ABSTRACT
Rapid urbanization necessitates reliable assessments of its impact on flooding. The two major challenges in data-sparse basins are the equifinality of the hydrological parameters and their dynamic updating with respect to changing land use and land cover. Existing calibration approaches typically address only one of these challenges. The existing approaches that allow for dynamic updating of parameters with changing land use and account for parametric uncertainty are referred to as Dyn_CN and uStat_CN, respectively. We propose a novel ‘urbanization-informed uncertainty-based calibration approach’, referred to as uDyn_CN, that combines the advantages of both these approaches. The existing and proposed calibration approaches were applied to India's data-limited, frequently flooded, urbanizing Adyar river basin comprising Chennai city. The performance assessment of the methodologies was done in an externally coupled hydrological-hydraulic modelling framework. In that light, the proposed hydrological calibration approach is more accurate than the existing approaches in both the hydrological and hydraulic simulations. For instance, the RMSE values for the simulated maximum water depths for uDyn_CN, uStat_CN, and Dyn_CN approaches are 0.22, 0.36, and 0.6 m, respectively. The study also demonstrates its subsequent application to analyze the impact of urbanization on flood hazard in the Adyar basin by 2050.
HIGHLIGHTS
The study proposed a novel urbanization-informed calibration approach that captures the uncertainty in data-limited regions.
The impact of urbanization on flooding can be modelled, considering uncertainty using the proposed approach.
The proposed approach performs better than the existing approaches in coupled hydrological-hydraulic flood modelling.
INTRODUCTION
Globally, floods are the most frequent disaster, causing havoc to life and property (IPCC 2021). The rapid growth in population in and around the urban areas (Schafer & Victor 2000) has resulted in the accelerated development of impervious surfaces and reduced rainwater storage in the soil, which in turn leads to flooding even for short and intense rainfall events. Thus, urbanization exacerbates the frequency and intensity of flooding, increasing the vulnerability of people inhabiting the floodplains (Halgamuge & Nirmalathas 2017), especially in developing nations. Hence, it is essential to accurately model the urban floods and ascertain the areas at flood risk. However, the majority of the river basins in the world that are ungauged or data-limited due to the lack of a network of gauging stations or the deterioration of the existing gauges pose a challenge to modelling the floods occurring in them (Sivapalan et al. 2010; Efstratiadis et al. 2014; Pool et al. 2017).
As flood dynamics primarily depend on the topography of the domain, geomorphological indices are often used to delineate flood-prone areas in large ungauged basins (Degiorgis et al. 2013; Manfreda et al. 2014; Nobre et al. 2016; Albertini et al. 2022; Annis et al. 2022; D'Angelo et al. 2022). The upslope contributing area, elevation, slope difference to the nearby channel, etc., are used in the process (Manfreda et al. 2014). These methods are generally used in identifying frequently flooded areas over large catchments (Hdeib et al. 2018). The applicability of such methods to identify flooding in relatively smaller, highly urbanized flood plains remains limited. This can be attributed to the multiple flow paths of floods caused by the smaller topographic features such as buildings, streets, major drains, and canals. Therefore, precise mapping of flood extents, water levels, flow rates, and the associated flood risk for various management scenarios can be achieved only with a hydrological-hydraulic flood simulation framework (Theiling & Burant 2013; Hdeib et al. 2018). In this process, the runoff hydrographs of a hydrological model can then be used as boundary conditions to a hydraulic model for simulating flood inundation and its dynamic evolution. Thus, in the urbanized flood plains, the utility of hydrological and hydraulic modelling frameworks is essential, as they help in a deeper understanding of the behaviour of the basins for floods of different intensities.
The simulation of flood inundation extents by linking the hydrological and hydraulic models is the first and foremost step in managing floods (Merwade et al. 2008; Theiling & Burant 2013). However, the parameters of the hydrological models need to be calibrated based on past flood data for accurate simulation of the floods (Efstratiadis et al. 2014; Manfreda et al. 2014). In that light, reliable information on spatio-temporal variation in rainfall and flood flows in the rivers, such as the discharge time series capturing peak flows and time to the peak, is essential for calibrating hydrological models (Koutroulis & Tsanis 2010) and subsequently generating accurate boundary conditions for hydraulic models in order to obtain flood inundation maps. For the calibration of hydraulic models, water-level time series and a post-flood survey of the floodwater marks are required (Gaume & Borga 2008). This information is generated by intensively recording the wrack, debris, or watermarks in trees, walls, pillars of bridges, etc., post-floods (Horritt & Bates 2002; Borga et al. 2008; Neal et al. 2009; Nithila Devi et al. 2019). Flooding extents can also be derived from satellite images (Bates & De Roo 2000; Bates 2004; Komi et al. 2017), though the peak timing of the floods may not be captured.
Particularly, with respect to data requirements for hydrological model calibration, the regions termed ‘data-poor’ lack the availability of spatially and temporally varying rainfall and streamflow data (Hdeib et al. 2018). The flow or rainfall data may be discontinuous and/or available at a coarse time step. The rainfall data may be point measurements instead of spatially varying data from a radar. It needs to be highlighted that recording peak flows during extreme floods is a big challenge (Fukami et al. 2008; Horritt et al. 2010; Koutroulis & Tsanis 2010). The inability to capture peak flows of extreme events due to recording at the coarse time step, failure of the measurement device, inaccessibility, etc., may result in poorly defined rating curves. Calibration in such data-limited areas may result in parametric equifinality or several parametric values providing agreeable simulations of past flood events (Aronica et al. 2002; Pappenberger et al. 2008). Comprehending such inherent uncertainty linked to the parameters in modelling is essential (Aronica et al. 2002; Pappenberger et al. 2008; Apel et al. 2010; Domeneghetti et al. 2013).
These limitations challenge modelling floods in data-sparse regions that are highly susceptible to flood damage. In the earlier studies (Eckhardt et al. 2003; Breuer et al. 2006), the uncertainty in the LULC-dependent hydrological parameters was investigated in analyzing the implications of land use change in hydrological variables such as stream flow and groundwater. These studies did not perform any calibration to identify the behavioural ranges of parameters that give reliable simulations. Instead, they used narrowed-down ranges based on literature, and it has to be mentioned that these parameters are not physically measurable. Therefore, it cannot be said that they represent realistic simulations. Consequently, an uncertainty-based calibration may aid in obtaining reliable simulations, and eventually, better flood management decisions can be made (Apel et al. 2010; Domeneghetti et al. 2013; Annis et al. 2020).
Typically, for modelling floods in the data-sparse regions, the semi-distributed hydrological model with the hydrologic unit or sub-basin-specific parameter values is calibrated to simulate past flood events (Koutroulis & Tsanis 2010; Garambois et al. 2013; Massari et al. 2014; Mishra et al. 2018; Nithila Devi et al. 2019; Daksiya et al. 2021). Such studies that have looked into addressing hydrological parametric uncertainty in data-sparse regions are discussed in the following text. The parameters of the hydrological model, for example, HEC-HMS for the data-sparse basin of Lebanon, were calibrated using the available past storm data (Hdeib et al. 2018). Monte-Carlo simulations were used to find the behavioural parameter ranges for each sub-basin, thereby addressing the problem of equifinality in data-limited regions. A single land use land cover (LULC) map was used, and the sub-basin-wise fixed parameter ranges were identified. According to the study, 6–12 events were considered for calibration in the data-limited study area, and stable parameter ranges were determined. On the other hand, some studies (Nithila Devi et al. 2019, 2020) used literature-based ranges of hydrological parameters that are expressed as a function of LULC for simulating the impact of LULC change in flood modelling in a data-limited basin in India. These studies did not consider uncertainty in the hydrological parameters but permitted the updating of hydrological parameters with the changing LULC.
It can be thus concluded that with regards to flood simulation, two conventional hydrological model calibration approaches are practised in the data-sparse regions: (i) using several historical events to arrive at temporally constant parametric ranges (Koutroulis & Tsanis 2010; Hdeib et al. 2018) and (ii) using one or two historical events to estimate one single deterministic value for each parameter (Daksiya et al. 2017; Mishra et al. 2018; Nithila Devi et al. 2019, 2020). The first calibration approach addresses uncertainty in data-limited regions by estimating behavioural parameter ranges for each sub-basin of the hydrological models. The lacuna with the sub-basin-specific calibration approach is that the calibrated values remain fixed for the sub-basin and will not be calculated as a function of LULC. Therefore, though this approach accounts for uncertainty, it does not allow for the dynamic updating of hydrological parameters with the changing land use, for example, in urbanization. Hence, uncertainty can be addressed using the first calibration approach, but the impact of urbanization or changes in LULC on flooding cannot be quantified. The second approach allows for dynamic updating of the parameters with the changing LULC, thus facilitating the analysis of the impact of urbanization on flooding. However, it does not account for the uncertainty typical of the data-sparse regions. For this reason, a calibration approach that accounts for both the parametric uncertainty and the changing land use needs to be explored in rapidly urbanizing catchments that are also data-sparse. This may eventually help in reliable modelling of the implications of future urbanization on flooding.
For this reason, we propose an uncertainty-based calibration framework that estimates the schema or the look-up table linking land use and the narrowed-down parameter values instead of temporally fixed sub-basin-wise averages. A Monte-Carlo sampling of the schema and past flood data is involved in this uncertainty-based calibration. Thus, when an urban sprawl projection for the future is simulated, sub-basin-wise counts of different land uses can be made. Based on the calibrated schema, hydrological parameters can be estimated for each sub-basin depending on its land use area histogram, and reliable flood simulations for future urban scenarios can be performed for data-limited regions. Therefore, the novel calibration schema can be termed as an ‘urbanization-informed uncertainty-based calibration approach’.
To illustrate the utility of the proposed uncertainty-based calibration framework in modelling the impact of urbanization on flooding, the widely used HEC-HMS is chosen as the hydrological model. A suite of models under the name IROMS (Integrated River and Ocean Modelling System) has been developed at the Hydraulics Laboratory of the Indian Institute of Technology Madras. The two-dimensional (2D) local-inertial model in IROMS (Sridharan et al. 2020; Nithila Devi and Kuiry 2024), which is a modified version of the LISFLOOD model (Bates et al. 2010; De Almeida & Bates 2013), is used herein for hydraulic flood simulations. The data-limited and rapidly urbanizing Adyar River basin in Tamil Nadu of India, which comprises the south of the Chennai Metropolitan Area (CMA), is chosen as the study area. The CMA is densely populated, and the population is expected to reach 10 million by 2025 (Sekar & Kanchanamala 2011). The urban areas in the CMA have increased several times in recent decades (Bharath et al. 2017; Padmanaban et al. 2017; Nithila Devi et al. 2019), and urbanization is still happening. The vulnerable floodplains of the Adyar River in CMA are exposed to recurrent flooding during the North-East monsoon seasons. The basin remains ungauged, and the inflows into the Chembarambakkam reservoir located on one limb remain the only source of flow data for model calibration. The proposed calibration framework has used the inflows of past flood events and post-flood watermarks survey data. A cellular automata-based algorithm is used to predict future land use by 2050. Then, the calibrated hydrological and hydraulic models used the projected land use to analyze the impact of urban sprawl on flooding scenarios of the CMA.
This study comprises the following two objectives:
uncertainty-based calibration framework for estimation of the schema linking LULC categories and their corresponding narrowed-down hydrological parameter ranges, and
usage of the calibrated schema in analyzing the impact of urbanization on flooding in the data-limited study area.
For comparison, not only the proposed novel urbanization-informed calibration approach but also the aforementioned existing approaches have been utilized in historical flood simulations. This is done to bring out the advantages of the proposed calibration scheme in modelling the impact of urbanization on flooding. Thus, the paper's importance lies in its attempt to address equifinality in data-limited areas and facilitate dynamic updating of the hydrological parameters with the changing land use to reliably simulate the impact of urbanization on flooding.
STUDY AREA
Study area map of the Adyar basin with the hydrological and hydraulic domains. The sub-basin names, prefixed with ‘W’, generated by the hydrological model are also shown in the figure.
Study area map of the Adyar basin with the hydrological and hydraulic domains. The sub-basin names, prefixed with ‘W’, generated by the hydrological model are also shown in the figure.
METHODOLOGY
Overall methodology describing the steps involved in module 1: uncertainty-based calibration (left side) and module 2: impact analysis of urbanization on flooding (right side).
Overall methodology describing the steps involved in module 1: uncertainty-based calibration (left side) and module 2: impact analysis of urbanization on flooding (right side).
Module 1: uncertainty-based calibration of hydrological parameters
Data availability
The flooding events of the years 2005, 2009, 2010, 2015, 2016, and 2020 are chosen for calibration. For all these events, daily rainfall data at the gauge stations (Table 1) were obtained from the Public Works Department (PWD). The locations of these gauges and the Chembarambakkam reservoir are shown in Figure 1. The inflows into the Chembarambakkam reservoir are the only available data during these events, as no other flow recordings are available for the Adyar basin. The inflow time series obtained from the Chennai Metropolitan Water Supply and Sewerage Board (CMWSSB) is calculated by budgeting the storage, release, and previous inflows at daily time steps. Nevertheless, the catchment for Chembarambakkam reservoir constitutes nearly 47.2% of the entire catchment area upstream of CMA. Also, in the Adyar basin, both the reservoir's catchment and non-catchment areas upstream of CMA are composed of rural areas. The hourly streamflow time series estimated using a rating curve is available for the flooding event of the year 2008 at the outlet of the basin (Josephine et al. 2014; Anandharuban et al. 2019; Duraisekaran et al. 2021). These are the data available for hydrological calibration and validation in the data-limited Adyar basin.
Past storm events and the data availability of the Adyar basin
Flood event . | Rain gauge data availability . | Total rain (highest one-day rain) (mm) . | Streamflow data availability . | Usage . | LULC . |
---|---|---|---|---|---|
02Dec2005–05Dec2005 | CH, ME, PA, SR, TAa | 360 (283) | reservoir inflows | CALIBRATION | 28Jun2002 |
25Nov2008–03Dec2008 | CH, ME, PA, SR, TA | 460 (188) | basin outlet | VALIDATION | 09Mar2011 |
02Nov2009–13Nov2009 | CH, ME, KA, AUa | 686 (151) | reservoir inflows | CALIBRATION | 09Mar2011 |
01Nov2010–09Dec2010 | CH, ME, KA, AU | 241 (75) | reservoir inflows | CALIBRATION | 09Mar2011 |
26Nov2015–04Dec2015 | CH, ME, KA, AU | 511 (383) | reservoir inflows | CALIBRATION | 25Mar2017 |
10Dec2016–18Dec2016 | CH, ME, KA, AU | 244 (214) | reservoir inflows | CALIBRATION | 25Mar2017 |
11Nov2020–25Nov2020 | CH, ME, KA, AU | 295 (132) | reservoir inflows | CALIBRATION | 25Mar2017 |
Flood event . | Rain gauge data availability . | Total rain (highest one-day rain) (mm) . | Streamflow data availability . | Usage . | LULC . |
---|---|---|---|---|---|
02Dec2005–05Dec2005 | CH, ME, PA, SR, TAa | 360 (283) | reservoir inflows | CALIBRATION | 28Jun2002 |
25Nov2008–03Dec2008 | CH, ME, PA, SR, TA | 460 (188) | basin outlet | VALIDATION | 09Mar2011 |
02Nov2009–13Nov2009 | CH, ME, KA, AUa | 686 (151) | reservoir inflows | CALIBRATION | 09Mar2011 |
01Nov2010–09Dec2010 | CH, ME, KA, AU | 241 (75) | reservoir inflows | CALIBRATION | 09Mar2011 |
26Nov2015–04Dec2015 | CH, ME, KA, AU | 511 (383) | reservoir inflows | CALIBRATION | 25Mar2017 |
10Dec2016–18Dec2016 | CH, ME, KA, AU | 244 (214) | reservoir inflows | CALIBRATION | 25Mar2017 |
11Nov2020–25Nov2020 | CH, ME, KA, AU | 295 (132) | reservoir inflows | CALIBRATION | 25Mar2017 |
aCH, ME, PA, SR, TA, KA, AU stands for Chembarambakkam, Meenambakkam, Padappai, Sriperambadur, Tambaram, Kattupakkam, and Anna University, respectively.
Hydrological model
Though grid-based or fully distributed models are more accurate, the data limitations pose challenges in providing spatially varying rainfall inputs and parameter estimation. Therefore, a semi-distributed HEC-HMS model version 4.9 (USACE 2000) is used in this study. The model uses Soil Conservation Service (SCS) Curve Number (CN) and SCS Unit Hydrograph (UH) methods to estimate the effective rainfall excess of infiltration and runoff, respectively. The SCS CN method calculates effective rainfall based on the parameter CN, which is estimated based on LULC and Hydrological Soil Group (HSG). The SCS UH method estimates runoff based on the lag parameter, which is a function of CN, basin slope, length, and area. Hence, these methods primarily rely on only one parameter, namely CN. For LULC data, the multi-spectral Landsat 7 and 8 satellite imageries for 28 June 2002, 09 March 2011, and 25 March 2017 were acquired. Google Earth-assisted unsupervised classification was performed to obtain six LULC categories: water, trees, urban, agricultural fields, shrubland, and barren lands. For the six calibration events, these three LULC maps were used (Table 1) under the assumption that there won't be an appreciable land use change within 3 years. The HSG soil map was created based on the soil data from the National Bureau of Soil Survey and Land Use Planning (NBSSLUP), India.
Calibration approaches
As mentioned in Section 2.1.3, the CN that is estimated as a function of LULC is the calibration parameter. It is, therefore, required to allow the CN to be updated dynamically based on the changing LULC to quantify the impact of urbanization on flooding. The equifinality of the parameters in data-sparse regions also needs to be addressed. As the conventional calibration approaches account for either the uncertainty (Koutroulis & Tsanis 2010; Hdeib et al. 2018) or the dynamicity (Daksiya et al. 2017; Mishra et al. 2018; Nithila Devi et al. 2019, 2020) of the CN and not both, the question arises related to the reliability of using them to simulate floods under future urbanization projections. In that light, an urbanization-informed uncertainty-based calibration approach is proposed in this study, and its advantages are analyzed in comparison with the existing approaches.
To demonstrate the superiority of the proposed calibration approach, two existing approaches are used in this study to compare the calibrated values of the hydrological parameter, CN, in the data-limited Adyar basin comprising Chennai city. The approaches are briefly described in the following.
Existing approach – 1: In this approach, the schema or the look-up table that establishes the association of CN with the LULC is calibrated based on one or two historical flood events. Consequently, this allows for dynamic updating of CN with urbanization but does not account for the equifinality or uncertainty in model calibration, which may be the case in data-limited regions. The notation ‘Dyn_CN’ is used throughout the manuscript to denote this approach. Such a methodology has been used by Nithila Devi et al. (2019) for the data-sparse Adyar basin, which is the study area in this paper, too.
Existing approach – 2: The CN range for each sub-basin was narrowed down by evaluating several samples of CN for various historical flood events. Thus, considering the uncertainty, the behavioural CN range is estimated for each sub-basin. However, as the CN range remains fixed for each sub-basin, they cannot be varied dynamically with urbanization. This approach is denoted by the notation ‘uStat_CN’ in this paper. Hdeib et al. (2018) have proposed this methodology for simulating floods in a data-sparse basin in Lebanon.
Proposed urbanization-informed uncertainty-based approach: Here, the range of CN values, instead of deterministic CN values, in the schema that links it with the LULC is narrowed down using several historical flood events. For this reason, the merits of the earlier mentioned approaches, namely accounting for uncertainty and allowing for dynamic updating of CN with urbanization, have been combined in the proposed approach. This novel approach is denoted using the notation ‘uDyn_CN’ hereafter.
For the Dyn_CN approach, the calibrated CN look-up table that was validated with the Chembarambakkam reservoir inflows for the years 2005 and 2015, as reported in Nithila Devi et al. (2019), is used as such, as the study area is the same. The CN look-up table is nothing but the unique combinations of LULC and soil in the study area. Using this table and the distribution of the LULC and soil group within each sub-basin, the average CN is estimated for each sub-basin. The following steps are adopted for the proposed uDyn_CN calibration approach to narrow down the CN ranges in the CN-LULC schema (Figure 2). The first step is to create 10,000 random samples of CN schema for the different combinations of LULC and HSG. A range of 30–100 is used to create samples of CN schema using Python code. A value of 100 is assigned to CN for the water land use, and the Adyar Basin does not have the HSG class ‘B’. Individual HEC-HMS run is then set up for the past storm events (Table 1) by providing the corresponding rainfall data. A sub-basin-wise histogram of the various LULC and HSG combinations for the past storm HEC-HMS run based on the respective LULC map is estimated. A Python code is then used to generate 10,000 instances of sub-basin-wise averaged parameter estimates (CN and lag) based on its LULC and HSG histograms and the sampled schema for each storm event. The Uncertainty Analyses tool within HEC-HMS uses the Monte-Carlo algorithm to randomly select and run the instances. A uniform distribution of the instances is assumed in the process to avoid making any prior assumptions about the distribution (Freer et al. 1996; Hdeib et al. 2018). The corresponding simulated reservoir inflow volumes are estimated when the 10,000 instances of the parameters are run for each past event. The schema samples that produce < 5% volume error for the Chembarambakkam inflows are then selected and averaged for each storm event. Finally, the overall range of the CN schema is populated based on the event-wise averages. Thus, the narrowed-down range of the CN schema, presented in Table 2, is estimated for the data-limited Adyar basin considering the equifinality or parametric uncertainty.
Udyn_CN CN parameter ranges for different LULC and HSG combinations
LULC . | Hydrologic soil group . | ||
---|---|---|---|
A . | C . | D . | |
Water bodies | 100 | 100 | 100 |
Trees/Forest | 46–55 | 47–54 | 54–57 |
Urban | 71–81 | 73–79 | 76–77 |
Agriculture | 67–73 | 62–78 | 63–76 |
Barren | 73–76 | 68–75 | 68–75 |
Shrub | 46–60 | 56–67 | 58–64 |
LULC . | Hydrologic soil group . | ||
---|---|---|---|
A . | C . | D . | |
Water bodies | 100 | 100 | 100 |
Trees/Forest | 46–55 | 47–54 | 54–57 |
Urban | 71–81 | 73–79 | 76–77 |
Agriculture | 67–73 | 62–78 | 63–76 |
Barren | 73–76 | 68–75 | 68–75 |
Shrub | 46–60 | 56–67 | 58–64 |
Urban LULCs have higher CN values, and the trees/forest LULCs have lower CN values.
In the case of uStat_CN, the behavioural schema samples with less than <5% volume error that are selected in the uDyn_CN approach are used to arrive at sub-basin-wise CN values based on the particular sub-basin's LULC and HSG histograms. The sub-basin-wise CN and CN lag values are estimated in this way for each storm event. Based on this, eventually, the overall sub-basin-wise CN and lag ranges are estimated for the uStat_CN approach.
Module 2: analysis of the impact of urbanization on flooding
Urbanization by 2050 in the Adyar basin is modelled using a cellular automata model called SLEUTH (Clarke & Gaydos 1998; Dietzel & Clarke 2004; Chaudhuri & Clarke 2013). Cells are the fundamental processing units. The model evolves by changing the states of all the cells at each time step following the transition rules and cells' states at the previous time step. The transition/growth rules represent four different growth processes: diffusive growth, new spreading centre, edge growth, and road-influenced growth. Several studies have employed SLEUTH for prediction at a time span greater than that covered by the inputs (Hua et al. 2014; Bihamta et al. 2015; Clarke & Johnson 2020; Saxena & Jat 2020; Varquez et al. 2023). Further details on the inputs and the calibration process can be found in Nithila Devi et al. (2020). Nithila Devi et al. (2020) report nearly a 2.3 times increase in urban areas in the Adyar basin from 2015 to 2050.
Rainfall analysis: (a) flood frequency analysis for the Adyar river basin, (b) distribution of daily rainfall during past storm events whose peak day rainfall is greater than the 10-year return period rainfall depth, and (c) average distribution of normalized rainfall fractions of the past storm events with respect to peak day rainfall.
Rainfall analysis: (a) flood frequency analysis for the Adyar river basin, (b) distribution of daily rainfall during past storm events whose peak day rainfall is greater than the 10-year return period rainfall depth, and (c) average distribution of normalized rainfall fractions of the past storm events with respect to peak day rainfall.
Looking at the long-term daily rainfall record, it can be seen that the peak 1-day rainfall does not occur in isolation during the monsoon season in the Adyar basin. Figure 3(b) presents the normalized rainfall distribution before and after the peak day rainfall for events with the peak rainfall greater than the 10-year return period rainfall depth (224.7 mm). The mean normalized fractions of pre- and post-peak rainfall days for the past events are presented in Figure 3(c). These mean fractions are then used to generate the pre- and post-peak day rainfall depths for 50- and 100-year one-day return period rainfall scenarios. To test the applicability of the mean normalized fractions, they are applied to estimate pre- and post-peak rainfalls for the past events with peak one-day rainfall greater than the 10-year return period rainfall depth. The estimates are then compared with the observed pre- and post-peak rainfall values, and coefficient of determination (R2) and Root Mean Squared Error (RMSE) values of 0.81 and 52 mm, respectively, are obtained. This method may help get a more reliable visualization of flooding for various one-day return period scenarios and may help formulate realistic flood management measures. A more sophisticated method using machine learning algorithms can also be developed and applied in the future to estimate pre- and post-peak day rainfall values. Along with pre- and post-rainfall scenarios for 50- and 100-year return periods, historical flood events of 2005 and 2015 are also considered. This is done to better understand the changes in realistic flood dynamics in the region due to urbanization.
Scenarios 50YR:TYI and 100YR:TYI use one day 50-year and 100-year return period rainfall, respectively.
Scenarios 50YR:TYII and 100YR:TYII use one day 50-year and 100-year return period rainfall, respectively, along with prior and posterior rains (Figure 4(b)).
Scenarios 2005 and 2015 use observed rainfall of 2005 and 2015 floods, respectively.
Calibrated CN values and the proportion of vegetation and urban areas for the years 2002, 2011, 2017 and 2050 for the W160 sub-basin.
Calibrated CN values and the proportion of vegetation and urban areas for the years 2002, 2011, 2017 and 2050 for the W160 sub-basin.
To quantify the impact of urbanization on these rainfall scenarios, the simulations for each scenario were performed using the baseline 2017 and the projected 2050 LULC. BL and U50 represent the baseline 2017 and urban sprawl 2050 simulations of various rainfall scenarios. Thus, a scenario described as 50YR: TYI-U50 refers to a hydrological model simulation with one day 50-year return period rainfall and 2050 LULC.
RESULTS AND DISCUSSION
Uncertainty and dynamic nature of CN in different calibration approaches
To better understand how well uDyn_CN, uStat_CN, and Dyn_CN approaches incorporate uncertainty and dynamicity with urbanization, the estimated CN values from different approaches for the sub-basin W160 (refer to Figure 1) are plotted in Figure 4 for the years 2002, 2011, 2017, and the future 2050. The W160 is the sub-basin that contributes to nearly half of the inflows into the Chembarambakkam reservoir that was used for calibration. Urban areas constitute 3.8, 6.5, 7.4, and 26.5% of the total sub-basin area in 2002, 2011, 2017 and 2050, respectively. The vegetation areas (including agriculture, forest and shrub) constitute around 80.7, 67.8, 69.9, and 55.6% of the total sub-basin area in 2002, 2011, 2017 and 2050, respectively. This shows a clear decreasing trend in vegetation and an increasing trend in urban areas due to urbanization. It can be seen that though the CN values change with urbanization in the Dyn_CN approach, it does not encompass a range of values to address the possible uncertainty in a data-limited region. On the other hand, the uStat_CN method offers a range of CN values for the sub-basin, but these values remain constant over time, making it challenging to quantify the increase in flood flows due to urbanization.
Validation of the hydrological parameterization schema
Volume errors for the reservoir inflows based on the three calibration approaches for the past storms
Flood event . | Observed volume (106 m3) . | Volume error (%) . | ||
---|---|---|---|---|
uDyn_CN . | Dyn_CN . | uStat_CN . | ||
02Dec2005–05Dec2005 | 78.6 | 17.2 | 27.9 | 23.5 |
02Nov2009–13Nov2009 | 19.8 | 2.2 | 8.3 | 2.7 |
01Nov2010–09Dec2010 | 28 | 16 | 33.9 | 36.5 |
26Nov2015–04Dec2015 | 112.6 | 7.1 | 7.6 | 8 |
10Dec2016–18Dec2016 | 16.4 | 3.7 | 15.1 | 22.8 |
11Nov2020–25Nov2020 | 35.3 | 11.1 | 16.8 | 17.4 |
Average | 9.5 | 18.3 | 18.5 |
Flood event . | Observed volume (106 m3) . | Volume error (%) . | ||
---|---|---|---|---|
uDyn_CN . | Dyn_CN . | uStat_CN . | ||
02Dec2005–05Dec2005 | 78.6 | 17.2 | 27.9 | 23.5 |
02Nov2009–13Nov2009 | 19.8 | 2.2 | 8.3 | 2.7 |
01Nov2010–09Dec2010 | 28 | 16 | 33.9 | 36.5 |
26Nov2015–04Dec2015 | 112.6 | 7.1 | 7.6 | 8 |
10Dec2016–18Dec2016 | 16.4 | 3.7 | 15.1 | 22.8 |
11Nov2020–25Nov2020 | 35.3 | 11.1 | 16.8 | 17.4 |
Average | 9.5 | 18.3 | 18.5 |
Only the volume errors were quantified considering the observed reservoir inflows. This is because the recordings are based on daily budgeting of average inflows, storage, and releases in the reservoir. Moreover, if the hourly recordings of the inflows or river discharges had been available for the flood events, it would have made sense to estimate peak flow error, too. Therefore, as the daily reservoir inflows were only available, calibration was performed based on volume error. An average of 9.5% volume error has been estimated for the simulations from the proposed uncertainty-based calibration approach, uDyn_CN. The volume error is 18.3% for deterministic calibration (Dyn_CN), almost double that of the proposed uncertainty-based calibration. Meanwhile, the average volume error for the uStat_CN approach is around 18.5%. This shows that the proposed uncertainty-based calibration approach, uDyn_CN, performs better than the conventional calibration approaches (uStat_CN and Dyn_CN) for the study area for events of different rainfall intensities.
Peak flows and inundated areas under various hazard zones for different scenarios
Scenarios . | Mean peak flows (m3/s) . | Percentage increase in peak flows (%) . | Mean inundated area (km2) . | Percentage increase in mean inundation area (%) . | ||||
---|---|---|---|---|---|---|---|---|
Low . | Moderate . | High . | Very High . | Total . | ||||
50YR:TYI-BL | 1,116.3 | – | 13 | – | – | – | – | – |
50YR:TYI-U50 | 1,257.6 | 12.7 | 14.6 | 15.6 | 7.4 | 9.2 | 6.1 | 8.8 |
50YR:TYII-BL | 2,390.5 | – | 27.9 | – | – | – | – | – |
50YR:TYII-U50 | 2,619.6 | 9.6 | 30.3 | 6.4 | 16 | 6.5 | 5 | 8.5 |
100YR:TYI-BL | 1,345.5 | – | 15 | – | – | – | – | – |
100YR:TYI-U50 | 1,512.5 | 12.4 | 16.7 | 20.5 | 10.1 | 9.8 | 6 | 11 |
100YR:TYII-BL | 2,633.6 | – | 30.9 | – | – | – | – | – |
100YR:TYII-U50 | 2,887.8 | 9.7 | 33.4 | 2 | 15.8 | 6.7 | 5.7 | 7.9 |
2005-BL | 1,875.8 | – | 19.6 | – | – | – | – | – |
2005-U50 | 2,249 | 19.9 | 26.5 | 59.1 | 52.2 | 22.6 | 16.6 | 26.5 |
2015-BL | 3,387.7 | – | 39.2 | – | – | – | – | – |
2015-U50 | 3,700.4 | 9.2 | 41.9 | 1.2 | 6 | 11.6 | 7.1 | 6.9 |
Scenarios . | Mean peak flows (m3/s) . | Percentage increase in peak flows (%) . | Mean inundated area (km2) . | Percentage increase in mean inundation area (%) . | ||||
---|---|---|---|---|---|---|---|---|
Low . | Moderate . | High . | Very High . | Total . | ||||
50YR:TYI-BL | 1,116.3 | – | 13 | – | – | – | – | – |
50YR:TYI-U50 | 1,257.6 | 12.7 | 14.6 | 15.6 | 7.4 | 9.2 | 6.1 | 8.8 |
50YR:TYII-BL | 2,390.5 | – | 27.9 | – | – | – | – | – |
50YR:TYII-U50 | 2,619.6 | 9.6 | 30.3 | 6.4 | 16 | 6.5 | 5 | 8.5 |
100YR:TYI-BL | 1,345.5 | – | 15 | – | – | – | – | – |
100YR:TYI-U50 | 1,512.5 | 12.4 | 16.7 | 20.5 | 10.1 | 9.8 | 6 | 11 |
100YR:TYII-BL | 2,633.6 | – | 30.9 | – | – | – | – | – |
100YR:TYII-U50 | 2,887.8 | 9.7 | 33.4 | 2 | 15.8 | 6.7 | 5.7 | 7.9 |
2005-BL | 1,875.8 | – | 19.6 | – | – | – | – | – |
2005-U50 | 2,249 | 19.9 | 26.5 | 59.1 | 52.2 | 22.6 | 16.6 | 26.5 |
2015-BL | 3,387.7 | – | 39.2 | – | – | – | – | – |
2015-U50 | 3,700.4 | 9.2 | 41.9 | 1.2 | 6 | 11.6 | 7.1 | 6.9 |
The bold numbers in the table indicate highest values in each scenario.
Validation of the three calibration approaches for the 2008 flooding event. For better understanding, the inset plot shows the peak discharge values of the event.
Validation of the three calibration approaches for the 2008 flooding event. For better understanding, the inset plot shows the peak discharge values of the event.
Hydraulic validation
In data-limited regions, the errors in the hydrological parameter estimation outweigh the model structure (Hdeib et al. 2018). The errors in estimating hydrological parameters also propagate into the hydraulic simulations in the coupled hydrological-hydraulic modelling. In that light, to understand the efficiency of the three calibration approaches, the mean hydrographs from the hydrological model setups at the entry of the Adyar River into the city are routed in the hydraulic domain using a 2D local-inertial hydraulic model in IROMS (Bates et al. 2010; De Almeida et al. 2012; Sridharan et al. 2020). Figure 1 shows the location of the inflow boundary condition and the hydraulic domain consisting of CMA. A high spatial resolution LIDAR DEM of 2 m is used to capture finer terrain features such as highways, canals, and major drains. The building data from Open Street Map (OSM) (http://www.openstreetmap.org) was acquired. The pixel values of the DEM that correspond to buildings were increased by a height of 5 m to avoid the simulation of unphysical flows through them. This simple approach efficiently captures flow dynamics around urban features (Reshma et al. 2024). Previous studies (Ghosh et al. 2019; Nithila Devi et al. 2019; Sridharan et al. 2020) on the Adyar River basin have used a Manning's friction value of 0.035 m−1/3s, and the same value is adopted in this study.
Hydraulic model validation: (a) map showing the location of post-flood survey marks and the maximum inundation depth simulated by the uDyn_CN approach for the 2015 floods, and (b) the absolute difference between the observed and simulated maximum inundation depths for the 2015 floods for different approaches. Values closer to zero on the y-axis indicate lesser error.
Hydraulic model validation: (a) map showing the location of post-flood survey marks and the maximum inundation depth simulated by the uDyn_CN approach for the 2015 floods, and (b) the absolute difference between the observed and simulated maximum inundation depths for the 2015 floods for different approaches. Values closer to zero on the y-axis indicate lesser error.
Given the limited data available, the novel uDyn_CN calibration approach that accounts for urbanization effects and uncertainty in the coupled hydrological-hydraulic framework yields more accurate flood simulations. Consequently, this approach can provide a reliable means of analyzing the influence of urbanization on flooding. Therefore, the uDyn_CN calibration approach is further used to analyze the impact of urbanization by 2050 on flooding in the data-scarce Adyar basin.
Impact of urbanization on flooding
The impact of urbanization on flooding is analyzed with respect to (i) peak flows, (ii) inundated areas, and (iii) inundated areas under different hazard zones for various scenarios. The discussion in the following text elaborates on the findings derived from both the baseline and future urbanization flood scenarios in these aspects.
Impact on peak flows
(a) Peak flow ranges and (b) percentage increase in mean peak flows for various urbanization scenarios.
(a) Peak flow ranges and (b) percentage increase in mean peak flows for various urbanization scenarios.
Impact on inundation area
The mean flows are routed in the hydraulic domain using the IROMS model, and the maximum inundation areas are extracted for various baseline and urbanization scenarios. Table 4 presents the inundation areas for all the scenarios. It also shows the percentage increase in inundation areas for the urbanization scenarios compared to the corresponding baseline scenarios. The percentage increase in inundation areas for TYI: 50- and 100-year flood scenarios are 8.8 and 11%, respectively. In the case of TYII: 50-year and 100-year flood scenarios, the percentage increase in inundation areas is 8.5 and 7.9%, respectively. It can be seen that similar to the trend in peak flows, urbanization has caused the mean 50-year flood to inundate the domain like that of a baseline 100-year flood. Also, the increase in inundated areas is greater in the TYI scenarios that are just one-day rainfall simulations without any pre- and post-event rainfall. However, in the study area, the monsoon rains of considerable magnitude mostly resemble TYII, as shown in Figure 3. The 2005 and 2015 flood scenarios have increased the total inundation area by 26.5% and 6.9%, respectively.
Impact on flood hazard zones
To understand the extent of damages caused by floods, threshold depths of 0.6, 1.4, and 3.5 m are chosen to define hazard zones of various categories. The following are the various hazard categories adopted in this study,
Low hazard zone with a depth ranging between 0.1 and 0.6 m and flood waters enter the buildings in this zone.
Moderate hazard zone with a depth ranging between 0.6 and 1.4 m and half of the ground floor gets inundated up to the height of the furniture. Also, the cars nearly get submerged.
High-hazard zone with a depth ranging between 1.4 and 3.5 m, where the ground floor of the buildings gets submerged and flood waters may enter the first floor.
Very high-hazard zone with a depth greater than 3.5 m, where the first floors of the buildings get affected
Studies have used depth-based classification to indicate the flood severity level in coastal areas (Nithila Devi et al. 2019; Mitchell et al. 2022). This is relevant as the flood velocities may not vary drastically in the slow-rising monsoon floods of the coastal Chennai city. The damage to residential, industrial and commercial buildings due to floods is also typically estimated based on flood depths and the socio-economic status of the considered area (Huizinga et al. 2017).
(a) Inundated area and (b) percentage increase in areas under different hazard zones for various urbanization scenarios.
(a) Inundated area and (b) percentage increase in areas under different hazard zones for various urbanization scenarios.
CONCLUSIONS
The Dyn_CN, uStat_CN and uDyn_CN calibration approaches use observed daily inflows into a reservoir for six historical flood events (2005, 2009, 2010, 2015, 2016, and 2020) of various intensities. Of the two approaches capable of modelling the implications of urbanization on flooding by allowing dynamic parameter changes, the uDyn_CN approach outperformed Dyn_CN across various flood intensities, reducing the average inflow volume error by nearly half. For the overall validation, the flooding event of the year 2008 is considered as the hourly discharges available at the outlet of the catchment. The uDyn_CN approach managed to capture the observed peak flow within its simulation bounds and reported less volume error than the other approaches. Hence, the proposed ‘urbanization-informed uncertainty-based calibration approach’ can be considered robust in the hydrological simulation of varying flood intensities in a data-limited region.
The mean inflows from the hydrological model for the three approaches are then routed into the hydraulic domain consisting of urban areas using the 2D local-inertial hydraulic model in IROMS. The post-survey flood marks for the 2015 floods have been used to assess the hydraulic model setups based on the three hydrological calibration approaches. This exercise was performed to assess how the errors in hydrological model calibration propagate into the hydraulic model simulations. It was seen that the RMSE values for the uDyn_CN, uStat_CN and Dyn_CN approaches are 0.22, 0.36 and 0.6 m, respectively. Thus, the proposed hydrological calibration approach, uDyn_CN, performs better than the existing approaches in the overall coupled hydrologic-hydraulic modelling framework in a data-limited, rapidly urbanizing basin.
To assess the impact of urbanization on flooding in the Adyar catchment, the 2017 and the projected 2050 LULCs (obtained from a cellular automata model called SLEUTH) are considered baseline and future urbanization scenarios, respectively. Following are the major inferences from the urbanization impact analyses carried out for the study area,
Due to urbanization, both the peak flows and maximum inundation extents exhibited a significant increase for moderate floods (50YR:TYI, 100YR:TYI and 2005) than extreme floods (50YR:TYII, 100YR:TYII and 2015).
The 50-year return period flood scenarios behaved like 100-year return period flood scenarios in terms of mean peak flows and inundation extents.
Under urbanization, for the extreme flood scenarios (2015 and 100YR:TYII), the hazard zones of high flood severity show greater increases than hazard zones of low severity.
As the calibration framework arrives at the schema linking LULC and hydrological parameter ranges, it can be applied to a grid-based or fully distributed hydrological model with the availability of spatially distributed rainfall data from a radar or satellite. In that way, the proposed calibration approach can aid in addressing the complexity of parameterization in fully distributed models. This has not been explored in the current paper; hence, it presents a scope for further research. The study has not looked into how the parameter ranges vary when calibrated with a different number of events, as the data were available for only seven events. Even for the seven flood events, only daily reservoir inflows were available. Alternatively, if the hourly hydrographs were available close to the basin outlet for several events, we might have got narrower parameter ranges. Nonetheless, compared to the existing approaches, the constrained parameter ranges estimated from the proposed approach perform well from hydrological and hydraulic perspectives for a data-limited basin. As another possible extension of this study, the superiority of the proposed approach can be analyzed in well-gauged basins, either only in the hydrological simulations or in the coupled hydrological and hydraulic modelling framework. Though the proposed framework uses the HEC-HMS model, any hydrological model that links its parameter values and LULC can be employed. This proposed calibration approach addresses both the uncertainty associated with data-limited regions and the dynamicity of the hydrological parameters associated with urbanization. Therefore, it is compared with the other two relevant approaches, which do not account for either of them. Comparing the proposed calibration framework with other approaches is beyond the scope of this paper and needs to be investigated in the future.
The key takeaway from this research is that the novel ‘urbanization-informed uncertainty-based calibration approach’ proposed in the paper enables dynamic updating of hydrological parameters based on the changing land use. Additionally, this approach accounts for parametric uncertainty. Thereby, it facilitates reliable flood simulation and hazard estimation considering the impact of urbanization in a data-limited region. This, in turn, may aid better flood preparedness, planning and management in the rapidly urbanizing data-limited catchments known for their high vulnerability to floods.
AUTHOR CONTRIBUTIONS
N.N.D. contributed to conceptualization; data curation; formal analysis and investigation; methodology; visualization; writing – original draft preparation and editing. S.N.K. contributed to funding acquisition; supervision; resources; writing – review and editing.
FUNDING DETAILS
This work was supported by the Ministry of Electronics & Information Technology, the Government of India, under the project titled, ‘Urban Modelling: Development of Multi-sectorial Simulation Lab and Science-Based Decision Support Framework to Address Urban Environment Issues’ (Sanction Number: MeitY/R&D/HPC/2(1)/2014).
DATA AVAILABILITY STATEMENT
Data cannot be made publicly available; readers should contact the corresponding author for details.
CONFLICT OF INTEREST
The authors declare there is no conflict.