Niger is a country located within the semi-arid Sahelian zone of West Africa, south of the Sahara Desert. Southern Niger is characterized by intense rainfall events from June to September with little rainfall in the dry season from October to May (Boko et al., 2020). Temperatures range from 15 to 45°C on average with a mean annual rainfall of approximately 500 mm (Andersson et al., 2017; Boko, et al., 2020). In recent years, land cover change and variable rainfall have resulted in higher flood peaks that arrive progressively earlier in the season (Mahe et al., 2009; Amogu et al., 2010; Aich et al., 2016). These climatic changes, combined with continued occupation of flood-prone areas and rising population growth rates of over 3% (May 2019), have led to an increase in the number of people vulnerable to floods in Niger (Fiorillo et al., 2018), although it is difficult to quantify such vulnerability due to the lack of direct and indirect studies of flood impacts on rural communities in remote areas across Niger.

The region for this study is located along the border of Mali and Niger, about 180 km northwest of Niamey, the capital of Niger (Figure 1). This area is approximately 1,500 square kilometers and located along the Niger River and at an elevation range of 200–300 m. Many areas in Niger are of difficult access due to poor infrastructure and terrorism making it difficult to obtain ground data that can be used for validating DEMs. This study area was selected due to the availability of LiDAR data, collected by the Kandaji Dam Commission Authority, the availability of in-situ data, and the climatic similarity to other flood-prone areas in Niger.

Two standard methods to compare DEMs are through Root Mean Square Error (RMSE) and spatial differencing. RMSE represents the absolute vertical variance, while a spatial differencing represents a spatial analysis of DEM elevation differences across the landscape. To appropriately compare the DEMs to one another, they were re-projected to the one common projection and vertical datum (Figure 2).

The Height Above Nearest Drainage (HAND) is a terrain model which normalizes the topography to the local relative heights along a drainage network and is widely accepted for simple hydrologic applications (Rodda, 2005; Nobre et al., 2011). HAND has been applied to a variety of applications from the creation of synthetic rating curves (Zheng et al., 2018) to simple flood modeling (Garousi-Nejad et al., 2019). The primary inputs into HAND are a flow accumulation raster and an extracted stream elevation raster (Figure 2). These inputs require additional DEM pre-processing including hydrologically conditioning by filling for sinks and calculating the flow direction before determining the flow accumulation. The resultant HAND raster shows the height of each pixel above the nearest drainage point. From the HAND raster, the surface area at incremental flood depths is calculated by selecting the pixels that are below a specified height above the nearest drainage.

The AW3D30 2.2 had the lowest RMSE with a value of 1.19 m while the ASTERv3 had the highest overall RMSE of 4.22 m. SRTM 1 Arc Sec and MERIT fell in between with RMSEs of 2.38 and 1.98 m, respectively. The low RMSE of the AW3D30 2.2 indicates the DEM differed the least from the LiDAR DEM while the higher RMSE of the ASTERv3 indicates the DEM differed the greatest from the LiDAR DEM.

The RMSE was also evaluated for 10-m elevation increments from 200 to 280 m to determine if there was a relationship between the RMSE and the elevation heights. The elevation ranges were equal in elevation range but not size; therefore, sample points were randomly selected in proportion to the percentage of the total area. The higher in elevation the band, the larger the RMSE for the given DEM. The AW3D30 had the smallest range of RMSEs from 1.74 to 4.61 while the MERIT DEM had the largest range of RMSEs from 1.63 to 12.55 (Table 2).

Based on the histograms of the DEM difference rasters in Figure 3, the SRTM and MERIT DEMs appear to underestimate the height, the AW3D30 overestimates the height, and ASTER both overestimates and underestimates the height across the study area. On average, SRTM and MERIT underestimated by 2.82 and 4.44 m, respectively, while AW3D30 overestimated by 1.92 m. Generally, the AW3D3 DEM 0slightly overestimated across the entire study area (Figure 3A). The SRTM DEM underestimated the elevation primarily in the northern part of the study area (Figure3B), overall these differences are fairly small compared to the MERIT DEM. The MERIT DEM underestimated primarily around the rivers, particularly along the braided channel in the southern part of the study area (Figure 3C).

In contrast, the ASTER elevation differences across the study area were on both extremes of overestimating and underestimating the elevation which averaged to an overall underestimation of 0.92 m. ASTER overestimated elevation along the primary Niger River riverbeds and underestimated in the land surrounding the braided channel (Figure 3D). These elevation differences are highlighted in Figure 3D as the only DEM with the largest range of elevation differences.

The 1,500 square kilometer study area is characterized by low topographic relief and small tributaries that flow into the larger Niger River. The flood surface area derived from the HAND model was calculated at 1 m flood depth increments, shown in Table 3. At the 1-m flood depth, ASTER covered the most surface area with 156.48  k m 2 while MERIT had the lowest surface area with 80.25  k m 2 . This was a similar pattern up through the 3-m flood height, where the LiDAR DEM surface area exceeded all other DEMs at 242.13  k m 2 . The surface areas for ASTER, AW3D30, and SRTM were fairly similar to one another for the 3-m flood depth at 238.90 k m 2 , 232.29 k m 2 , and 238.67  k m 2 , respectively. As detailed in Table 3, the pattern of flooded surface area in relation to the DEM resolution continued through the 10 m flood depth where the flood extent surface areas were, from greatest to least, 441.37, 413.21, 412.08, 400.33, and 394.35  k m 2   for LiDAR, AW3D30, SRTM, ASTER, and MERIT, respectively.

The RMSE was calculated for the HAND values of the global DEMs were compared to the LiDAR HAND values at 1-m increments. There is no distinctive pattern of a higher or lower HAND value being better in the global DEMs. The RMSEs of the HAND values for the whole study area were 2.98, 3.49, 2.44, and 3.05 for the SRTM, ASTER, AW3D30, and MERIT DEMs respectively. The ASTER RMSE was on the lower side for the first two HAND levels, but one of the highest for the remainder. The AW3D30 RMSEs remained one of the lowest RMSEs throughout all HAND levels (Table 4).

The flood extent values for the DEMs were graphed to better depict how the extents changed per the DEM as the flood level increased (Figure 4). The general trend seen at higher flood levels (greater than 3 m) is the LiDAR DEM had the greatest surface area, while the MERIT DEM had the least surface area. Under 3 m flood depth, the MERIT DEM still had the least surface area, but the LiDAR DEM also had less surface area than the AW3D30, ASTER, and SRTM DEMs in contrast to situations in which the flood depth is greater than 3 m (Table 2; Figure 4).

The differences in the flooded surface area are more evident in certain parts of the study area, such as at the convergence of the rivers in the middle of the study area at the 6-m flood depth (Figure 5). The 6-m flood depth depicts the DEM flood extent differences well, where the LiDAR has the largest flooded surface area, the 30 m DEMs are about the same surface area, and the MERIT has the least flooded surface area.

The primary driver of the differences in the flood surface area is the resolution of the DEM (5, 30, or 90 m). Figure 5 depicts an example of how the DEM affects the flood surface area when the flood depth is at 6 m. There is one area in the middle of the image, northwest of the larger part of the river, where a large difference can be seen between the DEMs. The LiDAR flooded area almost covers the whole section, whereas the MERIT is not flooded. The AW3D30, ASTER, and SRTM flooded areas have somewhere in between the LiDAR and MERIT, with SRTM having a more flooded surface area in the region than ASTER and AW3D30.

Differences in flood extents derived from the HAND model were caused by the elevation differences (Figure 6). At the 2 m flood height, the overall surface area of ASTER was greater than the LiDAR DEM flood surface area at 201.63 k m 2 and 163.38  k m 2 respectively. The areas where the ASTER underestimated the elevation height, indicated in blue in Figure 6C, were areas that flooded in the ASTER 2-m flood height than in the LiDAR DEM. Additionally, one of the river beds for which ASTERv3 overestimated the elevation height did not show as flooded in the ASTERv3 flood surface map (Figure 6A), but was shown as flooded in the LiDAR DEM (Figure 6B).

The range of RMSEs are a result of differences in land cover, processing methods, and date of data acquisition. The four global DEMs were edited and/or released in the last 5 years; however, much of the input data to create the DEMs are 10–20 years old. Landscape changes due to tree growth, the formation sand dunes, and flood plain erosion and sedimentation, all contribute to uncertainties in freely available global DEMs. Land use changes and other built structures can heavily influence flood extent in ways that global DEMs cannot capture. Vegetation can increase the DEM elevation error by 5 m or more, depending on the vegetation type (Weydahl et al., 2007; LaLonde et al., 2010). In semi-arid regions, such as the Sahel in West Africa, it is common for the presence of riparian vegetation and sparse vegetation throughout the rest of the landscape. The presence of vegetation along the rivers, and limited elsewhere, can significantly impact the flow of the rivers and the delineation of the drainage basin (Oksanen and Sarjakoski, 2005).

Along with date of data acquisition and land cover, the input data were acquired and processed differently. SRTM 1 data was acquired via C-band radar while ASTERv3 data were stereo-pair optical imagery. The acquisition processes result in variations in the portrayal of elevation in the DEM. The optical imagery represents the tops of trees as the elevation at a given point, while the radar can penetrate partially into the tree canopy, resulting in a lower elevation. The AW3D30 DEM data were also optical stereo images, while the MERIT DEM is an ensemble of multiple DEMs with different input data types. The input data for the DEMs determined the data processing techniques which are characterized by specific errors. InSAR, the process to create the DEM from radar data, is characterized by random errors through speckle noise (Farr et al., 2007), while the stereo-pair processing can lead to blunders and stripping in the DEM (Tarekegn and Sayama, 2013). These errors impact the overall accuracy and RMSE.

The RMSEs for all four DEMs were lower than those of previous studies as well as the global RMSEs (Florinsky et al., 2018; Kramm and Hoffmeister, 2019; Yap et al., 2019). All four of the DEMs had a RMSE of at least 4 m less than their globally published RMSEs (Table 1). SRTM products differed from previous literature for RMSE and the overall elevation. In this study, on average, the SRTM DEM was lower than the LiDAR DEM by 2.82 m; however, studies found that previous versions of the SRTM DEM are on average higher globally than truth datasets which is why local studies such as this one are important (Li and Wong, 2010; Jarihani et al., 2015; Bhuyian and Kalyanapu, 2018). The ASTER DEM coincided with these studies as on average it was lower in elevation than the LiDAR DEM. The discrepancy in the average elevation differences for the SRTM DEM and the RMSEs for all DEMs could be attributed to the version of the DEM used in the study as well as the study area characteristics. This study used the most recently released versions of the DEMs, while most of the studies used earlier versions. Additionally, most DEM comparison studies are in areas with high topographic relief and vegetation, in contrast to the low topographic relief and sparse vegetation of Niger (Florinsky et al., 2018; Watson et al., 2019). There are strong correlations between vegetation height and DEM error (Kellndorfer et al., 2004), therefore the presence or absence of vegetation is a plausible driving factor in the RMSE and elevation differences between previous studies and this study.

The study corroborated previous studies that found the DEM resolution may impact the flood model (Poulter and Halpin, 2008; Gallegos et al., 2009) as demonstrated in the flood depths from three to 10 m where as spatial resolution increased, the flood extent area decreased (Table 2). However, the flood depths below the 3 m HAND value did not correlate to the resolution. Before the 3-m depth, the 30 m resolution DEMs had the largest flood extent, then the 5 m DEM, followed by the 90 m DEM. This suggests that DEM resolution is not the only factor and may not be the dominant factor that influences variations among flood results (Li and Wong, 2010).

While resolution and total flood extent surface area are related, there are differences amongst the DEMs for where flooding occurs at a given flood depth which is not directly correlated with the resolution. For example, at the 6-m flood depth (Figure 5), the intersection of the two main branches in the study area is flooded in the LiDAR DEM and not flooded in the MERIT DEM based HAND model flood extents. While these differences coincide with the resolution, the 30 m DEMs vary in their flood extents. ASTERv3 is almost completely flooded in this area, like the LiDAR DEM, while AW3D30 and SRTM are less flooded. If resolution were the only factor influencing flood estimation, these DEMs should have similar flood estimates (Hawker et al., 2018). The variations in the flood extent within the 30 m resolution DEMs must then be related to the other sources of error such as random error, blunders, or processing methodologies.

ASTER, SRTM, and AW3D30 all had different spatial distributions of elevation differences (Figure 3) which translated into variations in flood extent at the different flood heights. For example, the elevation of the DEM has an impact flood extent where the ASTER elevation is higher than the LiDAR along the central western edge of the study area. AW3D30 overestimates elevation in this area which translates into no flooded area in the 6-m flood height, while SRTM underestimates the elevation and is flooded for this area (Figures 3, 4). The location and distribution of the elevation differences for a given DEM impact where the model simulates flood extent. In the southern part of ASTER, the riverbed elevations are higher than the LiDAR DEM (Figure 6). This overestimation of elevation leads to one of the rivers not being flooded in the 2-m flood height while in the LiDAR DEM the river is flooded. The relative accuracy and relationship of the elevation measurements to one another are further demonstrated with the RMSEs of the HAND values. As each of the DEMs over or underestimates elevations are various points in the study area in comparison to the LiDAR, their HAND values also differ from the LiDAR DEM. The AW3D30 DEM had the lowest RMSE for both the absolute accuracy and the relative accuracy when compared to the LiDAR DEM while the ASTER DEM had the highest RMSE for both. While these two assessments agree which of these two DEMs are most like the LiDAR DEM, the SRTM and MERIT DEMs are flipped. This illustrates that it is important to understand the holistic impact the choice of the DEM may have on the analysis in selecting one for analysis. While a higher resolution DEM such as the LiDAR DEM would provide the most detail and information for a decision maker, this is not realistic, especially in data sparse areas. Therefore, the information from this study can be used to assist in the selection of the input DEM for the flood analyses.

Local water management authorities often rely on one flood model with one DEM input to understand which areas are most vulnerable to flooding and to make decisions on providing support. The differences in DEM resolution as well as the overall absolute and relative accuracy impact the results of flood models. The coarser the resolution of the DEM, the less the smaller tributaries are captured in the topography and the less the smaller differences in elevation change are captured resulting in less flooded surface in the models. In areas with low topographic relief, such as Niger, where 2-m RMSE is the difference to a house flooding, high-resolution DEMs are important for flood planning. The recently released Forest and Building Removed Copernicus DEM (FABDEM) (Hawker et al., 2022) may improve upon the effects of the riparian vegetation impacts. Additionally, the high-resolution 2-m EarthDEM is an upcoming DEM that will be open to the public (NCSA, 2021) and can be used in applications that require elevation data higher resolution than 30 m. As these new technologies become available it is important to increase the understanding of the impact of DEMs on flood modeling and move beyond the impact of the other hydrologic components (Bhuyian and Kalyanapu, 2018) so decision makers can plan accordingly with the best information possible.