Volume 55, Issue 10 p. 7983-8009
Research Article
Free Access

Terrain Analysis Enhancements to the Height Above Nearest Drainage Flood Inundation Mapping Method

Irene Garousi-Nejad

Corresponding Author

Irene Garousi-Nejad

Department of Civil and Environmental Engineering, Utah Water Research Laboratory, Utah State University, Logan, UT, USA

Correspondence to: I. Garousi-Nejad,

[email protected]

Search for more papers by this author
David G. Tarboton

David G. Tarboton

Department of Civil and Environmental Engineering, Utah Water Research Laboratory, Utah State University, Logan, UT, USA

Search for more papers by this author
Mahyar Aboutalebi

Mahyar Aboutalebi

Department of Civil and Environmental Engineering, Utah Water Research Laboratory, Utah State University, Logan, UT, USA

Search for more papers by this author
Alfonso F. Torres-Rua

Alfonso F. Torres-Rua

Department of Civil and Environmental Engineering, Utah Water Research Laboratory, Utah State University, Logan, UT, USA

Search for more papers by this author
First published: 24 July 2019
Citations: 46

Abstract

Flood inundation remains challenging to map, model, and forecast because it requires detailed representations of hydrologic and hydraulic processes. Recently, Continental-Scale Flood Inundation Mapping (CFIM), an empirical approach with fewer data demands, has been suggested. This approach uses National Water Model forecast discharge with Height Above Nearest Drainage (HAND) calculated from a digital elevation model to approximate reach-averaged hydraulic properties, estimate a synthetic rating curve, and map near real-time flood inundation from stage. In 2017, rapid snowmelt resulted in a record flood on the Bear River in Utah, USA. In this study, we evaluated the CFIM method over the river section where this flooding occurred. We compared modeled flood inundation with the flood inundation observed in high-resolution Planet RapidEye satellite imagery. Differences were attributed to discrepancies between observed and forecast discharges but also notably due to shortcomings in the derivation of HAND from National Elevation Dataset as implemented in CFIM, and possibly due to suboptimal hydraulic roughness parameter. Examining these differences highlights limitations in the HAND terrain analysis methodology. We present a set of improvements developed to overcome some limitations and advance CFIM outcomes. These include conditioning the topography using high-resolution hydrography, dispersing nodes used to subdivide the river into reaches and catchments, and using a high-resolution digital elevation model. We also suggest an approach to obtain a reach specific Manning's n from observed inundation and validated improvements for the flood of March 2019 in the Ocheyedan River, Iowa. The methods developed have the potential to improve CFIM.

Key Points

  • Comparison of flood inundation mapped using Height Above Nearest Drainage (HAND) to inundation observed by Planet high-resolution imagery
  • Improvements in HAND flood inundation mapping by conditioning the underlying digital elevation model using high-resolution hydrography
  • Potential to use satellite observed inundation to infer distributed hydraulic roughness parameters for HAND-based hydraulic routing

Plain Language Summary

Flood inundation is difficult to map, model, and forecast because of the data needed and the computational demand. Recently an approach based on the Height Above Nearest Drainage (HAND), which is derived from a digital elevation model, has been suggested for both flood mapping and obtaining reach hydraulic properties. This approach was tested for a recent snowmelt flood on the Bear River and compared to inundated area mapped using Planet RapidEye satellite imagery. The identified differences were reduced by addressing shortcomings in the HAND terrain analysis evaluation, in terms of both the digital elevation model resolution and the method used to condition the digital elevation model using streamline information.

1 Introduction

Floods are responsible for billions of dollars of damage and loss of life worldwide (Parker, 2017), and much hydrologic research has focused on improving the ability to predict and model flood inundation, prepare for and mitigate flood effects, and warn people at risk. Mapping and modeling flood inundation extent with high precision is challenging because it requires a comprehensive description of computationally demanding and data limited hydrologic and hydraulic processes. Satellite observations of the inundated area, as well as detailed digital elevation model (DEM) topography, offer the opportunity to examine and improve currently available flood inundation mapping methods.

In February 2017, a flood occurred in Box Elder County, Utah, USA. A combination of rainfall and warmer temperatures caused significant snowmelt in northern Utah, resulting in record flows for this time of year in the Bear River. The 2017 discharge was the third largest on record (1952–present) and the largest since 1987. Here we use this event as a case study to evaluate and develop improvements in empirical methods for Continental-Scale Flood Inundation Mapping (CFIM) based on the Height above Nearest Drainage (HAND). Following detailed analysis of this case study, we validated the improvements we developed for another flood that occurred in March 2019 in the Ocheyedan River near Spencer City in Iowa, USA.

In contrast to comprehensive hydraulic models, empirical approaches offer alternatives that have fewer data demands and perhaps offer a more practical alternative for generating flood inundation maps. Researchers such as Rodda (2005) started to incorporate DEM information and grid cell based operations to calculate flood depth and then the probability of insured losses from floods. Rennó et al. (2008) redesigned Rodda's (2005) concept and introduced a terrain descriptor called HAND, which uses DEMs to define the height of each grid cell on the land surface above the cell in the nearest stream to which the drainage from that land surface cell flows. Thereafter, researchers applied HAND as a descriptor to determine soil water potential (Nobre et al., 2011), groundwater potential (Rahmati et al., 2018), and flood potential (Nobre et al., 2016).

To calculate HAND, a hydrologically conditioned DEM and a representation of the DEM flow field are required. A hydrologically conditioned DEM is one for which internally draining areas have been removed (or true internally draining areas are marked and retained), and each grid cell can drain following a non-increasing elevation path to the edge of the DEM (or true internally draining sink). Pit filling (Jenson & Domingue, 1988), breaching (Soille et al., 2003), and hybrid filling-breaching algorithms to hydrologically condition DEMs have been developed (Lindsay & Creed, 2005; Martz & Garbrecht, 1999; Soille, 2004). Also, recent work has advanced hydrography-driven coarsening to retain hydrographic fidelity in a high-resolution DEM when the DEM needs to be reduced for computational reasons (Moretti & Orlandini, 2018).

To represent the flow field, the earliest method is the D8 single-flow direction model initially proposed by O'Callaghan and Mark (1984). In this method, flow moves from each grid cell to one of its eight neighbors along the steepest downward slope direction. This method limits the precision with which flow direction is represented and introduces grid bias (Costa-Cabral & Burges, 1994; Fairfield & Leymarie, 1991; Tarboton, 1997).

Considerable work has been done to overcome D8 limitations (Freeman, 1991; Orlandini et al., 2003; Orlandini et al., 2014; Orlandini & Moretti, 2009; Quinn et al., 1991; Seibert & McGlynn, 2007; Tarboton, 1997). Among these alternative approaches is the D (D-infinity) multiple-flow direction model developed by Tarboton (1997). The D model shares the flow from a grid cell between two adjacent down slope grid cells based on flow direction angle proportioning.

Taking advantage of the D flow model, Tesfa et al. (2011) presented an algorithm that uses DEMs to derive a set of hydrological proximity measures that include distances up from or down to target grid cells. These distances may be evaluated horizontally or vertically using the D flow model with weighted averaging applied where flow is proportioned between multiple grid cells. The D distance down function, calculated vertically with the target grid cells being stream grid cells, may be used to calculate HAND (Zheng, Maidment, et al., 2018). This D distance down evaluation of HAND differs from earlier HAND approaches (Nobre et al., 2011; Nobre et al., 2016) that use D8. The D averaging across multiple-flow paths results in a smooth HAND surface and avoids step-like discontinuities that can result from using D8. Tesfa et al. (2011) described an efficient parallel implementation of these calculations that is available as part of the TauDEM software (Tarboton, 2017).

Liu et al. (2016) implemented the D distance down function to generate a HAND raster for the continental United States using a high-performance computer at the University of Illinois CyberGIS facility. Given this HAND raster, Zheng, Tarboton, et al. (2018) proposed a method to compute the reach-averaged channel geometry properties and estimate a synthetic rating curve to relate flow to water level in a stream reach. The Zheng, Tarboton, et al. (2018) method is an alternative to traditional approaches, where river geometry properties are computed using surveyed river cross sections at points along the stream reach that omit terrain detail between cross sections. This approach also avoids the intensive data requirements for river cross-section-based hydraulic modeling. At the same time, Liu et al. (2018) designed a workflow based on Liu et al. (2016) and Zheng, Tarboton, et al. (2018) for CFIM using discharge forecasts from the National Oceanic and Atmospheric Administration National Water Model (NWM). In this method, NWM discharges are converted to stage using a rating curve, and then the stage is used to map inundation based on HAND. Each reach in the National Hydrography Dataset (NHDPlus medium resolution; McKay et al., 2012) has a separate forecast discharge, and HAND is used to evaluate rating curves and map inundation for the NHDPlus catchment draining into each reach. In another recent study, Zheng, Maidment, et al. (2018) implemented the HAND approach with high-resolution topographic data derived from light detection and ranging using a geodesic minimization technique to map the streams (Passalacqua et al., 2010; Sangireddy et al., 2016). This approach maps streams by selecting a flow path to minimize a metric that is a combination of contributing area, curvature, and distance from NHDPlus (medium resolution) streams. The results showed that the inundation extent produced by their approach, called GeoFlood, is able to capture the Federal Emergency Management Agency flood plain coverage with 60–90% overlap accuracy. Given the potential for using the HAND method to map flood inundation over large areas based on available DEM data and discharge forecasts from the NWM, it is important to evaluate the performance of the approach in many different locations and settings.

Satellite-based remote sensing is a useful source for evaluating modeled flood inundation mapping because of its ability to provide a synoptic perspective across a large range of scales and resolutions. Remote sensing offers a practical solution to observing the location and extent of inundation for many flooding events (Policelli et al., 2017). Recent deployments of small CubeSat satellites by companies such as Planet (2017) advance hydrological remote sensing by providing an unprecedented combination of high-temporal- and high-spatial-resolution imagery at the global scale (Cooley et al., 2017). Remote sensing for river discharge is a topic receiving increasing attention, for example, Surface Water and Ocean Topography (SWOT) satellite mission, https://swot.jpl.nasa.gov/hydrology.htm (Biancamaria et al., 2010; Tourian et al., 2017). In particular, Sichangi et al. (2016) used Moderate Resolution Imaging Spectroradiometer (MODIS) and satellite altimetry to estimate river discharge and also suggested its use in rating curves and hydraulic parameter optimization. Overall, there is broad potential for applying remote sensing of surface properties, including inundation mapping and monitoring in hydrological research (Cooley et al., 2017; McCabe et al., 2017).

In this study, we evaluate the performance of CFIM in terms of its accuracy in representing flooded and nonflooded areas when comparing the results with inundation observed by high-resolution Planet RapidEye Satellites. We first use the HAND data and rating curves from Liu et al. (2018) to compute flood inundation based on a published methodology (Liu et al., 2018; Zheng, Tarboton, et al., 2018). Then, we examine the causes of discrepancies between mapped and observed flood inundation and develop and evaluate important improvements to the published CFIM method. These improvements include (1) dispersing nodes approximately uniformly along the reach to avoid the sometimes small and irregular-sized NHDPlus catchments that degrade synthetic rating curve estimation; (2) using a high-resolution hydrography (NHD high-resolution, i.e., 1:24,000) to condition the DEM and breach DEM barriers, often due to roads; and (3) using a high-resolution (i.e., 1/9th arc-sec [3 m]) DEM that is available for this area. We apply these sequentially, and at each step quantify the improvement in flood inundation mapping fidelity. We also evaluate the stage that best matches HAND-based flood inundation with observed inundation and use this to quantify inherent uncertainty in this evaluation and infer reach specific information about Manning's n that may minimize errors introduced with this approach. This opens an opportunity for direct estimation of reach specific Manning's n from observed inundation, overcoming the current CFIM limitations of using a single Manning's n everywhere. This builds on ideas from Sichangi et al. (2016).

The following section describes the study site and the February 2017 flood that occurred in the Bear River, Utah, USA. The methodology section describes the CFIM procedure for generating a flood inundation map based on HAND and introduces improvements to address the identified CFIM challenges. Next, we present the workflow for creating a flood inundation map based on RapidEye imagery, along with metrics to compare HAND-based flood inundation maps with high-resolution satellite imagery. The results section compares HAND-based flood inundation from CFIM with RapidEye observations and then sequentially quantifies the improvement due to each innovation introduced. Thereafter, results for a validation case study (a March 2019 flood that occurred in the Ocheyedan River, Iowa) based on both the published CFIM method and developed improvements are presented and compared with the observed flood inundation. We conclude with a discussion of the advantages and limitations of the new developments presented and ideas for future work.

2 Study Site and Data

2.1 Topography and Hydrography

Box Elder County is a mountainous area in northern Utah, USA. Flooding that occurred in the area in 2017 had a marked influence on people and properties. Figure 1a illustrates the topography of a region located within Box Elder County, Utah, that includes the Bear River. The contour lines represent the 10-m DEM available from the National Elevation Dataset (NED), from which Liu et al.'s (2018) HAND data were derived. The raster layer, also from NED, shows the available 3-m DEM covering the Bear River. The dataset includes DEMS of different resolution, with 10-m resolution data available over the entire continental United States, whereas 3-m data are available only for specific areas. The 3-m domain covers a part of the Bear River between two stream gages: the upstream gage operated by PacifiCorp at Collinston and the lower gage operated by the U.S. Geological Survey (USGS)–station 10126000 Bear River near Corrine. We focused on the reach of the Bear River between these gages. The Malad River (Figure 1b) enters the study reach close to its downstream (southern) end. Three main highways cross the Bear River within the domain. This is of note because the road-top elevations recorded in the NED DEM that result in artifacts in Liu et al.'s (2018) published HAND layer have been corrected by the high-resolution hydrography breaching procedure applied here.

Details are in the caption following the image
(a) Location and topography of the study area with 10-m National Elevation Dataset (NED) as grey contour lines (10-m interval) and the available 3-m NED with raster layer (color region); (b) Hydrography of the study site from the NHDPlus dataset and the locations of two stream gages (i.e., the upper northern PacifiCorp gage at Collinston and the lower southern U.S. Geological Survey [USGS] gage at Corrine). Flow is from north to south.

According to the National Hydrography Dataset (NHDPlus), the contributing area at the PacifiCorp gage at the upstream end of the Bear River reach is 15,545 km2. The contributing area at the Corrine USGS gage at the downstream end is 17,868 km2. The contributing area of the Bear River at its junction with the Malad River is 15,728 km2, and the contributing area of the Malad River is 2,110 km2. Therefore, the area draining directly and from other tributaries into the Bear River reach is 213 km2 or about 10 times less than the Malad River.

2.2 Historical and Observed Streamflow

Historical peak discharges (Figure 2a) indicate that the February 2017 discharge was the third largest on record (1952–present) and the largest since two prior floods in 1984 and 1986 at the USGS Corrine gage. Discharges observed at each gage (the PacifiCorp and USGS gages) for the flood of 2017 are shown in Figures 2b and 2c, and the date (15 February) when Planet RapidEye satellite imagery is available is indicated as well. The daily average discharges on 15 February 2017 were 224.33 and 251.77 m3/s observed at the PacifiCorp (http://bearriverbasin.org/rivers/rivers/) and the USGS gages (https://waterdata.usgs.gov/nwis/uv?site_no=10126000), respectively. Note that while the USGS gage recorded 15-min discharges, only daily discharges were available at the PacifiCorp gage. However, this event was gradual enough that daily discharges are sufficient for this analysis. The instantaneous discharge at the USGS gage at noon on 15 February was 250.6 m3/s, essentially identical to the daily average.

Details are in the caption following the image
(a) Historical annual peak discharges observed at the U.S. Geological Survey (USGS) 10126000 Bear River near the Corrine gage from 1952 to present and (b) discharges observed during February 2017 at the PacifiCorp gage near Collinston (blue line = daily average) and the USGS 10126000 Bear River gage near Corrine (green line = 15-min). The vertical dotted line indicates the date for which Planet RapidEye imagery is available closest to the peak flow, and horizontal dotted lines are the values used for inundation mapping.

2.3 Planet RapidEye Satellite Imagery

A Planet RapidEye image from 15 February 2017 was selected as being closest in time to the peak (Figure 3a). The RapidEye satellite constellation consists of five satellites collecting remotely sensed information within five spectral bands (blue: 440-510 nm; green: 520-590 nm; red: 630-685 nm; red edge: 690-730 nm; and near infrared: 760-850 nm) with the daily off-nadir and 5.5-day at-nadir revisit time. The specific image selected was a RapidEye Analytic Ortho Tile at a spatial resolution of 5 meters captured on 15 February 2017 at 18:42 UTC (i.e., 11:42 a.m. MST). This tile included 100% coverage of the study area and cloud cover of less than 1%. The RapidEye Analytic Ortho Tile product is geometrically corrected by the Planet (2017) team. These corrections remove distortions due to image perspective (tilt) and relief (terrain) prior to it being made available to research users.

Details are in the caption following the image
(a) The true color RapidEye Analytic Ortho Tile image captured on 15 February 2017 from Planet. (b) Zoomed in RapidEye image. (c) 30-cm resolution world imagery from ESRI (2018) for the same area as shown in (b).

We present a zoomed-in plot of the yellow box (Figure 3b) to show a portion of the flooded area (mostly around the river corridor) and the corresponding high-resolution 30-cm world imagery (Figure 3c) from ESRI (2018) as a reference image to indicate the general land cover of this area.

2.4 National Water Model Analysis and Assimilation Data

The Current CFIM approach implements the Zheng, Tarboton, et al. (2018) and Liu et al. (2018) HAND method using the NWM flows for each NHDPlus reach. The NWM is a hydrologic model that simulates the water cycle over the entire continental United States (http://water.noaa.gov/about/nwm). The model produces an analysis and assimilation discharge that represents a snapshot of current hydrologic conditions, in addition to short, medium, and long-range forecasts. Since we were working with a historical flood, we used the NWM assimilated flow at 18:00 UTC on 15 February 2017 (Figure 4) to create flood inundation maps representative of the current CFIM method.

Details are in the caption following the image
Assimilated National Water Model daily average flows (Q) on 15 February 2017 for selected reaches in the Bear River study site and Malad River tributary inflow. Observed flows shown for reference.

The Bear River has 32 NHDPlus reaches between the two gages in the study site. Discharge is shown for a selection of reaches to illustrate the modeled increase along the reach. The upstream reach (close to the PacifiCorp gage) had a daily average discharge of 146.8 m3/s on 15 February 2017 from the assimilated NWM flows, whereas the daily average observed value for the same date at the gage was approximately 224.3 m3/s. Subtracting the upstream flow (146.8 m3/s) from the flow just prior to the Malad River junction (183.4 m3/s), we infer that model tributary inflows are about 36.7 m3/s along the study reach upstream of the Malad River junction. The NWM Malad River inflow was 99.0 m3/s, consistent with the increment in the Bear River flow across this junction. However, a discontinuity exists in the reported assimilated NWM discharges upstream and downstream of the lower stream gage. These discharges differ substantially from observed discharges. Notably, the NWM discharge at the upper end of the study reach (daily average of 146.8 m3/s) is about 35% less than daily average observed by PacifiCorp (224.33 m3/s). At the downstream end, the discharges are closer (daily average of 251.77 m3/s observed and daily average of 285.8 m3/s NWM). These discrepancies may affect CFIM results.

3 Methodology

This section describes (1) the workflow to generate a flood inundation map based on HAND, along with the proposed improvements; (2) the flow direction conditioning method; (3) the workflow used to map observed flood inundation based on RapidEye imagery using a supervised classification method; (4) evaluation metrics used to compare HAND-based flood inundation mapping with high-resolution RapidEye imagery; and (5) the improvements to flood inundation mapping developed and evaluated as part of this effort.

3.1 HAND-Based Flood Inundation Mapping and Proposed Improvements

The HAND raster is a drainage-normalized and flowpath-coherent version of a DEM (Nobre et al., 2016). Generating a HAND-based flood inundation map primarily consists of three procedures:
  1. Calculate HAND.
  2. Estimate reach-averaged hydraulic properties and synthetic rating curve from HAND for each stream reach.
  3. Given discharge and the rating curve, calculate stage and map inundation at locations where HAND is less than the calculated stage.

According to the CFIM described by Liu et al. (2018), calculating HAND starts with hydrologically conditioning the DEM by pit filling. Then D8 flow directions are computed. A raster representation of the stream network is derived using head points of streams mapped in the NHDPlus medium-resolution dataset as stream starting points. This is done using a weighted D8 flow accumulation calculation with weights taken as 1 at stream heads, and then applying a threshold to the result. This produces a stream raster, which is effectively the grid cells along downslope flow paths traced along flow directions from each starting point. This procedure maps streams at a drainage density close to that of the NHDPlus stream network, but it differs from NHDPlus in that streams are located along DEM elevation valleys, which resolves discrepancies that arise due to misalignment between the DEM and cartographically mapped NHDPlus streams. Finally, the TauDEM D distance down function is used with the hydrologically conditioned (pit filled) DEM and D flow directions and the raster representation of the drainage network to generate the HAND raster. For details see Zheng, Tarboton, et al. (2018) and Liu et al. (2018).

The second procedure estimates hydraulic properties from the HAND map as described by Zheng, Tarboton, et al. (2018). The procedure uses the HAND raster map to extract/estimate the river geometry (i.e., surface area, wetted bed area, and volume of flood) and then compute the reach-averaged hydraulic properties (i.e., cross section area, wetted perimeter, top width, and hydraulic radius) for each reach. This procedure operates over the local catchments, defined as the areas draining directly to each stream reach. Zheng, Tarboton, et al. (2018) and Liu et al. (2018) used NHDPlus catchments. Here, in addition to NHDPlus catchments, we evaluated using catchments derived directly from the higher resolution DEMs (10 and 3 m) using nodes dispersed approximately uniformly along the study reach. NHDPlus catchments were derived from a 30-m DEM and may not align well with 10- and 3-m DEM-derived catchments in locations with complex topography. Hydraulic properties are calculated for a series of water depths or stages h, ranging from 0 to greater than the maximum water depth reasonably possible. For each catchment and water depth h, we first identify the cells where the HAND value H is the less than h. Then, the river geometry and the reach-averaged hydraulic properties are computed, and Manning's equation (equation 1) is used to calculate the discharge associated with each depth value h to establish the synthetic rating curve for each reach.
urn:x-wiley:00431397:media:wrcr24106:wrcr24106-math-0001(1)
where Q is discharge, n is the hydraulic roughness parameter in Manning's equation, A(h) is the cross section area of a channel as a function of stage h, R(h) is the hydraulic radius as a function of stage h, and S is the slope of the channel.

The final (third) inundation mapping procedure uses discharge as input and the synthetic rating curve to determine stage h*. Then, inundated extent is mapped as those grid cells for which h* > H. Each inundated cell is classified as 1 (flooded), and the rest of the cells are classified as 0 (nonflooded). This is accomplished separately for the catchment associated with each reach.

Our application of this approach for the Bear River study reach identified a number of limitations where improvements are possible:
  1. The CFIM used by Liu et al. (2018) utilizes NHDPlus medium-resolution catchments in the evaluation of HAND hydraulic properties for the associated NHDPlus stream. Problems arise where these catchments are small or variable in size, sometimes due to the presence of canals/ditches in the NHDPlus dataset. Other discrepancies are due to the NHDPlus streams not being well aligned with the DEM used in the HAND calculation. This is because NHDPlus used a 30-m DEM, while HAND used a 10-m DEM to capture additional detail. This can lead to discrepancies in hydraulic properties calculated for NHDPlus reaches, which propagate into the inundation mapping.
  2. The pit filling method used in the CFIM can result in large flat areas behind barriers in the DEM, where the DEM represents the top surfaces of bridge crossings.
  3. The results, especially for hydraulic properties, are only as good as the DEM used. The CFIM approach is based on the 1/3rd arc-sec (10-m) DEM. In some places a 1/9th arc-sec (3-m) DEM is available and may provide an opportunity for improvement.
Here we introduce enhancements to the HAND-based flood inundation mapping approach to address each of the noted limitations and improve the accuracy of the flood inundation mapping result while relying on the concepts introduced by Zheng, Tarboton, et al. (2018) and Liu et al. (2018). The improvements, which explicitly aim to reduce the occurrence of unrealistic HAND values and diminish the impact of DEM errors on flood inundation mapping, are as follows:
  1. Remove the canal/ditch features from the NHDPlus medium-resolution flow lines prior to the preparation of HAND inputs because canal/ditches are generally not evident in the DEM.
  2. Derive hydraulic properties and synthetic rating curves using the DEM-based drainage network and catchments based on evenly spaced nodes along a stream reach to avoid inconsistencies due to the size variability of the catchments. One challenge is that the slope can be zero for reaches that are across flat areas, and these become artifacts in the DEM. In these cases, we adjusted the local elevations of the junctions in the digital representation of the stream network to shift elevation changes between stream reaches and impose a nonzero slope on each stream reach. For example, when the reach downstream of a junction has a positive slope, but the reach upstream is flat; the elevation of the junction can be lowered to make these reaches have an equal but smaller slope than the downstream reach did initially. We extended this idea both upstream and downstream to account for the occurrence of adjacent reaches with zero slope. The result was a set of reach slopes that are all positive but do not alter the overall elevation differences and hence slopes in the stream network.
  3. Incorporate a hybrid filling-breaching algorithm to hydrologically condition the DEM consistent with a high-resolution hydrography dataset (such as NHD high-resolution). We developed a new flow direction conditioning approach for this purpose.
  4. Use a higher-resolution DEM (i.e., 1/9 arc-sec or 3-m resolution) to enhance accuracy, along with the enhancements above, which are using evenly sized stream reach and high-resolution hydrography flow direction conditioning etching approach.

The next section details the flow direction conditioning approach.

3.2 Flow Direction Conditioning

The approach for flow direction conditioning is to first determine the set of flow directions, represented as a grid with the same dimensions as the DEM, that follows the downward flow direction of the given hydrography dataset, and then to adjust the DEM so that elevations, moving downslope along these flow directions, are nonincreasing. The flow directions were determined using the sequence of TauDEM and GIS grid manipulations described below, while a new TauDEM tool was developed to adjust (flow direction condition) the DEM. To determine flow directions along streams, the DEM was modified by lowering elevation values significantly for grid cells on streams, a process sometimes referred to as burning, after which pits are filled and flow directions computed for this DEM. The burned-in streams manifest as canyons, and flow directions are constrained to remain within the canyon (stream) until they exit the domain. These on stream flow directions thus follow the path of the rasterized hydrography dataset. Off stream flow directions are set aside, and only the on stream flow directions are used in the new flow direction conditioning tool. The detailed steps used to determine along stream flow directions are as follows:
  1. Designate the input DEM as Z.
  2. Convert the hydrography stream lines to a raster that has the same dimensions (columns, rows, cell size, and edge coordinates) as the DEM denotes srfv (stream raster from vector) with values 1 on stream and 0 off stream. The stream vector dataset used here should not include streams that enter the domain (extent of the grid) from outside and must include streams where they leave the domain.
  3. Burn srfv into Z using the cell-by-cell grid calculation Zb = ZB * srfv, where B is a big number set as described below. The resulting grid is Zb, a DEM with deeply burned canyons along the rasterized streams. The ArcGIS raster calculator tool is used here. The big number B is somewhat arbitrary and should be big enough to burn the stream hydrography into the DEM to a sufficient depth that when pits are filled and flow directions determined, the only path available for flow is along the rasterized hydrography. One way to decide on B is to fill pits and then subtract the original DEM and use a number that is larger than the maximum difference but does not produce numerical overflow. Our application with high-resolution NHD streams has different classes of flowlines, specifically StreamRiver, ArtificialPath, CanalDitch, Pipeline, Connector, Coastline, and Underground Conduit. We excluded all except StreamRiver, ArtificialPath, and Pipeline/Connector and used B=1,000 m for StreamRiver, 700 m for ArtificialPath, and 400 m for Pipeline/Connector, which gives preference to flow along the StreamRiver line that we took to be the main flow path in braided situations. Then, an ArtificialPath is prioritized over a pipeline/connector.
  4. Fill pits in burned DEM (Zb) using the TauDEM Pitremove function. The result is Zbfel, which is now hydrologically conditioned but with deeply burned canyons along the streams. To avoid completely filling in the deeply burned canyons during pit filling, they should extend only to the edge of the DEM where flow exits the domain. This is why the stream vector dataset should not include streams that enter the domain and must include streams that leave the domain.
  5. Calculate D8 flow directions using the TauDEM D8FlowDir function with input Zbfel. The output flow directions, which are designated as raster p, are constrained by the burning to be within the burned canyons along the streams. Because the stream vector data set is required to include streams leaving the domain and not streams entering the domain, the conditioned DEM from (4) and flow directions will define paths that follow downslope within the stream raster to the edge of the domain where streams leave.
  6. Mask the D8 flow directions to have only the flow directions on streams, setting all other flow directions to no data. The raster calculation pm = p/srvf achieves this.

The result from this process is a grid with D8 flow direction values set along the streams for the grid cells intersecting the input hydrography dataset. This is used as input to the flow direction conditioning tool. The significantly altered DEMs (Zb and Zbfel) are not used further, and neither are the flow directions that were computed for parts of the domain outside streams.

The second step in flow direction conditioning is to adjust DEM elevations so they are strictly nonincreasing in the downstream direction along the flow directions generated from the first step above. A new TauDEM function, “flow direction conditioning,” was written to achieve this (Algorithm 1). This takes as input the original DEM, Z, and conditioned flow direction raster, pm, and produces as output a conditioned DEM, Zc.

urn:x-wiley:00431397:media:wrcr24106:wrcr24106-math-0002(2)

This recursively evaluates Z to the least of the incoming neighbor Z values and the cell value itself. The result is a grid of elevation values Z that are conditioned to drain along the flow directions given.

3.3 RapidEye-Based Flood Inundation Mapping

Satellite-based flood inundation mapping is generally performed using water detection algorithms. Water in satellite images is detected using three methods: (1) single-band, (2) multiband, and (3) classification. The single-band approach involves choosing one characteristic band from a multispectral image, a band for which the spectral signature of water is unique and representative. Then, to discriminate water from other surfaces, a threshold, often derived from the histogram analysis of the image for the characteristic band, needs to be defined. Separating surface water from other land types based on a single threshold in a single unique band is frequently problematic (Verpoorter et al., 2012). Identification of surface water can be improved using multi-band methods where a combination of different bands is used through a so-called spectral index. Multiband methods also require definition of a threshold for the selected spectral index to determine whether a pixel value is categorized as water or not. The subjective selection of the threshold may lead to an overestimation or underestimation of surface water (Xu, 2006).

Classification methods are frequently applied for classifying surface water in images. The classification approaches can be categorized into two main groups: supervised and unsupervised. The major task of the supervised methods is to segregate the spectral domain into classes (different land covers) according to their spectral similarities. Unsupervised methods assign a class to each pixel without any prior knowledge of the names (types) of those classes. Unsupervised techniques derive their result using the statistical properties of the data. In other words, unsupervised techniques group pixels if they have similar statistical properties. In cases where it is possible to obtain a labeled dataset (with class names such as water class) for training the classification algorithm, supervised classification is suggested as it significantly outperforms unsupervised classification (Laskov et al., 2005). Ireland et al. (2015) provides an example of the successful application of supervised classification of flooded areas from Landsat imagery.

In this study, we used a supervised classification method within the ArcGIS Pro software from ESRI (2018) on a RapidEye image for 15 February 2017 (which is 1 day after the first peak of the flood). Visually examining the true color image, we developed a training sample that included areas within open water and non-water areas. The RapidEye image was captured in wintertime (February 2017), when it is sometimes difficult to discriminate between water and shadow pixels visually. To mitigate this difficulty and increase the accuracy of the training sample, we used a normalized difference vegetation index (NDVI) map computed using equation 3, recognizing that negative NDVI values theoretically correspond to water.
urn:x-wiley:00431397:media:wrcr24106:wrcr24106-math-0003(3)

Within the ArcGIS Pro Classification Wizard, the Classification Method was set to supervised classification, the Classification Type to Pixel based, and the schema to default. Next, through the Training Samples Manager, we selected the training samples and chose Maximum Likelihood as the classifier. The spatial resolution of the classified image was the same as the RapidEye image (i.e., 5 m). Since the resolution of the HAND-based flood inundation maps in the present study was 10 or 3 m, we used nearest neighbor re-sampling in ArcGIS Pro to obtain both 10- and 3-m classified maps, which were then used for calculating the evaluation metrics as described in the following section.

3.4 Evaluation Metrics

In general, evaluation metrics are used to validate the results of a model against observations to measure how well the model performs. In this study, we used Correctness (C; equation 4) and Fit (F; equation 5) to indicate the degree-of-overlap between model and observed flood inundation maps (Horritt & Bates, 2002; Merwade et al., 2018; Sangwan & Merwade, 2015).
urn:x-wiley:00431397:media:wrcr24106:wrcr24106-math-0004(4)
urn:x-wiley:00431397:media:wrcr24106:wrcr24106-math-0005(5)

Both statistics should ideally be 1 (100%). C is an overall area metric and F is a location-specific metric. C, the correctness metric, quantifies the degree to which the total modeled and observed areas classified as inundated (wet) match. F is a stricter statistic that quantifies whether modeled and observed locations match (i.e., intersection), scaled by the total area mapped as inundated by either (i.e., union; Sangwan & Merwade, 2015). Here we took the Modelwet as HANDwet, the grid cells inundated in a HAND-based flood inundation map (such as CFIM), and Observedwet as RapidEyewet, the grid cells classified as water (inundated) from the RapidEye imagery.

3.5 Improvements Developed

We developed HAND-based flood inundation maps for each of the following scenarios to evaluate the HAND approach and suggested terrain-processing improvements through comparison to the classified RapidEye image. The first scenario is the current CFIM base case used as a starting point for evaluating improvements.
  1. Model inundation based on publically available CFIM information, where the hourly-assimilated NWM flows on 15 February 2017 at 18:00 UTC are used. The HAND raster, hydraulic properties, and rating curves were obtained from the NFIE Continental Flood Inundation Mapping data repository (https://web.corral.tacc.utexas.edu/nfiedata), and the codes that we used are the ones available on GitHub (https://github.com/cybergis/nfie-floodmap). This serves as an indicator for how well the current CFIM approach may be expected to perform.
  2. Model inundation based on HAND, hydraulic properties, and rating curves from scenario (1) but with observed discharges in the main river (i.e., Bear River) and negligible discharges from side tributaries except for the Malad River (because the drainage area of other tributaries is much smaller than the Malad River). We assumed a flow of 224.33 m3/s in the Bear River reach from the upper PacifiCorp Collinston gage to the Malad junction, with a flow of 251.77 m3/s in the Bear River reach downstream of the Malad junction. This is our best estimate of what actual discharges were on 15 February 2017, and it serves to separate the effect of discharge errors from DEM and HAND and rating curve errors.
  3. We delineated only the main Bear River and Malad tributary channels from the DEM, initiating them at channel heads near where the NHDPlus medium-resolution (i.e., 1:100,000) channel network enters the domain. We distributed nodes evenly along the main river to delineate catchments from the 10-m DEM for use in the HAND process. This results in streams comparable to CFIM but more consistent catchments, and hence, more consistent channel properties than used in CFIM where NHDPlus medium resolution catchments are based on the 30-m DEM and are quite variable in size.
  4. We etched the high-resolution hydrography (NHD high-resolution, i.e., 1:24,000) into the 10-m DEM using flow direction conditioning (section 3.2). This removes DEM barriers often due to roads. Catchments were delineated using evenly distributed nodes. This scenario allows us to evaluate the flow direction conditioning approach at the same DEM resolution as operational CFIM.
  5. We repeated the procedure from scenario (3) but with the high resolution (1/9th arc-sec, 3 m) DEM that is available for this area. This scenario allows us to evaluate the potential benefit from higher resolution DEM data.
  6. We etched the high-resolution hydrography (NHD high-resolution, i.e., 1:24,000) into the 3-m DEM using flow direction conditioning. This repeated the procedure from scenario (4) but with the high-resolution (1/9th arc-sec, 3-m) DEM that is available for this area and allowed us to evaluate the potential benefit from higher resolution DEM data with flow direction conditioning.

4 Results

4.1 Conditioned Topography Through Flow Direction Conditioning Approach and Its Effect on HAND

To illustrate the effect of the flow direction conditioning (etching) on the DEM, we zoomed in to locations that illustrate the effect well (Figure 5). The flow direction conditioning approach etches the path of the high-resolution NHD flowline into the 10-m DEM (Figures 5a-5c). The DEM etching that is apparent here was negligible or not discernible over much of the 10-m DEM area, indicating that the etching only affects the DEM at places where barriers exist. Road barriers seemed more prevalent in the 3-m DEM, and etching provides a way to resolve (punch through) these barriers (Figures 5d-5f).

Details are in the caption following the image
Examples of the effect of flow direction conditioning (etching). (a) 10-m digital elevation model (DEM) close to Bear River City, (b) 10-m stream raster of the high-resolution National Hydrography Dataset (NHD), (c) 10-m etched DEM close to Bear River City, (d) 3-m DEM around Highway 102, (e) 3-m stream raster of the high-resolution NHD, and (f) 3-m etched DEM around Highway 102.

We created HAND rasters for each DEM (Figure 6). The results show that the etching method affects the HAND raster for both selected areas. Without etching, a considerable part of the streambed is flat due to pit filling, resulting in a HAND map with values close to 0 for most of the river corridor (Figures 6a and 6c). This is of concern because it impacts extraction of river hydraulic geometry and the synthetic rating curve from the HAND map such that for the same flowrate, the water depth estimated from a HAND-derived synthetic rating curve might be unrealistic.

Details are in the caption following the image
The HAND map for the regions shown in Figure 5 computed based on (a) 10-m pit-removed digital elevation model (DEM), (b) 10-m pit-removed etched DEM, (c) 3-m pit-removed DEM, and (d) 3-m pit-removed etched DEM for areas close to Bear River City (a-b) and Highway 102 (c and d)

4.2 Flood Inundation

The length of the study reach is almost 60 km, with 32 NHDPlus reaches and corresponding catchments along the main stem, although some of these are very small. Other studies (Godbout, 2018; Zheng, Maidment, et al., 2018) have suggested that reach lengths of 1.5 to 5 km may be optimal for use with the HAND approach. We placed 20 nodes approximately evenly along the main stem, including at the downstream end, upstream end, and at the Malad River junction. This number of nodes was chosen to obtain stream reaches about 3 km long, within the range suggested by Godbout (2018) and Zheng, Maidment, et al. (2018). This resulted in 18 catchments upstream of the Malad River junction and one catchment from the Malad River junction to the downstream gage being delineated, each draining to a reach about 3 km long (Figure 7).

Details are in the caption following the image
Location of nodes and associated catchments in (a) NHDPlus-derived Continental-scale Flood Inundation Mapping (CFIM) approach and (b) evenly distributed nodes to get uniform length reaches.

Flood inundation maps for each scenario that we evaluated are shown in Figure 8, and comparisons with classified inundation are detailed in Figure 9. We compared observed flood inundation (a) with modeled flood inundation using NWM discharges (b), observed discharges (c), DEM-derived catchments from evenly distributed nodes (d), etched high-resolution hydrography (e), 3-m DEM (f) and etched hydrography with 3-m DEM (g). The results show that the modeled inundation extent (Figures 8b-8g) is able to capture the majority of the observed inundation extent from Planet RapidEye satellite (Figure 8a) in all scenarios. Evaluation metrics C and F (Table 1) quantify the performance of each scenario.

Details are in the caption following the image
Modeled and observed flood inundation maps. (a) RapidEye-OBS: inundation classified from Planet RapidEye satellite imagery; (b) Scenario 1, Continental-Scale Flood Inundation Mapping-National Water Model (CFIM-NWM): CFIM methodology with NWM assimilated flows; (c) Scenario 2, CFIM-OBS: CFIM methodology with observed flows; (d) Scenario 3, 10m-Uniform: evenly distributed nodes and TauDEM used to redelineate channels and catchments from 10-m DEM; (e) Scenario 4, 10m-ETCH: high-resolution National Hydrography Dataset (NHD) flow paths etched into 10-m DEM; (f) Scenario 5, 3m-Uniform: evenly distributed nodes and TauDEM used to redelineate channels and catchments from 3-m DEM; and (g) Scenario 6, 3m-ETCH: high-resolution NHD flow paths etched into 3-m DEM.
Details are in the caption following the image
Comparison of each scenario with inundation classified from Planet RapidEye imagery. (a) Continental-Scale Flood Inundation Mapping-National Water Model (CFIM-NWM), (b) CFIM-OBS, (c) 10m-Uniform, (d) 10m-ETCH, (e) 3m-Uniform, and (f) 3m-ETCH.
Table 1. C = Correctness Metric and F = Fit Metric Values for Each Scenario
Fig 9 a. b. c. d. e. f.
Scenario 1. CFIM-NWM 2. CFIM-OBS 3. 10m-Uniform 4. 10m-ETCH 5. 3m-Uniform 6. 3m-ETCH
C 1.77 1.65 1.71 1.67 1.60 1.49
F 0.47 0.45 0.47 0.52 0.54 0.59
  • Abbreviation: CFIM-NWM, Continental-Scale Flood Inundation Mapping-National Water Model.

The results (Figure 9 and Table 1) show that the adoption of measured discharge values improves C as it reduces from 1.77 in scenario 1 (CFIM-NWM) to 1.65 in scenario 2 (CFIM-OBS). However, F degrades slightly (i.e., 0.47 to 0.45). The improvement in C is due to better representation of tributary flows, which, at least for the NWM assimilation we had, are inconsistent with the observations. The reduction in F is due to overall greater inundation because of higher observed flows, and the DEM having fairly extensive flat areas where pits were filled.

Using evenly distributed nodes to create uniform-length reaches and associated catchments for use in synthetic rating curve estimation, going from scenario 2 (CFIM-OBS) to scenario 3 (10m-Uniform), does not lead to an improvement in overall inundation extent prediction as C increases due to the overestimation of stage caused by flat areas near the downstream end of the study area close to Bear River City (see arrow in Figure 9c). On the other hand, F increases slightly from CFIM-OBS to 10m-Uniform (Table 1) indicating that location-specific overestimation and underestimation are slightly improved. One reason for this is that catchments are better aligned with the DEM (Figure 10).

Details are in the caption following the image
The effect of using evenly distributed nodes and digital elevation model (DEM)-derived catchments and streams on flood inundation for a zoomed in region in the study domain. (a) A selected NHDPlus catchment and the HAND map based on 10-m DEM as used in the Continental-Scale Flood Inundation Mapping-National Water Model (CFIM-NWM) and CFIM-OBS scenarios. (b) A selected DEM-derived catchment and the HAND map based on 10-m DEM as used in the 10m-Uniform scenario. (c) Modeled flood in CFIM-OBS. (d) Modeled flood in 10m-Uniform.

In the CFIM-OBS scenario (Figure 10a), the discrepancy between NHDPlus catchments and HAND derived from 10-m DEM is quite evident. An area within the catchment, according to the 10-m DEM, drains to another nearest stream, and this impacts the calculation of hydraulic properties and mapping of inundation (Figure 10c). On the other hand, the catchment derived based on the 10-m DEM is more consistent (Figure 10b) and results in improved mapping of inundation.

The results (Figure 9 and Table 1) show that the flow direction conditioning of the 10-m DEM (scenario 4) improves both C and F metrics (Table 1). The overestimation of stage caused by flat areas near the downstream end of the study area close to Bear River City (see arrow in Figure 9d) is reduced. Additionally, using a high-resolution DEM (3-m DEM in scenario 5, i.e., 3m-Uniform) improves both metrics compared to 10m-ETCH (F: 0.54>0.52 and C: 1.60<1.67). However, overestimation still occur in some areas (such as areas close to Highway 30; Figure 9e) due to a barrier in the 3-m DEM caused by Highway 30. Overall flow direction conditioning of the 3-m DEM (scenario 6) improves the mapping of inundation in terms of both C and F metrics (Table 1 and Figure 9f). In particular, the F metric improves noticeably since the overestimation (red color) of the modeled flooded areas is reduced around the areas where road crossings exist.

To illustrate the effect of flow direction conditioning (etching) with the 3-m DEM, we zoomed in on an area where 3m-Uniform has mapped inundation that was not observed (Figure 11). In the 3m-Uniform scenario (Figure 11a), the HAND map shows a notable flat area, which then caused this entire area to be flooded (Figure 11c). In 3m-ETCH (scenario 6), where etching removed the DEM barrier, the overestimation of the modeled flood was reduced (Figure 11d).

Details are in the caption following the image
The effect of flow direction conditioning on flood inundation for an area around Highway 30. (a) HAND based on 3-m digital elevation model as used in the 3m-Uniform scenario, (b) HAND based on 3-m etched digital elevation model as used in the 3m-ETCH scenario, (c) modeled flood in 3m-Uniform, and (d) modeled flood in 3m-ETCH.

4.3 Sensitivity Analysis: Optimal Stage and Roughness Coefficient

The CFIM method uses a uniform value of 0.05 for Manning's n in the calculation of synthetic rating curves. We also used this value in the six scenarios reported so far. However, this assumption may be responsible for uncertainties that remain in the modeled inundation extent. To investigate this, we used catchments delineated with the 3m-ETCH DEM and searched over a range of stage (h) values for the stage that provided the best fit (highest F metric) compared to the Planet RapidEye observed flood inundation extent. This provides a quantification of the best inundation mapping possible using the HAND approach with the DEM and catchments chosen, separate from uncertainties in discharge and synthetic rating curve. The stage evaluated to produce best fit F (Figure 12a) was then used to back calculate the Manning's n that would produce this stage using equation 1 and the observed discharge (Figure 12b) for each reach. The corresponding fit metric, F, (Figure 12c) and correctness metric, C, (Figure 12d) are also shown for each catchment. Overall, the optimal stage resulted in F = 0.72 and C = 1.11. The average of the fitted Manning's n values was n = 0.02 and stage, as well as the Fit metrics for this n, applied as a uniform value to all catchments, and the a-priori n = 0.05 are shown (Figures 12 a, 12c, and 12d). Applying the average of the fitted Manning's n (n = 0.02) to each catchment with the 3m-ETCH scenario produced an inundation map where F and C were 0.63 and 1.12, respectively. Note that this notably improves both metrics relative to the best from Table 1 (i.e., F = 0.59 and C = 1.49). Thus, an improvement in F of about 7% (from 0.59 to 0.63) may be obtained simply by calibrating n, keeping the value the same everywhere. A further improvement (from 0.63 to 0.72) may be obtained by letting n vary spatially. Furthermore, recognizing that F=0.72 is the optimal fit of stage given the DEM and RapidEye observations, we can interpret F values from Table 1 relative to this value. Specifically, for this study the impact of flow direction conditioning at the 10-m scale was an improvement of F from 0.47 to 0.52, a 7% improvement. At the 3-m scale, flow direction conditioning improved F from 0.54 to 0.59, an 7% improvement. In moving from 10- to 3-m DEM, F with flow direction conditioning improves from 0.52 to 0.59, an 10% improvement.

Details are in the caption following the image
Optimizing stage and estimating Manning's n. (a) Stage values for each stream reach using (I) synthetic rating curve with n = 0.05; (II) Best fit F from optimization of h to where F is maximized; (III) synthetic rating curve with n = 0.02, the n value obtained by averaging the n values obtained from the optimal stage. (b) Manning's n obtained from stage corresponding to best fit F. (c) Fit metric F for each of the cases in (a). (d) Correctness metric C for each of the cases in (a).

4.4 Validation of Developed Improvements

To validate the improvements developed, we applied the approach to a reach of the Ocheyedan River (about 13 km) close to Spencer City in Iowa, USA, that experienced a flood in March 2019. This validation case was selected using the following criteria: (1) the reach of interest should have at least one active observation gage to provide the observed streamflow, (2) it should have high-resolution DEM (preferably 3m), (3) the region of interest should have high-resolution and cloud free satellite imagery on the flood date, and (4) it should include roads crossing the stream rivers such that the effect of etching on HAND calculations makes a difference. This validation case was identified by searching waterwatch.usgs.gov to identify recent floods and at the same time searching Planet.com for the availability of Planet satellite imagery and nationalmap.gov to check whether the region a had high-resolution DEM (3m). After several attempts, we selected a reach of the Ocheyedan River (about 13 km) close to Spencer City in Iowa, USA, as our validation case study because it met all criteria mentioned above (Figure 13a).

Details are in the caption following the image
Validation case study and the observed flood of March 2019 in the Ocheyedan River in Iowa. (a) Hydrography of the study site from the NHDPlus dataset and the location of the USGS 06605000 gage close to Spencer City (Flow is from west to east), the true color Sentinel-2 Tile image captured on (b) 20 April 2019 and (c) 21 March 2019 from Planet, and (d) the daily average discharges observed during March 2019 at the USGS 06605000 gage. The vertical dotted line indicates the date for which Planet Sentinel-2 imagery is available closest to the peak flow, and the horizontal dotted line is the value used for inundation mapping (i.e., 75.61 m3/s).

The flood of March 2019 on the Ocheyedan River reach (Figure 13d) shows the daily average peak value of 172.45 m3/s on 15 March observed at the USGS 06605000 Ocheyedan River near Spencer gage. The National Weather Service flood stage for this gage is 2.44 m (or 8.0 ft.), and the observed gage heights at the USGS 06605000 gage were above the flood stage during 14–23 March (for example 3.66 m on 15 March and 3.05 m on 21 March at noon).

A Planet Sentinel-2 image from 21 March 2019 captured at 17:21 UTC (i.e., 12:21 pm CT) was selected as being closest in time to the peak (Figure 13c). This image is within the flood period (i.e., 14–23 March) where the gage heights were above the flood stage. When compared to another Sentinel-2 image captured on 20 April 2019 (Figure 13b), the flooded region can be seen in the images. The Sentinel-2 satellite (different from the RapidEye satellite used in the Bear River) is a wide-swath, high-resolution and multi-spectral satellite that contains 13 spectral bands with different spatial resolution (10, 20, and 60m). The specific image selected was S2A_MSIL1C_20190321T171011_N0207_R112_T15TUH_20190322T001148. This tile includes 100% coverage of the study area with 0% cloud coverage. As in the workflow used in the Bear River case study, we used the Visible and Near Infrared bands (i.e., band B2 [blue], B3 [Green], B4 [Red], and B8 [NIR] with spatial resolution of 10m) for the supervised classification.

We created a HAND raster for the 10-m DEM (as used in the published CFIM), 3-m DEM, and 3-m etched DEM and zoomed in to a location (the red circle in Figure 13) about 4-km downstream of the USGS 06605000 gage, where a road crosses the Ocheyedan River reach (Figure 14). Results illustrate the importance of using a high-resolution DEM on the HAND raster as well as the impact of the etching approach on removing barriers when using such a high-resolution DEM dataset. With the 10-m DEM (Figure 14a), a considerable portion is flat due to pit filling, resulting in a HAND map with values close to 0 for most of the river corridor. Without etching (Figure 14b), the existence of the road barrier seems more prevalent in the high-resolution DEM, which affects the HAND raster. Once again, this is of concern because it influences the extraction of river hydraulic geometry and the synthetic rating curve from the HAND map. The flow direction conditioning approach (Figure 14c) etches the path of the high-resolution NHDPlus flowline into the 3-m DEM and provides a way to resolve (punch through) the barrier.

Details are in the caption following the image
Example of the effect of flow direction conditioning (etching) at a location about 4 km downstream of the USGS 06605000 gage. The HAND map based on (a) a 10-m pit-removed digital elevation model (DEM) as used in the published Continental-Scale Flood Inundation Mapping, (b) a 3-m pit-removed DEM, and (c) a 3-m pit-removed etched DEM.

We followed the procedure of dispersing nodes approximately uniformly along the Ocheyedan River reach to avoid the sometimes small and irregular-sized NHDPlus catchments (Figure 15a) that degrade synthetic rating curve estimation. We placed six nodes approximately evenly along the main stem (Ocheyedan River reach between the upstream junction of Stony Creek and the Ocheyedan River and the downstream junction of the Ocheyedan River and Little Sioux River). This resulted in five catchments each draining to a reach of about 3 km long (Figure 15b).

Details are in the caption following the image
Location of nodes and associated catchments in (a) NDHPlus derived Continental-Scale Flood Inundation Mapping (CFIM) approach and (b) evenly distributed nodes to get uniform length reaches. Comparison of the result of the modeled (HAND-based) flood inundation with inundation classified from Planet Sentinel-2 imagery. (c) CFIM approach and (d) 3-m etch digital elevation model with all improvements involved.

We used the daily average discharge value of 75.61 m3/s (observed on 21 March) in the main river and illustrated the comparison of the modeled inundation based on HAND for both CFIM and the 3-m etched DEM with classified inundation from Planet Sentinel-2 satellite imagery (Figures 15c and 15d). In order to prevent inconsistency in the different domain due to misalignment of the NHDPlus and DEM-derived catchments, we chose a region that is available for both scenarios (i.e., the NHDPlus catchments) for calculating evaluation metrics. Results show that the modeled inundation extent is able to capture the majority of the observed inundation extent from Planet Sentinel-2 in both scenarios. However, discrepancies between modeled and observed flood inundation are more apparent in the CFIM method (Figure 15c).

Computed evaluation metrics (shown in Figures 15c and 15d) for each scenario shows that using evenly distributed nodes to create uniform-length reaches and associated catchments for use in synthetic rating curve estimation along with flow direction condition to resolve barriers in a DEM improve both C and F metrics compared to CFIM (C: 1.17<1.44 and F: 0.67>0.56). The discrepancy between NHDPlus catchments and HAND derived from a 10-m DEM resulted in overestimation of flood inundation in CFIM (see arrows labeled as “A” in Figure 15c). These areas within the NHDPlus catchment drain to other nearest stream according to the 10-m DEM. This affects the calculation of hydraulic properties and mapping of inundation. In addition, due to the existence of DEM barriers, a considerable portion is flat due to pit filling, resulting in a HAND map with values close to 0 and overestimation of the estimated stage (see the arrow labeled as “B” in Figure 15c). Results show our improvements reduce the occurrence of unrealistic HAND values and diminish the impact of DEM errors on flood inundation mapping.

Evaluating the stage that best matches HAND-based flood inundation with observed inundation and letting Manning's n vary spatially resulted in F = 0.69 and C = 1.10. This improves both metrics relative to the best scenario (i.e., F = 0.67 and C =1.17). Thus, an improvement in F of about 3% (from 0.67 to 0.69) may be obtained simply by calibrating Manning's n. Applying the average of the fitted Manning's n values (n=0.04) as a uniform value to all five catchments produced an inundation map where F and C were 0.67 and 1.13, respectively. Recognizing that F=0.69 is the optimal fit of stage given the DEM and Sentinel-2 observations, we can interpret F relative to this value. In moving from CFIM to 3m-ETCH, F with flow direction conditioning improves from 0.56 to 0.67, a 16 % improvement.

5 Discussion

This work follows on from the work of Zheng, Maidment, et al., (2018); Zheng, Tarboton, et al., (2018) and Liu et al. (2018) that have developed and advanced the HAND approach for use on the U.S. NHDPlus network by the National Water Model. In our work, the streams to which HAND was calculated, and the catchments over which hydraulic properties were evaluated, were derived purely from the DEM after conditioning. This is important to ensure consistency between streams and catchments in the terrain analysis processing for enriching the content of DEM data for use in hydrologic modeling. This should be the case no matter the resolution or accuracy of the DEM. There will always be DEM errors to some degree. Flow direction conditioning alters the DEM to be consistent with the given hydrography, taken to be a better source for flow direction than the DEM, which suffers from artificial barriers.

The approach evaluated here differs from Zheng, Tarboton, et al. (2018) in two ways. First, Zheng, Tarboton, et al. (2018) did not use conditioning on high-resolution streams to etch, or punch through, road barriers, and thus HAND evaluated by Zheng, Tarboton, et al. (2018) will still be limited by the occurrence of flat areas upstream of these barriers. Second, streams mapped using the geodesic minimization approach are not guaranteed to align with topographic minima the way that streams derived directly from the DEM flow directions do, and thus, it cannot be guaranteed that they are good targets with which to evaluate HAND. In the calculation of HAND, it is important that the “stream” used as a target for HAND be consistent with the DEM, or, more specifically, be located at the bottom of whatever valley or channel is represented in the DEM, so that HAND is a positive quantity measured down to this target stream.

The HAND process used in this study has uncertainty due to the simplified representation of flow hydraulics through the assumption of uniform flow and application of Manning's equation over irregular stream reaches. This is acknowledged as a shortcoming, but it is also advocated as a useful approximation for regional- and national-scale modeling, as the information needed to apply more rigorous hydraulic approaches is rarely available over large areas. The use of reach-averaged hydraulic properties derived from HAND is seen as an advantage over cross-section-based approaches, such as are commonly used in Federal Emergency Management Agency HEC-RAS studies, because the reach-averaged hydraulic properties account for, and integrate, through the aggregation of volume and wetted bed area associated with each flow depth, the variability within a stream reach that can be lost between cross sections in a cross-section-based approach.

The inundation mapped from remote sensing, from Planet RapidEye satellites or Planet Sentinel-2 in this study, also has uncertainties. The best match that would be possible with the Planet RapidEye classified inundation used in the Bear River reach or Planet Sentinel-2 classified inundation used in the Ocheyedan River reach case studies and HAND mapped inundation has fitness scores of 0.72 and 0.69, respectively, less than the theoretical optimum of 1. This is partly due to errors and uncertainties in satellite mapped inundation, and such inundation being inconsistent with the topography. For example, adjacent grid cells of equal HAND value should either both be inundated or not, but this was not always the case with Planet RapidEye inundation.

Adjusting the HAND inundation threshold, h, to achieve the best fit between Planet RapidEye or Sentinel-2 inundation and HAND inundation provides an independent estimate of the stage in each stream reach. When combined with observed discharge (or modeled discharge where observations have been assimilated), this provides an independent estimate of the Manning's n channel hydraulic roughness parameter consistent with other assumptions in the HAND approach. Fitted roughness values generated by doing this were on average less than the default value used in CFIM for both Bear River and Ocheyedan River case studies, although, again, this inference is limited to these study reaches. This may reflect a bias in the CFIM value, or may also reflect part of the channel being missing in the DEM due to the DEM representing water surface elevations. This observation opens some new questions for future research. First, it emphasizes the importance of having DEMs that represent the bathymetry (channel bed topography) as closely as possible. Second, there is a question as to whether, in locations where the DEM does not represent bed topography, the hydraulic geometry parameters of the missing part can be inferred from the inferred stage, discharge, and roughness parameter. Third, this suggests an opportunity to use observed inundation from past floods to infer stream reach hydraulic roughness in a way that is consistent with HAND hydraulic geometry and to use these values in hydraulic routing. We recognize that the specific Manning's n values fit here are limited because they have been estimated from one event. There would certainly be value in examining multiple events in estimating Manning's n. We should not want to overstate the importance of the specific Manning's n estimated, but rather note that this provides an approach or an opportunity for estimating a spatially variable Manning's n that warrants further investigation as a way to overcome the limitations and bias associated with a single roughness parameter that was evident in the results. This may provide a way to come up with distributed roughness parameters for use in distributed reach scale hydraulic routing, while noting that the presence of high-resolution satellite imagery might be a limiting factor. It is worth observing that Sichangi et al. (2016) have explored relationships between modified Manning's equation parameters and stage and discharge from remote sensing. Furthermore, the generalization of rating curves, such as addressed in this paper, for stream reaches anywhere there is a good DEM, has potential for use with remote sensing approaches that are being pursued with SWOT (Biancamaria et al., 2010). This is thus a rich area for future research where it is an open question as to the degree to which improved hydraulic parameters would help towards better discharge forecasts, as these also involve routing.

The flow direction conditioning method uses stream hydrography to condition the DEM and provides an improvement whenever the stream hydrography is of a resolution or quality to be a better indicator of flow than the DEM by itself. This method is particularly beneficial where there are artificial barriers that result in flat areas where topographic information is lost during pit filling, common at many transportation stream crossings. The flow direction conditioning method required only two inputs: a DEM and a high-resolution hydrography dataset. In some cases, the lack of a high-resolution hydrography dataset may hinder the applicability of the flow direction conditioning approach. However, in the United States, the high-resolution hydrography mapped at 1:24,000 scale, as used in this study, is available from the U.S. National Hydrography Dataset for the entire continental U.S., and the flow direction conditioning could be applied nationwide. The computational cost is significant. DEM processing steps are essentially done twice: first to determine flow directions in a burned in DEM and then to condition the DEM and repeat the process for the conditioned DEM. Computation time is thus expected to be about double. However, this is something that is done once to prepare the data for the HAND approach, and we feel the improvements merit the extra computation, which can be done quite quickly as described by Liu et al. (2018). There are no additional ongoing computational costs once the flood inundation is being modeled.

6 Conclusions

The first contribution of this paper is flow direction conditioning, a new DEM terrain analysis method that was developed using high-resolution hydrography data to alter, or condition a DEM so that elevation values do not increase moving downstream along hydrographic flow paths. This serves to remove artificial barriers in the DEM due to infrastructure such as road crossings, producing an important improvement in the calculation of HAND and mapping of inundated flooding. In this study, the fit metric that quantifies how well-modeled inundation matched high-resolution satellite observations was improved by 7% for a 10-m DEM and 8% for a 3-m DEM in the Bear River case study, an important improvement given the relatively small but important fraction of the area that flow direction conditioning impacts. Further evaluation of this approach for different study areas is certainly warranted. High-resolution hydrography mapped at 1:24,000 scale, such as used in this study, is available from the U.S. National Hydrography Dataset for the entire continental United States, and this approach could be applied, similar to Liu et al. (2018), at a continental U.S. scale to compute HAND and associated channel hydraulic properties.

The importance of DEM scale (3-m vs 10-m) was also quantified. Higher-resolution DEM data such as the 1/9 arc-sec (3-m) resolution from the U.S. NED for some areas can improve the precision with which flood inundation can be mapped using the HAND approach. In this work, the fit metric improved 10% and 16% with the higher resolution DEM in the Bear River and the Ocheyedan River case studies, respectively. This provides input to consider when evaluating the merit and additional expense of 3-m data collection.

The results presented here have shown that the misalignment between NHDPlus catchments and DEM-derived catchments can be a limitation in the application of the HAND approach to flood inundation mapping. Catchments and stream reaches derived from nodes evenly spread along streams to balance reach lengths helped resolve some of these problems. Using catchments derived from the DEM produced results that were improvements in comparison to those obtained using NHDPlus catchments derived from a coarser DEM.

Lastly, the fixed roughness parameter in CFIM can be a limitation, and this study introduced an approach to estimate reach specific Manning's n from observed flood inundation.

In an effort to make this study reproducible, the data and computational scripts used to produce the study results have been saved in HydroShare (Garousi-Nejad et al., 2019). The code for the flow direction conditioning tool is part of TauDEM and is available from the TauDEM GitHub repository (http://github.com/dtarb/taudem).

Acknowledgments

The authors would like to acknowledge support from the Utah Water Research Laboratory to conduct this research. The preliminary results of scenario 1 were obtained while the author was attending the University Consortium for Geographical Information Science (USGIS) summer school in 2017, supported by the National Science Foundation (NSF). The authors thank Shaowen Wang for his support and Yan Liu for providing insightful comments and help on the results.