus damage correlations (flood loss curve) are established for a given area.
7. Future Research
DEM is a raster image that represents the terrain only for the instance of data acquisition. However, river morphology is not static. Therefore, the dynamics of riverine morphology should be considered while a DEM is used for river FIA. A possible track of future research can include adopting a DEM with the temporal variability of riverine morphology and floodplain dynamics. Studies have shown that combining time-series images with remote sensing data (i.e., elevation, land use, vegetation, etc.) may be helpful for analyzing river planform, bathymetry of water body (i.e., lake, stream) and floodplain characteristics [
12,
31,
32,
33,
34]. Upcoming remote sensing missions such as SWOT can be immensely helpful in this regard, since it will provide seasonal water surface musk at a global scale [
22]. A data-driven algorithm that can compile data from multiple sources (i.e., SWOT, Landsat, etc.) to observe seasonal variation for a long period would be helpful in producing a better representation of river bathymetry and floodplain elevation. Additionally, it would reduce the reliance on field data (i.e., stage-discharge relation, channel roughness, channel shape, etc.) for implementing conveyance-based DEM correction techniques.