Effectiveness of UAV‐based DTM and satellite‐based DEMs for local‐level flood modeling in Jamuna floodplain

Open‐source, satellite‐based digital elevation models (DEMs) are widely used for flood modeling. However, studies on effectiveness of these DEMs in depicting local‐level flood processes are limited. This study generated a high‐resolution digital terrain model (DTM) based on unmanned aerial vehicle (UAV) photogrammetry and used in a two‐dimensional (2D) hydrodynamic model (HEC‐RAS) to simulate the flood processes in a floodplain environment of the Jamuna River in northern Bangladesh. The effectiveness of a few satellite‐based DEMs was also compared with this DTM by using the DEMs in the same hydrodynamic model. Field data for two flood seasons were collected to develop the model. The results indicate that the 2D model with UAV‐based DTM provides the flood parameters, such as flood arrival time, depth, duration and extent, better than those from the satellite‐based DEMs. Of the open‐source DEMs, the FABDEM and the WorldDEM™ have the least errors and provide better results compared to the SRTM30, ALOS PALSAR, and ASTER DEMs. The UAV technique with ground control points and field measurements for the tree‐canopy and water areas is very useful in generating a fit‐for‐purpose DTM. The findings of this study would be useful for terrain generation and DEM selection for local‐level flood modeling elsewhere.


| INTRODUCTION
Two-dimensional (2D) hydrodynamic modeling is needed for understanding flood propagation processes and improving flood forecasts at local levels (Iqbal et al., 2022;Meire et al., 2010).A 2D flood model can represent mixed flow regimes, highly dynamic flood waves, abrupt contraction and expansion, general wave propagation, and super-elevation around bends (Hydrologic Engineering Center, 2018).However, the results of the 2D model can be influenced by several factors, including topography, input data accuracy, catchment area characteristics, and infiltration (Yalcin, 2020).More importantly, the 2D model is highly sensitive to the quality of the utilized digital elevation model (DEM) (Jain et al., 2018;Saksena & Merwade, 2015).2D models typically rely on the open-source DEMs available from remote sensing satellites (Iqbal et al., 2022;Jain et al., 2018;Masood & Takeuchi, 2011;Rahman, 2015).
Different DEMs, such as advanced land observing satellite (ALOS) phased array L-band synthetic aperture radar (PALSAR), shuttle radar topography mission (SRTM), multiple-error-removed-improved-terrain (MERIT), advanced spaceborne thermal emission and reflection radiometer (ASTER), and TanDEM-X WorldDEM™, are widely available and used (Guan et al., 2020;Wessel et al., 2018;Yamazaki et al., 2019).Recently, the forest and building removed digital terrain model (FABDEM) from Copernicus GLO-30 has become popular (Santillan, 2023).The flood-modeling studies in developing countries like Bangladesh mostly depend on these satellite-based, coarse-resolution, open-source DEMs (Biswas & Mondal, 2022).However, those DEMs have issues with accuracy in spatial and temporal resolutions and vertical elevation as most of them (except for WorldDEM™ and FABDEM) are Digital Surface Model (DSM), which includes the top surfaces of trees and buildings (Franci et al., 2016;Kundu et al., 2014;Munir & Iqbal, 2016).Therefore, a vertical accuracy issue with the highest root mean square error (RMSE) from 12.62 m to 17.76 m is found for SRTM-90 m, SRTM-30 m, and ASTER-30 m DEM (Azizian & Brocca, 2019;Mukherjee et al., 2013;Rabus et al., 2003).These global DEMs also have issues with the detail of terrain features and microtopographic variations in relatively flat terrain (Chu & Lindenschmidt, 2017;Gallien et al., 2011).Sometimes, DTMs are generated from traditional ground-based spot-level measurements for local-level flood studies (Biswas & Mondal, 2022).However, this practice also cannot represent the full terrain surface accurately due to the sparse distribution of the spots, or low resolution between the points.These global or ground-based DEMs can be adequate for predicting the average flood depth and extent at relatively large scale such as basin, and may not be adequate for flood modeling at small scale in a dynamic setting such as a micro-watershed in a floodplain.However, the importance of fine spatial resolution and accurate vertical elevation is increasing day by day to meet the requirements of local-level flood assessment through inundation prediction and quantification (Saksena, 2015;Schumann & Bates, 2018).
Use of unmanned aerial vehicles (UAVs) is very helpful for generating an accurate high-resolution DEM (Ahmed & Mahmud, 2022;Hashemi-Beni et al., 2018).It generates the DEM from the images with close-range sensing, providing a platform for different sensors such as visible and infrared sensors, spectrum analyzers, and LiDAR sensors (Jakovljevic et al., 2019).When coupled with surveyed ground control points (GCPs), UAV-based photogrammetry can capture spatial data with a richness of detail that can meet high standards (Hashemi-Beni et al., 2018;Kardasz & Doskocz, 2016;Serban et al., 2016).Photogrammetry from aerial images collected by the UAV combined with postprocessing has become a very promising methodology because of the data acquisition speed, automation of data processing, and gathering more accurate geoinformation (Bazzoffi, 2015;Chiabrando et al., 2018;Watanabe & Kawahara, 2016;Westerveld, 2020).More importantly, the decreasing cost of the technology is making it available for a wide range of users.The UAV can give a DEM more accurate elevations with GCP surveys even in low grass vegetation areas (Coveney & Roberts, 2017).The importance of UAV in hydrology and water management has already been demonstrated (Acharya et al., 2021).A substantial increase in resolution and accuracy of flood information in floodplains can be achieved with UAVbased DTM (Langhammer et al., 2017).
Some research has been conducted to assess the performance of different satellite-based DEMs in landslide susceptibility, river morphology and watershed analysis, where different DEMs are shown to produce different results (Arash & Yasi, 2022;Chanu & Bakimchandra, 2022;Karabulut & Özdemir, 2019;Pryde et al., 2007;Rabby et al., 2020;Saleem et al., 2019).Also, studies are available on the use of UAV-based DTM in flood modeling in Indonesia (Derka et al., 2022), South Korea (Lee et al., 2022), China (Li et al., 2021), Italy (Annis et al., 2020), Philippines (Gagula et al., 2023), Ghana (Trepekli et al., 2022), Colombia (Escobar Villanueva et al., 2019), Turkey (Yalcin, 2020), and Poland (Karamuz et al., 2020).However, comparative studies on effectiveness of satellite-based DEMs and UAV-based DTM in local-level flood modeling are scarce (Adesina et al., 2022).Moreover, the available studies have reported only the differences in results between UAV-and satellite-based DEMs without providing further insights as to why and where such differences occur on the ground with primary measurements and observations.Such insights are useful for a hydrodynamic modeler to be aware of the level of potential errors that could arise due to the use of different DEMs and in identifying the DEMs to use in certain situations for plausible flood parameters.Furthermore, the effectiveness of these DEMs in Bangladesh's local-level flood propagation is yet to be reported.Since the floodplain of Bangladesh like the chars of the Jamuna River is very active and dynamic, and shifts frequently and changes elevation with time (Mondal et al., 2015), it may be very hard to capture those dynamic aspects with static satellite-based DEMs.Thus, this study was undertaken to compare the effectiveness of a UAV-based high-resolution DTM with that of different open-source DEMs for local-level flood simulation in Bangladesh.

| STUDY AREA
Bangladesh has 80% of floodplain land area and about 25% of the land area is flooded in an average flood year.Many communities reside in the unprotected, low-lying, riverside floodplains, which are flooded by the river water overtopping the banks.Such communities are exposed and vulnerable even to the flood events associated with low return periods (Haque et al., 2019;Nahin et al., 2022).Due to climate change, land use change and riverbed siltation, flood events are becoming increasingly unpredictable and severe.As a consequence, the historical practice of living with water is being threatened by rising flood impacts (Mondal et al., 2018;Zevenbergen et al., 2008).A proper assessment and modeling of floods in the floodplain is therefore necessary for appropriate flood management with effective planning, design and implementation of mitigation measures (Ali et al., 2018;Brammer, 2010;Hashimoto et al., 2021).
This study is conducted in a floodplain of the Jamuna River in the northern part of Bangladesh (Figure 1).The floodplain is located in the Sirajganj district, which is one of the most flood-prone areas of Bangladesh (Ali et al., 2018).It is flooded primarily by the riverine flood during the monsoon.The most vulnerable areas in the district are the low-elevation flat areas made up of deposited silt in the river.These areas are locally called the "char" (island) areas (Ali et al., 2018;Mondal et al., 2015).The Ranigram village, which is an attached char of the Jamuna River, located in the Khokshabari union (lowest tier of local administration) of the district is selected for this study.The village is located on the western side of the Jamuna River and has a strong hydrological linkage with the river.It has also some peri-urban characteristics as it is close to the district headquarters.
Ranigram is flooded every year and flood conditions are unpredictable from year to year.Along with the fragile environmental and hydrological conditions, many physical structures and interventions made in the floodplain over the years have worsened the inundation condition of Ranigram even in subcritical floods.The study area is 2.1 km 2 and is located outside the new floodcontrol embankment of Sirajganj town.The area is selected because different types of land features exist Location of the study area beside the Jamuna River in the Sirajganj district of northern Bangladesh.
simultaneously in the village.Figure 2 shows an orthoimage of the area taken with a drone survey.It is surrounded on three sides by embankments.There are two small canals beside the union parishad (UP) and at the culvert location through which floodwater can reach the village.Ranigram has agricultural lands, residential areas, khals (Bengali for "natural canals"), a culvert, and areas of tree canopy, which make the village a study area with diverse components (Figure 2).There is a sudden change in elevation in the agricultural lands as well as at the interventions like roads, marketplaces, and houses.Due to these very diverse and dynamic aspects of the area, an outdated, coarse-resolution, open-source DEM may not be adequate to understand the flood processes in the area, and a near-real time, fine-resolution topographic and bathymetric information is necessary for local-level flood assessment.Two crossbars exist in the Jamuna River on the upstream and downstream of the village area.Ranigram was not included for flood protection by the town protection embankment, and so the area faces extreme floods during the monsoon.Widespread flooding, shifting river channels, constantly eroding cultivated land and settlement, and displacement of people are the main problems in this area.Floodwater gets stored in the lowland area from UP to Culvert after entering the village through the breaching beside UP (Figure 2).With the further increase in water level, the flood water moves to the northern part of the village through the canal beside UP and to the southern part up to the Pilot Site through the canal below the Culvert.

| METHODOLOGY
The study was conducted using a structured methodology.At first, a UAV-based high-resolution DTM was generated and a few satellite-based DEMs were collected.Then, a 2D flood model was developed for Ranigram using the Hydraulic Engineering Center's River Analysis System (HEC-RAS 5.0.7)software.Thereafter, the flood parameters resulted from different model setups with different DEMs were compared.Finally, the vertical errors in the satellite-based DEMs were investigated for the floodplain of Ranigram by comparing them to the UAVbased DTM.The following subsections elaborate these steps.

| UAV-based high-resolution DTM
A high-resolution DTM was generated utilizing the UAV incorporated with measured topography and bathymetry data of Ranigram following a hybrid methodology (Langhammer et al., 2017).The Agisoft Metashape 1.5.2 photogrammetry and ArcGIS 10.5 software was used to generate and process the DTM (Serban et al., 2016;Yamazaki et al., 2017).The methodological framework used in the generation of the DTM is shown in Figure 3.
The processed raster data from UAV was classified into two categories: (i) open noncanopy area and (ii) treecanopy and water area.For the noncanopy area, the UAV-generated data were directly used.For the treecanopy and water area, measured topography and bathymetry data were used and the DTM was created by a triangulated irregular network (TIN) process.Thus, combining the UAV-based and field-measured elevation data by this hybrid approach, a more accurate and highresolution DTM of Ranigram was generated.

| UAV imaging, ground truthing, and postprocessing
Three days of field data collection including drone and GCP surveys were conducted in Ranigram in 2018.A Mavic 2 Pro drone was used in the survey.The camera of the drone was calibrated before flying over the Ranigram area.The flying direction was set to cover an area of 3.3 km 2 , and the vertical imaging technique was followed for taking the photos.Flying altitude was 115 m from the ground level, and the ground resolution was 2.16 cm/ pixel.The focal length pixel size was 2.41 Â 2.41 μm.A total of 3981 images were taken by the drone, having 3,347,840 tie points.
A total of 92 GCPs were taken for creating and validating the DTM, where 46 points each were used for creating and validating the DTM.The GCPs were taken in such a manner that the diverse land features and land uses of the area could be captured (Figure 4a).The locations of the GCPs on the land surface were marked by using colored spray cans (Figure 4b), and the coordinates were collected using a kinematic GPS machine.The points became visible in the UAV images and hence could be detected in the photogrammetry software.The reduced levels (RLs) of the GCPs were measured with a proper benchmark connection.The accuracy of the DTM was checked with the RLs of the validation GCPs (Figure 4a).The RMSE of the DTM at the GCPs was found to be very low (Table 1).Thus, the generated DTM could represent the terrain of the floodplain of Ranigram very well.
A georeferenced 3D point cloud of Ranigram was produced and a DTM as well as an orthophoto in a defined coordinate system were generated from the UAV image data in the Agisoft Metashape software.In the software, all the photos were aligned, and the geometry and texture were built for a realistic appearance.For a geomatic application, photos with the painted-spray-can marks were aligned to the locations and used to geo-reference the image.At the end of the image processing, a dense 3D point cloud was generated.

| DTM generation for open noncanopy, tree-canopy and water areas, and artifacts correction
As the photogrammetry images cannot penetrate the trees and water to reach the ground, photogrammetry is unable to determine the actual terrain level underneath the tree canopy and water areas.The area, where the terrain is clearly visible from the close sensing images of UAV, is accurately represented in the DTM.The whole area was divided into two categories (Figure 5a) for selecting the areas of different data.The shapefiles of the tree-canopy and water areas as well as the open noncanopy area were drawn from the field observations.The DEM data from the UAV were taken for the noncanopy area.For the tree-canopy and water areas, the topographic and bathymetric data, respectively, were measured during the field survey (Figure 5b).
From the dense point cloud generated after processing, the UAV data and validating with the GCP elevations, a DTM at a resolution of 17.5 cm per pixel (Figure 5c) was generated for the noncanopy area.The measured topography and bathymetry data for the treecanopy and water areas, respectively, were also processed for creating another DEM.With the point elevations, using the 3D analyst tool of ArcGIS, a TIN was then created (Watanabe & Kawahara, 2016).From the TIN, a DTM was made in the 3D analyst using the TIN to Raster tool (Figure 5d).The cell size was 17.5 cm, the same as the UAV DEM.
In the UAV, some artifacts were caused by the presence of water and shadow in the field.In some areas, no data value were identified and corrected from the field observed values.The correction of these artifacts comprises the extraction, void filling, and editing of these features, and the elevations of these areas were fixed with field observations and interpolation (Hashemi-Beni et al., 2018).
Some manual correction was also required to ensure the quality of the DTM.There was a small culvert below a road through which water could pass from the north to the south sides of the village.In the drone data, the top level of the culvert was treated as a road.Hence, the culvert location was selected in the DTM, and edited manually with the bed elevation as the land surface.
Finally, the two different DTMs from the UAV and measured elevations were mosaiced into a single raster to obtain a final high-resolution DTM of Ranigram (Figure 5e).The resolution of the combined DTM was 17.5 cm only.

| Selection and processing of satellite-based DEMs
Five satellite-based DEMs, namely, WorldDEM™, SRTM30, ALOS PALSAR, ASTER, and FABDEM, were used in this study.The horizontal coordinate system of all the DEMs is in WGS84, and the vertical coordinate system of the DEMs except for ALOS PALSAR is in mean sea level (MSL).The vertical unit of all the DEMs is in meters.A brief description of each DEM is given below: For this study, the vertical elevation of the DEMs was converted to the Public Works Datum (PWD) from the MSL datum as the collected flood water level data were in the PWD datum.The MSL datum is 0.46 m higher than the PWD datum.The DEMs are shown in Figure 6.

| Data collection
The daily floodwater data at the UP, Culvert, and Pilot Site locations of Ranigram were collected during the monsoon season of 2018 and 2020 by installing three water level gages (Iqbal, 2021;Iqbal et al., 2022).The gages were installed with proper referencing to the national benchmark maintained by the of Bangladesh.These flood data were used as boundary conditions for calibration and validation of the hydrodynamic model.A local gage reader was trained and engaged to collect the daily water levels from the three gages in the floodplain.Also, the flood photos at select locations of the village were collected marking the date, time, and location.These also helped identify the spatial distribution of the floods along the village for the calibration and validation of the 2D model.In addition, the daily flood data of the Jamuna River were collected from the Flood Forecasting and Warning Center of BWDB.The Ranigram area was flooded for about three and a half months from late June to early October in 2018, and for about 5 months from early June to late October in 2020.

| HEC-RAS model development
The HEC-RAS 2D model was used to understand the floodplain hydrodynamics of Ranigram (Farooq et al., 2019;Iqbal, 2021;Iqbal et al., 2022).HEC-RAS can simulate the water surface profiles for 2D unsteady flow (Biswas & Mondal, 2022;Chyon et al., 2023;Mehta et al., 2021).The RAS Mapper was used for the geometry data input and processing.The high-resolution DTM of the Ranigram area was imported for the terrain.The 2D flow perimeter of the study area was set up with the imported shapefile of Ranigram.Then, a computational mesh within the boundary layer was developed with 2 Â 2 m 2 computational point spacing, resulting in a total of 21,500 grid cells.
Two external boundary condition lines were drawn.The upstream boundary was beside the UP, and the measured daily flood water level was used as a stage boundary.The downstream boundary was at the south of the Ranigram village beside the Pilot Site where the normal depth was used as the boundary condition.The normal depth was calculated from the measured floodwater levels in the upstream and downstream locations, and the average normal depth was found to be 0.0001.An unsteady flow simulation was conducted for understanding the flood propagation in Ranigram.The option for using the full momentum equation sets based on the continuity and momentum equations (Quiroga et al., 2016) was selected.
The model was calibrated for 2018 with different Manning's roughness coefficient (n) values for the associated land uses.The village was divided into two land use classes as the agricultural and housing lands.The n values of 0.03-0.06were used for the calibration of the model with different trial values for different classes.
The water level at the culvert location and the spatial extent of the flood water were matched during the calibration of the model.The n values of 0.04 and 0.05 for the agricultural and housing lands, respectively, showed the best agreement in the calibration.Also, the computational time interval had a role in the calibration, which was finally set to 1 min.Other parameters were set as default.The calibrated model was validated independently for the 2020 flood.
The model was run for the full monsoon season from June to October and the outputs were recorded at a daily interval.The model was run again using different DEMs as the terrain to investigate the influence and effectiveness of the satellite-based open-source DEMs.The calibration parameters found for the UAV-based DEM were maintained in these DEMs.The maps of flood parameters, that is, flood extent, depth, duration, arrival time, and percent of time inundated, were exported from the different models and were compared.

| Assessing errors in vertical elevations of satellite-based DEMs
The errors in vertical elevations of the satellite-based DEMs were calculated by taking the UAV-based highresolution DTM of Ranigram as the base elevation model.At first, all the DEMs were converted to a 30 m resolution using the resample tool of the 3D analyst.Then the elevation of the satellite-based DEMs was subtracted from the UAV-based DTM by the raster calculator of the spatial analysis tool.The spatial distribution maps of the errors were exported.Finally, the raster was reclassified with the 3D analyst tool to count the cells with classified error range, and the area with a specified error class was calculated.

| Flood parameters with UAVbased DTM
The 2D HEC-RAS model built with the high-resolution DTM was calibrated and validated at the culvert location.In the calibration with the measured flood data of 2018, the R 2 value was found to be 0.98, the Nash-Sutcliffe efficiency (NSE) 0.91, mean absolute error (MAE) 0.043 m, and RMSE 0.068 m (Iqbal et al., 2022).The residuals ranged from À0.07 to 0.34 m.In the validation with the measured flood data of 2020, the R 2 value was found to be 0.96, the NSE 0.82, MAE 0.055 m, and the RMSE 0.12 m.The residuals ranged from 0.69 to À0.08 m.Thus, the model was found to represent the flood scenarios of Ranigram very well from the and validation.Further details on the calibration and validation of the model can be found in Iqbal (2021) and Iqbal et al. (2022).
The result of the model was also compared with the real flood scenario of 2020.The flood photos were collected from a few select locations of Ranigram.On June 28, July 8, July 15, and August 29, 2020, the flood photos were taken from the local market beside the UP, retrofitted house, Pilot Site, and Culvert.The flood maps for the corresponding dates were also exported from the HEC-RAS model results, and were compared at the locations of the photos (Figure 7).The model-simulated flood maps showed similar flood scenarios when compared with the actual flood photos.It again appeared that the HEC-RAS 2D model built with the UAV-based highresolution DTM performed very well in representing the flood scenario of the floodplain.
The simulation results of important flood parameters, that is, arrival time, duration, percent time inundated, and maximum depth and extent, for the 2020 flood season (Figure 8) were also found to match the field observations.The floodwater arrived very quickly in the lower lands of the central Ranigram area and stayed for a very long period in 2020.Entering from the breaching location, the floodwater reached the canals and lower agricultural lands within one day, and the medium-high agricultural lands within a week.The floodwater took 3 weeks to reach the elevated housing lands in the northern and southern parts of Ranigram.The lower agricultural lands of the central Ranigram area faced flooding for the whole flood period of about 5 months.The elevated housing lands were affected for about 2 months.The agricultural lands were inundated with 3-4 m depth, and the housing lands with 1-1.5 m depth.Overall, the flooding in Ranigram in 2020 was very severe, which was captured well with the HEC-RAS model built with the UAV-based DTM.

| Flood parameters with satellitebased DEMs
The HEC-RAS model was run again using the different satellite-based DEMs to compare its performance when From the maximum flood extent and depth maps, it is observed that the flood was extended on the lower agricultural lands for the WorldDEM™ and ASTER DEM, but there was no flood on the elevated housing lands.The maps with the SRTM and ALOS PALSAR DEMs show the flood extent only on the agricultural lands near the breaching point, and the flood depths were also very low.FABDEM shows flooding all over the village except for some southern housing lands.Though FABDEM shows closer results to the UAV-based DTM, there is high inundation in the culvert area and eastern side of the village.The higher inundation in these areas could be due to low elevation in the DEM is because of the effect of the canal in the riverside in this 30 m resolution DEM.The flood depths on the agricultural lands with the WorldDEM™ were similar to the depths with the UAVbased DTM.However, the flood depths with the ASTER DEM were very much scattered and did not follow the result of the UAV-based DTM.The maximum flood depths on the agricultural lands were 2 m to 4 m with the WorldDEM™ and UAV-based DTM, whereas the depths were 4-9 m with the ASTER DEM.
The overall model results with the SRTM and ALOS PALSAR DEMs deviate very much from the real field scenario.The canal beside the UP, through which the water reaches the northern part of Ranigram from the breaching location, is missing in these DEMs because of their low spatial resolution.So, the floodwater does not enter the northern Ranigram.Also, the opening below the culvert is not present in these DEMs as it cannot be observed from an oblique remote sensing.So, the flooding in the southern part of Ranigram cannot be captured by using these DEMs.

| Errors in vertical elevations in satellite-based DEMs
The errors in vertical elevations in different satellite-based DEMs for the Ranigram area are shown in Figure 9.A negative elevation in the figure indicates that the DEM elevation is higher than the actual, and vice versa.Most of the errors in the DEMs are found to be negative.Only in a few portions, where there are water bodies and canals, the errors are found to be positive.The highest negative errors are found for the tree-canopy areas, where remote sensing cannot penetrate onto the terrain.Also, the highest positive errors are found for the embankment portion, as it was not reflected in the low-resolution DEMs.A substantial portion of the SRTM and ALOS PAL-SAR DEMs has errors of more than 2 m.The errors in the tree-canopy areas in both DEMs are extremely high, ranging from 4 to 12 m.Even in the noncanopy land, the errors are from 2 to 4 m.The errors in the ASTER DEM are from 1 to 7 m, and those in the WorldDEM™ are from 1 to 4 m.The errors in the tree-canopy areas are from 4 to 7 m and 2 to 4 m in the ASTER DEM and WorldDEM™, respectively.The corresponding errors for the noncanopy areas are 1-4 m and 0-1 m, respectively.The error is comparatively less in the FABDEM with green color dominating where the error is within 1 m.The FABDEM is 1-2 m high in the cyan-colored area in the northern and southern housing areas and in some parts of the southern lowlands.In the yellow-colored area near the UP, culvert and pilot site, the FABDEM is 1-2 m lower than the UAV-based DTM.This DEM has lower elevation in the embankment areas of the mid-west and eastern Ranigram.So, the high-land features are not captured well in this lowresolution DEM.
The errors associated with the different DEMs were quantified to find out the areal extents of the errors.Figure 10 shows the area-wise error intensities in the DEMs for the Ranigram area.Only ASTER DEM and FABDEM have more than 1 m of positive error for an area of 0.2 km 2 and the positive errors in the other three DEMs are negligible.The lowest error of À1 to 1 m was for 0.55, 0.48, 0.60, 0.13, and 1.13 km 2 in the ASTER, WorldDEM™, SRTM, ALOS PALSAR, and FABDEM, respectively.A negative error of 1-4 m was prominent in the ASTER DEM and WorldDEM™, and that of 4-7 m was prominent in the SRTM and ALOS PALSAR DEMs.In the FABDEM, the negative error was 1-2 m for 0.53 km 2 .The WorldDEM™ has an area of 0.76 km 2 with negative error of 2-4 m, and the ASTER DEM and FAB-DEM have an area of 0.53 and 0.14 km 2 , respectively, with the same negative error.The SRTM and ALOS PAL-SAR DEMs have 1.1 and 0.81 km 2 areas, respectively, with negative errors of 4-7 m.These two DEMs have negative errors even for more than 7 m in about 0.42 km 2 areas.Though the errors in the ASTER DEM are lower than the SRTM DEM, the errors are highly abrupt for both the open and tree-canopy areas and do not follow any particular pattern which renders it extremely difficult for DEM correction.The FABDEM does not have errors more than 4 m.
The errors in vertical elevations for the Ranigram area indicate that most of the areas in the satellite-based DEMs are highly elevated compared to the UAV-based DTM, and the errors are higher in the tree-canopy area.The errors in the WorldDEM™ and FABDEM are the lowest among the DEMs analyzed, and the distribution of the errors across different error classes and land uses is relatively more plausible.Most errors in the FABDEM are within 1 m and some are within 2 m, indicating a better accuracy of this DEM.

| DISCUSSION
The HEC-RAS model results with the SRTM30 and ALOS PALSAR DEMs are very poor and have large errors.The SRTM DEM was made from the C-band radar images of the SRTM mission in 2000 and aimed for elevation accuracy of not more than 16 m (Rosen et al., 2001;Van Zyl, 2001).The errors found in this study are comparable to those of the open agricultural fields and forest cover areas of Greece, Saudi Arabia, and Norway (Elkhrachy, 2018;Hoffmann & Walter, 2006;Weydahl et al., 2007).The elevations are found to be mostly over-estimated in the tree-canopy areas and hence the errors are the highest, which echo the findings of Zandbergen (2008).The error in the ALOS PALSAR DEM is lower than the SRTM DEM as it is comparatively new, and both geometrically and radiometrically corrected using the SRTM data.Even with the corrections, ALOS PALSAR could not achieve the accuracy needed for local-level flood modeling.The SRTM mission data are quite old and because of the continued change in land-use pattern, the need for recent data with more accuracy is warranted.
The intensity of higher error in the ASTER DEM is lower than the SRTM30 DEM as it used stereo-pairs images acquired by nadir-viewing and backward-viewing in the near-infrared band from the TanDEM-X mission data.Gonz alez-Moradas and Viveen (2020) have also found the performance of the ASTER DEM good over the open areas and poor over the tree-canopy areas, and the error range is lower than that of SRTM.The anomalies and artifacts found in the floodplain of Bangladesh were also found in Estonia, Norway, New Zealand, and China (Uuemaa et al., 2020), which limit the usefulness of the ASTER DEM for local flood modeling.
A better accuracy of the WorldDEM™ in the open areas compared the other open-source satellite-based DEMs is reported in Becek et al. (2016).The accuracy is better in the WorldDEM™ because the TanDEM-X mission data were corrected with the thematic validation and according to the ISO 2859 standard (Becek et al., 2016).This DEM is the latest and has comparatively higher resolution.An error between 1 and 3.5 m for the WorldDEM™ is reported for Ukraine (Koppe et al., 2015), which is in the similar range of this study.The DEM was also used in the flood modeling in Pakistan (Farooq et al., 2019) and the results support the findings of this study.In spite of its better quality, the errors in the tree-canopy area are quite large for the local-level flood modeling.
The FABDEM is found to be more accurate all over the open, building and canopy areas of the Jamuna floodplain.The main reasons for the better accuracy are that the DEM was created from the TerraSAR-X data sources of the TanDEM-X mission by removing the forest and building to make it a terrain model (Hawker & Neal, 2021).A mean absolute vertical error of 1.12-1.61m was found for the FABDEM in built-up areas (Hawker et al., 2022), which conforms to the findings for the Ranigram area.The FABDEM was also found to be accurate in terrains having slope of <2 (Santillan, 2023), which is the terrain characteristic of the Ranigram area.Though this DEM does not have large errors (more than 2 m) in the Ranigram area, the low to moderate errors induce some uncertainties in the resulting flood parameters.
Most errors in the satellite-based DEMs are negative, which indicates that the DEM is higher than the actual land elevation.Because of this negative error, the flood models cannot detect the floods in higher areas, though these areas are flooded in real scenarios.Moreover, the small khals and canals through which water passes are not captured because of the low resolution.Also, the narrow embankments are not captured in the low resolution DEMs, which allow water to flow, but in reality, the flow is obstructed in those locations.As the land-use pattern is changing very frequently in the floodplain of Bangladesh, the outdated DEMs do not provide an accurate land elevation.Thus, regular updating of the global DTMs like the land use models are needed for a country like Bangladesh to capture dynamic aspects of floods with changing topography.Until then, the high volatility of the changing topography requires local (i.e., UAV-based) DTMs that can be updated frequently at minimum costs.Another potential use of the UAV technology could be for improving the satellite-based DTMs of a large area with the UAV-based DTMs of a smaller area of interest.
Erroneous satellite-based DEMs are not suitable for very local-level flood modeling where even half a meter flood detection is essential.The WorldDEM™ performs well in open areas, but has limitations in the tree-canopy areas.Therefore, this DEM can be used for local-level flood modeling only in the open areas.If the error in the tree-canopy area of the DEM can be corrected, it can then be used for local-level flood modeling.The FABDEM has the least error among the DEMs used, and the flood inundation throughout the study area is better captured with this DEM.Thus, this DEM is found to be relatively better for flood modeling of the Ranigram area.The usefulness of this DEM in hydrology and flood assessment studies has also been mentioned in a recent study (Hawker et al., 2022).Overall, the horizontal resolution of the satellite-based DEMs is very low, whereas highresolution DEM is necessary to capture the detailed flood information at the local level.
This level of detailed information representing realflood scenarios in the floodplain is crucial for local-level flood modeling.The UAV-based high-resolution DTM incorporated with field-measured elevation is fruitful for a flood model to represent the local-level flood scenario accurately.So, we find that a high-resolution DTM and frequent field-measured data can significantly improve the 2D hydrodynamic model performance.In this hybrid method, the measured topography and bathymetry for the DTM is only necessary if the land surface is not visible from the UAV due to tree-canopy and water.In riverside charlands of Bangladesh, only UAV technique with a few GCPs would be sufficient to represent the terrain accurately.The RMSE in low-vegetation areas for the UAV results is only about 5 cm (Gafurov, 2021).
The cost of a DJI Mavic Pro drone in Bangladesh is US$1150-$2000, which is cheaper than the comparable LiDAR technology.Moreover, the cost is decreasing which will make drones more readily available (Lee et al., 2022).The research and service organizations in Bangladesh and elsewhere can procure drones and enhance their technical capacity to perform drone survey and photogrammetry analysis.This can substantially reduce the cost of DTM generation and improve the local-level flood modeling.
The use of UAV technology is also constrained by some limitations.UAV cannot map the areas blocked by trees as the photogrammetry images cannot penetrate the trees to reach the terrain at the ground.Thus, it provides surface elevation only at the top of the trees (Westerveld, 2020).Since it cannot measure the terrain in the tree-canopy areas correctly, some vertical errors may occur in the tree-canopy areas.Hence, a land-based topographic survey is required in the canopy areas for generation of accurate terrain data.Moreover, the errors in an open-source DEM in a floodplain setting might not have a significant impact on flood parameters when a large domain is modeled.However, in a small domain for local-level study, these errors could create notable differin the results.The findings of this study are constrained by the use of a particular hydrodynamic model with particular types of boundary conditions.Also, measured data for two flood seasons only were used in developing the model.The study was conducted for a small area in floodplain environment.Moreover, since the timing of the satellitebased DEMs is not the same as the UAV-based DTM, the study in no way indicates the level of accuracy of the DEMs in their respective years of imaging.It only indicates the adequacy of those in local-level flood modeling in a dynamic setting for a recent year.

| CONCLUSIONS
A 2D HEC-RAS model for the floodplain of the Jamuna River was created utilizing a UAV-based high-resolution (17.5 cm) DTM.The hydrodynamic model with UAVbased DTM and surveyed field data performed very well in representing the local-level flood information.When the model was run with the satellite-based DEMs, the results were not accurate enough for the local-level flood modeling.The model results with the SRTM and ALOS PALSAR DEMs were found to be unsuitable for flood modeling at the local level.The WorldDEM™ and the ASTER DEM were better for the noncanopy areas, but the ASTER DEM had anomalies and abruptness in error distribution, which made it unsuitable for the local-level flood modeling.The errors in the vertical elevations of the low-resolution satellite-based DEMs made the hydrodynamic model represent the local-level flood information poorly.Overall, the WorldDEM™ with its 12 m resolution has less error and can be used for the locallevel flood modeling after suitable corrections for the tree-canopy and water areas.The FABDEM works better than the other satellite-based DEMs all over the study area.Though there are some errors in some areas, these can be accepted in the studies with a lack of funding for acquiring an accurate DTM.The 30 m resolution is still an issue that creates some errors in local-level flood model; otherwise, the FABDEM is the most preferable among the freely available DEMs.This study indicates that a high-resolution DTM prepared from UAV photogrammetry can be used in detailed hydrodynamic modeling at local level in floodplain environment.Bathymetric and topographic surveys are also essential for the tree-canopy and water areas, at the road and embankment locations with hydraulic structures underneath, as well as for ground truthing of the DTM.The findings of the study would be helpful in 2D hydrodynamic modeling, watershed delineation, inundation mapping with GIS, and hydrologic routing.Further studies are also needed in other catchments to check the general validity of the DEM-related conclusions drawn from this study.

F
I G U R E 2 Orthophoto of the Ranigram area in the Sirajganj district taken with a drone.

F
I G U R E 3 Methodological framework for creating a unmanned aerial vehicle-based high-resolution digital terrain model of Ranigram.
a. WorldDEM™ is a DTM developed based on the radar satellite data acquired during the TanDEM-X mission(2010)(2011)(2012)(2013)(2014)(2015) by postprocessing of the data to reduce the impacts of synthetic aperture radar (SAR)-specific F I G U R E 4 (a) Locations of the ground control points (GCPs) used and (b) the GCPs marked on the land surface with spray cans for the validation of the unmanned aerial vehicle photogrammetry.data features and artifacts where each editing quality control was checked by a thematic validation and is performed according to ISO 2859 (Airbus, 2018).The resolution of the DEM is 12 m.Becek et al. (2016) found the absolute vertical accuracy of the DEM to be less than 4 m, and the relative vertical accuracy <2-4 m. b.SRTM DEM was developed based on the C-band radar data from the images acquired by two interferometry SARs aboard Space Shuttle Endeavor in February 2000 (USGS, 2021).The DEM was released in 2003.It has a resolution of 30 m.The linear vertical absolute and relative height errors of this DEM were found to be <10 m (Jarvis et al., 2008) and 16 m (Farr et al., 2007), respectively.c.ALOS PALSAR was a satellite mission by the Japan Aerospace Exploration Agency during 2006-2011.It is a geometrically and radiometrically terrain corrected data product using the SRTM DEM by the Alaska Satellite Facility (ASF, 2021).The resolution of the DEM is 12 m.The vertical elevation of the DEM is in ellipsoidal height, which needed to be converted into orthometric height by a geoid undulations model (Takaku et al., 2014).An average geoid height of À55.32 m for Ranigram was used in the conversion.d.ASTER DEM was created photogrammetrically from a compilation of cloud-free ASTER stereo-pair images from the TanDEM-X mission (DeWitt et al., 2015; JPL, 2021).The sensor was carried aboard the Terra satellite launched in December, 1999, and the stereo-pair T A B L E 1 Root mean square error of the unmanned aerial vehicle-based digital terrain model of Ranigram based on the 46 ground control points data.U R E 5 (a) Land cover classified into tree-canopy and water area, and noncanopy area, (b) Measured point elevations in the treecanopy and water areas, (c) digital terrain model (DTM) of the noncanopy area with the unmanned aerial vehicle-based data, (d) DTM of the tree-canopy and water areas with the measured elevation data, and (e) Combined final high-resolution DTM of Ranigram.images were acquired by nadir-viewing and backwardviewing, and after-looking angles in the near-infrared band (Gonz alez-Moradas & Viveen, 2020; Hirano et al., 2003).The DEM was released in 2011 and reproduced with an updated algorithm in 2020.The resolution of the DEM is 30 m.The elevation error of the DEM is about 7-15 m (Hirano et al., 2003).e. FABDEM is the forest and buildings removed DEM from the Copernicus GLO-30 DSM.This DEM was originally from the TerraSAR-X data sources of the TanDEM-X mission during 2010-2015 (Hawker & Neal, 2021).The data were refined with different algorithms as well as ground-based measurements.This DEM was released by the University of Bristol.Because of the removal of the forest and buildings, this DEM can represent the DTM.The resolution of the DEM is 30 m.

F
I G U R E 7 Simulated flood maps for (a) June 28 (b) July 8 (c) July 15, and (d) August 29, 2020 compared with real-time flood photos of Ranigram.used with the globally available open-source lowerresolution DEMs.The comparative results of important flood parameters the 2020 flood are given in Figure 8.The arrival time maps of the model built with the SRTM and ALOS PALSAR DEMs indicate that the floodwater arrived only at the agricultural lands near the breaching location and could not reach the other lower lands in the north and south of Ranigram.The results with the WorldDEM™ and ASTER DEM show that the floodwater arrival in the agricultural and lower lands followed the result of the UAV-based DTM, but the floodwater did not arrive in the elevated housing lands.The FABDEM result shows an early arrival in the agricultural and lower lands, and 3-4 weeks to arrive in the housing lands except for a few south-eastern lands where the result followed the UAV-based DTM.The arrival time in some northern areas and south of the UP areas shows a very quick arrival compared to the UAV-based DTM.The duration maps with the WorldDEM™ and ASTER DEM show that the flood duration in the lower agricultural lands was the same as the UAV-based DTM, but no flood was shown on the elevated housing lands.The duration maps with the SRTM and ALOS PALSAR DEMs show that the flood water stayed only on the agricultural lands near the breaching point and could capture two-thirds of the UAV-based DTM result.The duration map with FABDEM shows similar results of the UAV-based DTM except for some northern areas and south of the UP areas.The maps of the percentages of time inundated with different DEMs show similar results as for the flood duration.

F
I G U R E 9 Errors in vertical elevations in different satellite-based DEMs for the Ranigram area of the Sirajganj district: (a) WorldDEM™, (b) ALOS PALSAR, (c) SRTM 30 m, (d) ASTER, and (e) FABDEM.

F
I G U R E 1 0 Area-wise vertical errors in different satellitebased digital elevation models for the Ranigram area in the Sirajganj district.