Providing solutions for data scarcity in urban flood modeling through sensitivity analysis and DEM modifications

Urban flooding, caused by excessive rainfall, has become one of the most common natural disasters in recent years. Developing an accurate urban flood model becomes imperative as flooding affects livelihoods and economies of the population in affected areas (Teng et al. 2017). Simulating urban flooding is challenging because of the complex interaction between hydro-meteorological, surface, and sub-surface factors, which are the major determinants of flood flow. However, the availability of sufficient and reliable observation data is an ongoing challenge, especially in developing countries. Due to these limitations, in addition to their limited technical capabilities, there has been comparatively little attention given to flood analysis and forecasting research in these countries, which is mainly attributed to their restricted financial resources (Nkwunonwo et al. 2020). Inconsistencies in surface data, particularly in land use and soil type, can significantly impact the surface overflow and infiltration capacity, affecting the accuracy of the flood simulation. Due to the financial constraints and substantial effort needed to conduct a detailed land characteristics survey, such efforts are often not prioritized in developing countries. As a result, much of the surface data is roughly estimated or outdated. This ongoing challenge has significantly contributed to the limitations in the development of accurate flood models in data-scarce domains.

However, recent innovative studies have developed methods to supplement input data, even in case of deficiencies in data gathered from in situ instruments. These research efforts involve the utilization of remote-sensing products to represent the ground data (e.g., Khaki et al. 2020; Xu et al. 2023), the application of artificial neural networks (e.g., Kim et al. 2019), and the combined usage of numerical simulations and machine learning methods (e.g., Yan et al. 2021). Some studies have also introduced calibration methods to optimize the usage of existing data in improving flood prediction. Reynolds Puga (2019) developed a method to improve the robustness of flood model calibrations using limited hydro-meteorological data. Reynolds et al. (2020) explored the utilization of limited calibration data and uncertainty forcing data in flood simulations, while Giuliani et al. (2017) introduced a synthetic data approach to generate large datasets based on limited observations.

In addition to the availability of adequate hydro-meteorological data, limitations in obtaining high-resolution digital elevation models (DEMs) pose a significant challenge in geospatial analysis. Recent studies have emphasized the importance of high-resolution DEMs for the accurate prediction of urban flooding (de Almeida et al. 2018; Hawker et al. 2018; Kim et al. 2019; Xing et al. 2021). For developing countries, producing a wider-scale fine-resolution DEMs (<5 m) is financially and practically challenging. In contrast to flood simulation for rural areas, where the surface overland flow can be determined by 2D overland flow based on shallow water equations (Hsu et al. 2000), simulating flood flow in urban domains imposes more complex dynamics because flow is altered by anthropogenic structures (Falconer et al. 2009; Paquier et al. 2015). The usage of digital terrain models (DTMs), which are synthetic representations of the bare earth ground surface, is widely used in flood modeling due to its plain topography and simple calculations according to McClean et al. (2020). However, for urban flooding, flood flow is primarily a combined effect of the distribution of urban features, such as roads and buildings. Fine-resolution digital surface models (DSMs) are preferable for simulating urban flooding as they can adequately capture these attributes (i.e., Priestnall et al. 2000; Mason et al. 2007; Meesuk et al. 2015; Wang et al. 2018; Yalcin 2018). However, the lack of high-resolution elevation data in developing countries presents difficulties in performing accurate flood simulations.

In recent years, a number of modeling simplifications have been attempted to overcome the limitations, particularly in conceptualizing the surface component of flooding where small catchments are interconnected through road networks. Mark et al. (2004) proposed a methodology that incorporates the streets into the DEM with the elevation corresponding to the curb height. Lee et al. (2016) adjusted the elevation by increasing the height of the building configuration and decreasing the height of the road and the curve. Schubert & Sanders (2012) compared methods for modeling unsteady flows through urban areas using building treatment. Arrighi & Campo (2019) investigated the influence of DTM resolution on the estimation of urban flood losses by representing comparison results involving raw DTM data, filled building footprints, waterproof block representations of buildings, and merged elevation datasets. Lee (2021) suggested a method for modifying the DEM using the rolling ball concept, which involves adding realized urban vertical data to existing DEM data. Xing et al. (2021) developed a correction method based on the geographic information system (GIS) to modify DEMs by adding building complexes and removing bridges to obtain a more realistic description of flood flow. The developed techniques provide adequate results; however, these approaches are applied to urban domains that are not larger than 20 km². These methods are not ideal for larger sub-catchments because of the considerable effort required to create individual building configurations for a domain of this size.

Therefore, to mitigate the limitations in the availability of high-resolution input data including elevation, rainfall, surface roughness, and infiltration parameters, alternative solutions have been proposed to improve the accuracy of the results. A more accurate representation of elevation is achieved in the larger-scale urban flood simulation by utilizing the information from both DTMs and DSMs. These datasets are utilized to represent the stream, road, and building barriers, and they have been processed in a GIS environment. Adequate hydro-meteorological parameters are derived for model input from limited observation data through remote-sensing alternatives, and the developed method has been used to adjust these parameters effectively through sensitivity analysis. The aims of this research are to (1) provide a method for modifying the mid-quality DEM data adequately for simulating urban floods to incorporate road elevation as routes for surface flood propagation in urban domains, (2) identify the adjustment parameters for the watershed roughness coefficient, infiltration, and rainfall distribution, (3) perform validations for the simulation by applying the method to an urban flood event, (4) analyze the effect of incorporating urban terrain properties in the accuracy of the flood simulations.

Sensitivity analysis was performed for the urban flood model input parameters due to the study constraints encountered in our previous work (Dasallas et al. 2022), where a significant difference was seen in the seen and simulated maximum peak of the upstream water level for the rainfall event. This study intends to present a solution for the scarcity of appropriate data necessary for simulating urban flooding, provides an adjustment method that can be applied to a wider-scale urban domain, and improves the datasets fit for more accurate urban simulations. Improving the input datasets for numerical simulations can be a valuable tool for the accurate assessment of people's safety during flood events (e.g., Arrighi et al. 2015; Lumbroso & Davison 2018; Bernardini et al. 2021) and evaluating the impact of flood flow in structure stability (Postacchini et al. 2019; Martinez-Gomariz et al. 2020). These limitations in the availability of detailed data are critical challenges for developing countries, some of which are prone to extreme flood events but have less capacity to produce high-resolution, hydro-meteorological, and elevation data critical for urban flood management, prevention, and mitigation.

Due to the lack of comprehensive data for observed rainfall, land use and soil type data within the domain, supplementary data, and correction processes were applied to the model input data. Remote-sensing-based precipitation estimates are employed to supplement the insufficient ground-based rainfall data. Parameter calibration is conducted based on the range of values for surface roughness and infiltration calculated from the works of Arcement & Schneider (1989) and Rawl et al. (1983), respectively. Finally, a GIS method is developed to improve the elevation accuracy of the available DEM data. While remote-sensing data have been increasingly utilized as an alternative to observational data in modern research, there is limited literature exploring the application of parameter ranges derived from existing experimental studies. Moreover, although small-scale building and road network adjustments have been applied in previous studies, difficulties arise when applying these methods on a larger scale. The methodology proposed in this study incorporates innovative solutions from recent research but also introduces a new approach to address the data scarcity for flood simulation in developing countries. The corrected input parameters are then used to simulate the rainfall-runoff model for the watershed-scale runoff simulation. The urban domain elevation was adjusted to incorporate the characteristics of road network and building configuration from the DTM and DSM, respectively. The modified DEM is then utilized as base topography for the urban flood simulation.

The use of satellite-retrieved precipitation estimates was applied due to insufficiencies in the ground-based data. The application of the retrieved estimates is only meant to supplement the estimated spatial distribution of rainfall within the watershed; the retrieved rainfall data were not employed for rainfall satellite data calibration. Four sets of hourly rainfall estimates were used to represent each of the synoptic stations within the watershed, namely Mt. Oro, Manila, NAIA, and Tanay stations. The rainfall time series were then used to simulate the inundation.

The Manning correction coefficient is exclusive for the watershed, as the values are expected to differ from different spatial characterizations of land use within watersheds.

The standard range values of the soil moisture content and wetting front from Rawl et al. (1983) are utilized to calibrate the simulated water level.

In flood modeling, DEM modifications are done to adjust or alter the elevation data to better represent the topographic features of the flood plain. Due to the limitations in the availability of high-resolution and accurate DEM data in developing countries, modifying DEM can be a valuable tool in improving the accuracy of the elevation model. The elevation models used in this research were acquired from the University of the Philippines-Disaster Risk and Exposure Assessment for Mitigation (UP DREAM) Program, which utilized the light detection and ranging (LiDAR) equipment, specifically the Airborne Laser Terrain Mapper (ALTM) (UP DREAM 2022). These datasets have horizontal and vertical accuracies of 50 cm (Training Center for Applied Geodesy and Photometry 2015) and underwent validation, pre-processing, calibration, orthorectification, and editing before their final use (UP DREAM 2022). However, despite these steps, DSMs can still be subjected to noise artifacts that may lead to misleading inundation results (Yunika et al. 2019). Noise refers to the presence of miscorrelated values distributed in the DEM data that can reduce its accuracy. To address this issue, a combined elevation of the DTM and DSM was used to simulate urban inundation, incorporating small streams and road networks for flood propagation.

First, the following data were obtained: 5 m resolution DTM and DSM data (ASCII file) from the UP NOAH Center, and road and stream data (shape file) from the Philippine GIS portal (https://www.geoportal.gov.ph/). The following steps were performed using the GIS software:

• (1)

  Create a polyline shapefile for roads and streams based on remote-sensing data,

• (2)

  Convert road and streamlines to polygons using the Buffer Algorithm (Note: road distance is set as 5 m, average width of small roads),

• (3)

  Trim the polygon file, removing unnecessary parts of the road polygon, such as small streets, out-of-boundary roads, and isolated roads,

• (4)

  Convert road and stream polygons to raster using Clip Raster with Polygon Tool (the elevation of streams and roads in the urban domain was assumed to be identical to the measured DTM data), and

• (5)

  Overlay the DTM road and stream raster onto the DSM raster using the Mosaic Raster Tool.

Sensitivity analysis is performed in flood modeling to determine how the different variations in the hydro-meteorological input variables can affect the resulting flood outputs given a set of different parameter value assumptions. The process is applied to identify the most appropriate values to use in order to improve the accuracy and reliability of the flood simulations. Due to the lack of high-accuracy hydro-meteorological parameters for the data-scarce domain, sensitivity analysis is performed to determine the exact values of infiltration and roughness coefficient parameters that can simulate the water level closely replicating the water level observed at the stream gauge station. To determine the most suitable input parameters to use for the water level calibration, several test case simulations were conducted to replicate the change in the water level recorded at the Sto. Niño station (refer to Figure 2 for the location of the station). The MDEM was used as the base topography for these simulations (Figure 6(c)).

For the rainfall input, the simulations were conducted under three sets of cases. First, recorded ground-based data were utilized for Cases 1–3. Second, the retrieved data (i.e., Case 4) from satellite estimates were used. And third, a combination of ground-based and retrieved data was employed for Cases 5–8. For example, the original, unaltered data for roughness coefficient, infiltration, and ground-based rainfall were used to simulate Case 1, minimum parameters and ground-based rainfall were used for the Case 2 simulation, and so on. For the ‘ground + retrieved’ column in Table 3, simulations were tested using two different approaches. In one approach, a single retrieved data averaged for the entire watershed was used for the simulation (watershed-based). In the other approach, retrieved data estimated for each rainfall station was used separately for the simulation (station-based).

Based on the comparison of water level deviation for the different test cases, it was observed that Test Case 6 (blue line) was able to replicate the behavior of the increase in the observed water level. In this case, the roughness coefficient and infiltration parameters were set based on their maximum values. In addition, the distribution of rainfall within the watershed was identified as a crucial factor in accurately simulating the water level, as evident from the comparison of Test Cases 1–3 and 4–8. Due to the limitations of the availability of reliable rainfall data recorded during the event, rainfall estimates had to be assumed for the other stations within the watershed through the use of global monitoring data. Rainfall estimates, together with the observed ground data, appeared to be more dependable flood input data and were therefore used to simulate the urban flood event.

Comparisons of the observed and simulated flood depth are shown for the upper stream (Figure 13), midstream (Figure 14), and downstream regions (Figure 15), which show the underestimated and overestimated points for the three scenarios. Based on the observed flood depth during the event, more than 2 m of flooding were observed along the low-lying areas surrounding the river. Significant levels of flood were also experienced for areas not necessarily near the river but with low-lying elevation (red areas in Figure 8). While the flooding surrounding the river was captured in the DTM simulations, DSM simulations, on the other hand, were able to encapsulate the flooding along small pseudo-catchments resulting from impedance from the urban natural and artificial barriers. Both of these observations are captured for MDEM simulations.

In Figure 16, the comparison of observed and simulated flood depth for individual flood points shows that the majority of points have underestimated (the simulated depths from the DTM (blue) are shown). DSM simulations (red) show some improvements; however, they still result in some over- and underestimations at several points. In MDEM simulations (green), however, most of the simulated flood points fall within the observed flood depth range.

The figure shows the fit correlation between the observed and simulated flood depth using DTM, DSM, and MDEM, and it also includes a linear regression plot for each. Based on Table 4 and Figure 17, DTM simulations perform relatively well for flood depths of 0–0.5 m, but not as well for flood depths exceeding 0.5 m. However, DSM performed DTM when flood depths exceeded 0.5 m. In the figure, the DSM exhibits a more closely clustered set of points along the regression line compared to the DTM, indicating a stronger correlation. However, MDEM simulation was able to replicate the observed flooding predominantly at all levels, except for the 0–0.5 m range where DTM performed relatively better. MDEM achieved percentage hits of 100, 60, 86.67, 96, and 86.21%, with corresponding RMSE values of 0.00, 0.15, 0.14, 0.17, and 0.63, respectively. This result aligns with the scatter plot correlation graph, where the regression line for MDEM shows a significantly stronger and more positive linear correlation, compared to the untreated DEM simulations. Overall, the range of RMSE improvements across all flood depth level varied from and for untreated elevation data to for the modified elevation data.

Based on these observations, the underestimated flood depths in Scenario 1 are assumed to be a result of the absence of impeding artificial barriers in the simulation, which causes the accumulation of floodwater toward the river. These were resolved in Scenario 2 simulations. The overestimations in Scenario 2 are presumed to be caused by limitations in determining distinguishable flood outlets, such as clearly defined roads, canals, and streams, which would serve as flood-flow paths from the confined areas toward the river. However, these limitations were adequately resolved in Scenario 3. This implies that, for moderate-quality elevation data, MDEM simulations were able to accurately represent flood depth and the extent of the urban flooding, in comparison to the untreated terrain and surface elevation models. The new procedure would enable large-scale pre-processing of ‘noisy’ urban DEM data, making it suitable for accurately representing inundation in urban flood simulations.

Due to the limitations in the availability of adequate data necessary for urban flood simulations, including rainfall, watershed roughness coefficient, infiltration, and high-quality digital elevation data, parameter calibrations are carried through sensitivity analysis. In addition, a method of modifying the DEM based on the existing 5 m resolution DTM and DSM, using GIS tools, is proposed. Numerical simulations were carried out to compare the resulting inundation using DTM, DSM, and MDEM as base topographies, which are applied in an extreme flood event in the Marikina–Pasig River basin, Philippines. Flood calculations were performed by applying 2D shallow water equations in the surface overflow, utilizing the implemented finite volume method (FVM) for numerical discretization.

The method for modifying the DEM through the GIS is summarized in the following steps: (1) creation of road and stream polylines based on remote-sensing data; (2) conversion of polylines into polygon shapefiles; (3) trimming of road and stream files; (4) conversion of the shapefile into raster format; and (5) overlay of DEM road and stream raster to DSM raster using the Mosaic Raster Tool. The resulting elevation raster, which possesses both the flood propagation and impeding characteristics of the DTM and DSM, is used as the base topography for flood simulation. Adjustments were made to calculate the watershed surface roughness coefficient, infiltration parameters, and rainfall distribution input. The roughness estimation utilized the Manning equation proposed by Arcement & Schneider (1989), which accounts for different roughness values depending on various factors such as base characteristics, surface irregularity, channel cross-section variation, obstruction, vegetation and flow conditions, and channel meandering correction. The infiltration adjustment follows the variation of change in soil moisture content and wetting front suction head, following the approach presented by Rawl et al. (1983) to solve the Green and Ampt equations. Input rainfall data were adjusted by incorporating observed rainfall data from a ground station and a rainfall estimate retrieved from the IMERG for the specific watershed. A sensitivity analysis was conducted for the simulated water level, revealing that the accurate reproduction of observed water level was achieved by utilizing maximum values for the Manning coefficient and infiltration parameters, along with the combined ground- and station-based rainfall estimate data.

For the 2009 case study, a total of 78 observed flood points indicated that the simulations utilizing the MDEM as base topography resulted in a more accurate prediction of inundation, with error values of ⁠, compared to DTM and DSM simulations with error value ranges of and ⁠, respectively. A correlation scatter graph also shows that MDEM simulations produced a stronger, positive linear regression as compared to the untreated DEM simulations. MDEM simulations successfully represented road flooding surrounding the river, streams, and low-lying regions, as well as the accumulation of floodwater in the urban catchments between the building structures. The modified DEM was capable of providing a more accurate representation of flood flow within road and stream networks, a crucial factor in identifying flash-flood and debris flow potentials.

While the presented approach offers valuable insights, it is important to acknowledge certain limitations in the applicability of this method to urban flood modeling. In particular, the lack of detailed information regarding the watershed's soil type and land use data utilized in this research. Although these data can still be used to calibrate the watershed surface roughness coefficient and infiltration within a large-scale hydrologic process, more detailed impervious surface data are required to conduct a much more comprehensive analysis of flood flow in the urban domain. The unavailability of high-resolution surface soil characteristics also hampers the exploration of the retention capability of green spaces and porous structures. In the developed technique, open spaces are treated as represented in the DSM data, without considering their potential role as flood propagation pathways or catchment for retention, which depends on the height of the surface relative to the flood depth. Furthermore, the approach does not account for the interaction between surface and sub-surface sewer pipes. Therefore, it is recommended to apply this approach for simulations of large-scale urban inundation, rather than detailed hyper-resolution (<5 m) urban flood analysis, considering the aforesaid limitations.

Despite these limitations, the utilization of MDEM in urban flood modeling offers significant advantages in simulating inundation more accurately, allowing the inclusion of road, stream networks, and building barriers in the simulations. The proposed technique can be instrumental in simulating urban flood events, providing critical information for flood early warning systems, mitigation strategies, traffic advisories, and long-term flood management actions. This methodology serves as a solution to address the challenges posed by the limitations in the availability of adequate resolution elevation and hydro-meteorological data, which is a crucial issue for under-developed and developing countries vulnerable to recurrent intense rainfall events that often lead to hazardous urban flooding.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.