A computationally efficient flash flood early warning system for a mountainous and transboundary river basin in Bangladesh

Flash floods are one of the most dangerous natural hazards due to their sudden nature of occurrence. They are generally caused by heavy or excessive rainfall over a very short period of time, where runoff generation is accelerated due to saturated or poorly infiltrating soil, or steep terrain. The time difference between the rainfall event and flood peak is often less than 12 hours, which makes forecasting extremely challenging (Lee & Singh 1999). Flash floods are the number-one cause of death among all storm-related hazards in the USA, with approximately 100 lives lost each year (Ashley & Ashley 2008).

Over the past few decades, flash floods have drawn attention in many parts of the world including the United States, European Union, Australia, and Egypt (Hapuarachchi et al. 2011; Abuzied & Mansour 2019). Flash floods are of particular concern in the northeastern region of Bangladesh (shown in Figure 1(a)), which experiences flash floods nearly every year during the pre-monsoon season (March–May). This area is home to extensive agriculture and aquaculture activities that support the country's economy and food security. In the dry season, the region contributes 16% of the total rice production of the whole country (Quddus 2009). However, the pre-monsoon sudden peak discharge as a flash flood from the adjacent upstream areas often causes immense damage to the harvest-ready crops. The March 2017 flash flood affected nearly one million people and caused a potential loss of approximately one million metric tons of crops, which accounts for 3.7% of the country's agriculture sector GDP. This event compromised the overall food security of the country (ReliefWeb Flood Situation Report 2017). The northeastern region has been experiencing the same type of flash flood every year with varying degrees of damages and consequences. An operational flash flood forecasting and early warning system is critical for ensuring the safety of the local population and protecting the food security of Bangladesh.

One challenge in developing a skillful forecasting framework in the northeastern region of Bangladesh is the complex nature of the flash flood generation mechanism. The region consists of interconnected stream networks, perennial and seasonal channels, lowland regions, and internal depressions locally known as Haors. In Figure 1(b), a schematic diagram of the area is shown to illustrate the hydraulic connectivity of channels and floodplains. The flash flood region is located in the downstream portion of the transboundary Meghna river basin. The upstream region consists of the mountainous Meghalaya, Tripura, and Barak river basins of India. In Figure 2, the topography of the basin shows the rapid transition from hilly areas in the northern upstream region to floodplains in the southern and southwestern parts inside Bangladesh. Tributaries in the north originate from Khasi and Jaintia Hills, which receive an average of 12,000 mm of rainfall annually and are some of the wettest places of the world (Parry 2013). In upstream channels where precipitation is orographically enhanced by mountainous terrain, hydrologic processes dominate the flood generation mechanism. This shifts in the downstream floodplain, where hydrodynamic processes control flood propagation. Thus, the forecasting of flash flood-inducing flow depends on the ability of hydrological and hydrodynamic models to represent the processes at the small scales of interest as well the availability and accuracy of rainfall, soil moisture, and topographical data used to drive the models (Yatheendradas et al. 2008; Simons et al. 2014). This also means that conventional techniques used in river-based flood forecasting are not suitable for flash floods in floodplains where two-dimensional hydrodynamic processes dominate the hydrological or channel flow processes (Sangati 2009). The characteristics of the rain (intensity, duration, amount, and time-space distribution) and the physical and hydrological characteristics of the watershed (area, length, slopes, shape, type of soil and land cover, and antecedent conditions) need to be accurately captured for any framework to forecast flash floods.

Another critical ingredient for flash flood forecasting is forecasted precipitation. Forecasted precipitation is a key parameter in forecasting flows for a lead-time longer than the time of concentration of a river basin (which is short for mountainous basins). Recent advancements in satellite earth observations and enhanced capabilities of computer models now enable quantitative precipitation forecasts (QPF) (Liguori et al. 2012; Liu et al. 2015). Forecasted precipitation produced by numerical weather prediction (NWP) models, with global scale coverage, and near real-time availability, are already being applied in river flow forecasting and monsoon flood forecasting (Verbunt et al. 2006; Roberts et al. 2009; Liguori et al. 2012). Some recent studies also dynamically downscaled the NWP forecasts by using high-resolution weather research and forecasts (WRF) model along with terrain and land-use features (Rao et al. 2007; Hsiao et al. 2013; Sikder & Hossain 2016). Although WRF-based downscaling methods improve the accuracy of the forecasts, computational resources, and limited internet accessibility are a concern in developing countries like Bangladesh (Sikder & Hossain 2016). Thus, it is also crucial for flash flood forecasting methods to be computationally efficient for use in an operational setting without compromising accuracy.

Among the available methods, the most widely used operational method for flash flood forecasting is flash flood guidance (FFG) method (Georgakakos 1986). Estimation of threshold runoff volume of various durations and soil moisture accounting (Sweeney 1992) is required in this method. It is usually lumped over the basin; spatial and temporal distribution of threshold runoff within the basin and influence of soil and land cover characteristics are not considered. Other approaches have used multisensor data and neural networks (NN) for flash flood forecasting (Kim & Barros 2001; Piotrowski et al. 2006; Chiang et al. 2007). Kim & Barros (2001) used NWP forecasts, wind, and pressure data along with rain gauge data in an artificial neural network to produce streamflow forecasts with up to 24 h lead-time. Chiang et al. (2007) developed a data-based approach using recurrent neural network using gauge observations and satellite-derived PERSIANN-CCS (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System) precipitation (Hong et al. 2004). Chen & Yu (2007) proposed a probabilistic approach for flood stage forecasts based on hourly water stage and rainfall data using support vector regression and probability distribution of forecast error based on fuzzy inference. In most methods published in the literature, hydrological and topographical characteristics of the catchments do not appear to be explicitly represented, except for rainfall and discharge. Also, all of the statistical and probability-based approaches (i.e., Sajikumar & Thandaveswara 1999; Wang et al. 2015; Kratzert et al. 2018) require a long time series for training the model which is not available in case of ungauged and transboundary basins, like northeastern Bangladesh.

The most recent advancement in computing has made modeling approaches to flash flood forecasting more feasible (Liang & Smith 2015). Jasper et al. (2002) first used numerical weather prediction models to produce meteorological observations and then used them as forcing data in a distributed hydrological model to forecast flash floods in an alpine watershed. England et al. (2007) applied a physically based distributed two dimensional, runoff, erosion, and export (TREX) model to simulate (hourly) extreme floods in semi-arid regions in the western United States. There are also atmosphere–hydrological coupled models, such as the National Center for Atmospheric Research (NCAR) developed WRF-Hydro modeling system (Gochis et al. 2015). For flash floods, WRF-Hydro can be applied to simulate high-resolution hydrometeorological processes such as surface overland flow, saturated subsurface flow, channel routing, and baseflow processes. The lack of representation of interconnection between stream networks is however a common limitation in all hydrological models, which makes WRF-Hydro unsuitable for northeastern Bangladesh. It is necessary to include hydrodynamic/hydraulic models in conjunction with hydrologic models to properly represent the hydraulic processes in the river channels and floodplains.

Besides the discharge and river flow forecasts, it is also necessary to produce an accurate mapping of the spatial extent of inundation forecasts to trigger location-specific warnings, damage assessment as well as to plan relief and rehabilitation. A very limited number of studies used the hydrodynamic modeling approach, and those that did were limited to one-dimensional (1D) flow simulation (Carpenter et al. 1999; Ghoneim & Foody 2013) to produce station-based forecasts. To the best of our knowledge, there are no such published studies where two-dimensional hydraulic processes have been represented along with hydrological models in an operational framework for flash flood/flood forecasting. Possible reasons for this might be due to the reasons of (i) complexity in modeling development, calibration, and validation of multiple models that depend extensively on meteorological, hydrological, and topographic data and (ii) computation time of simulation of a complex framework consisting of meteorological–hydrological–hydraulic processes.

In this study, the objective is to explore the optimal combination of scale, time step, and complexity in topographic and hydrometeorological observations for developing a flash flood forecasting system. The core component of the framework is based on making use of globally available weather forecast data in real-time from NWP models. These precipitation forecasts can drive the hydrologic–hydrodynamic framework without any computationally expensive downscaling. As a first step, the approach here is to investigate how publicly available forecasts on precipitation from the global NWP model like GFS perform and whether skill remains acceptable without any regional NWP-based downscaling using WRF. The specific research questions that this study answers are: (1) What is the optimal combination of scale, time step, and complexity for topographic and hydrometeorological data to produce skillful flash flood forecasts? and (2) What is the baseline accuracy of a flash flood forecasting system that is achievable without dynamic downscaling of NWP weather forecasts? In the paper, datasets and detailed methodology are described in the section Data and methods, followed by Results and findings. Major findings, conclusions, and recommendations are presented in the final section.

We used Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS, link: http://legacy.chg.ucsb.edu/data/chirps/index.html) (Funk et al. 2015) as the reference precipitation product during the model calibration–validation phase. The temporal accumulation is of daily scale, and the spatial resolution of the dataset is 0.05° by 0.05°. As a nowcast source of precipitation for the operational framework, we explored two near-real-time precipitation products, namely, GSMaP-NRT (link: https://sharaku.eorc.jaxa.jp/GSMaP_NOW/index.htm) (Okamoto et al. 2005; Kubota et al. 2007; Aonashi et al. 2009; Ushio et al. 2009) and Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG, link: https://pmm.nasa.gov/data-access/downloads/gpm) Early Run Version 05B. The spatial resolution of GSMaP and IMERG precipitation is 0.1° by 0.1° and temporal resolution is half hour (IMERG) and 1 hour (GSMaP). Both of the products are available with a latency of 6 hours.

Solar radiation, average wind speed, maximum and minimum temperature, relative humidity datasets are also collected from various sources. Gridded monthly solar radiation data were collected from NASA Prediction of Worldwide Energy Resources (link: https://power.larc.nasa.gov/data-access-viewer/). Average wind speed, maximum and minimum temperature, and relative humidity data were retrieved from the National Climatic Data Center of NOAA (https://www7.ncdc.noaa.gov/CDO/cdoselect.cmd?datasetabbv=GSOD). In-situ observed WL were provided by Bangladesh Water Development Board (BWDB) (www.bwdb.gov.bd). Rated discharge was generated using the rating curve provided by the same institution.

NWP products are the key parameter used in this study. The Global Forecasting System (GFS) developed by the National Oceanic and Atmospheric Administration (NOAA) was used as the source of forecast precipitation from a global NWP model. Historical dataset of GFS forecast precipitation was downloaded from https://www.ncdc.noaa.gov/data-access/model-data/model-datasets/global-forcast-system-gfs. In the nowcast system we developed, the datasets are downloaded from https://nomads.ncep.noaa.gov/cgi-bin/filter_gfs_0p25.pl utilizing the online spatial and parameter-based subsetting facility. The GFS weather forecast products were updated every 6 hours up to 16 days lead-time at a spatial resolution of 0.25° by 0.25°. The temporal accumulation of precipitation products used in this study was 3 hours.

The Shuttle Radar Topography Mission (SRTM) developed void filled Version 3.0 SRTM Global 1 arc second product was used as the DEM, downloaded from https://earthexplorer.usgs.gov/. Land cover data for the hydrologic and hydrodynamic models were derived from Climate Change Initiative-Land Cover Maps of 2015 by using MERIS of Envisat along with SPOT-Vegetation, AVHRR, and PROBA-Vegetation mission datasets which are available at http://maps.elie.ucl.ac.be/CCI/viewer. Soil data and information related to soil properties were retrieved from http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/en/.

Sentinel-1 synthetic aperture radar (SAR) imagery products were used to compare the framework-generated flood inundation maps. SAR is an all-weather satellite with the advantage of operating at wavelengths not impeded by cloud cover and with day–night image acquisition. The images are at 10 m spatial resolution with an average temporal resolution of 10 days. All of the Sentinel images were processed using Google Earth Engine (GEE) (Gorelick et al. 2017). A backscatter coefficient of −14 db (Ahmad et al. 2019) was used to differentiate between water and non-water features. Finally, the extracted maps were used to assess the accuracy of the generated forecasts.

The key steps of our flash flood forecasting framework development are DEM correction and recondition, base hydrological and hydrodynamic model calibration–validation, and operationalization of the system. Each of these steps is discussed in the following sub-sections.

Generally, rivers in a high terrain region can easily be delineated using any hydrological analysis tool (e.g., Automatic Watershed Delineation of SWAT, ArcHydro tool of ArcGIS). However, in the floodplain, it is quite difficult to distinguish between the floodplain and the river channels using automated tools. For accurate representation, the river network in the floodplain was manually derived first using Google Earth. The automatic watershed delineation tool of the ArcSWAT model was also used to delineate the river network in the basin. The delineated network and manually digitized networks were then merged to create a more accurate river network map. Using this river network, the SRTM DEM was reconditioned and the watersheds were delineated for further processing. The DEM was reconditioned using a sharp drop of 10 m (burning) and a smooth drop of 5 m along the channels. This reconditioned DEM was finally used as the topography input into the hydrological model. During the development of the base hydrodynamic model, sensitivity analysis was performed using different amounts of systematic errors in SRTM DEM tested (i.e., 2.0 m, 2.5 m, 3.0 m, 3.5 m, and 4.0 m). Finally, it was found that an assumption of 3.5 m of error in SRTM DEM produced the best estimation of flood inundation when compared with independent and in-situ observations.

The Soil and Water Assessment Tool (SWAT) was used here to develop the hydrological model. SWAT is a river basin-scale hydrological model used to simulate the quantity of surface and groundwater. The hydrologic cycle simulated by SWAT is based on the water balance equation. It is a semi-distributed physically based model that divides the basin into a number of hydrologic response units (HRU). An HRU is a unit of area with uniform land use, soil type, elevation, and slope. We used the SWAT model at daily time step to simulate streamflow at the floodplain boundary locations. The model was developed using the reconditioned DEM, land use, and soil data. For the SWAT model, the following parameters were computed: monthly climatology of precipitation; average number of days of precipitation; standard deviation of precipitation; skew coefficient of daily precipitation; probability of wet day following a dry day; probability of wet day following a wet day; maximum 0.5 h rainfall; maximum and minimum air temperature; average wind speed; daily dew point temperature; and daily solar radiation. Most of these parameters (monthly climatology of precipitation; average number of days of precipitation; standard deviation of precipitation; skew coefficient of daily precipitation; probability of wet day following a dry day; probability of wet day following a wet day; maximum 0.5 h rainfall) were calculated using CHIRPS precipitation between 1980 and 2017 or climatology datasets from the National Climatic Data Center (NCDC). Except for precipitation, climatology of all other datasets was used during simulation of the SWAT model. During the calibration phase, CHIRPS precipitation dataset was used in this model. The model was calibrated using the observed data from 2004 to 2012. The optimized parameters were used to validate the model from 2013 to 2017. In Figure 2, the model domain is shown along with basin boundary and modified stream network. Inflow calibration locations of the hydrological model are also shown in the same figure.

Using the Climate Change Initiative (CCI) developed land cover data, Manning's n values were defined for overland flow in the floodplain. Hydraulic Engineering Center-developed HEC-RAS version 5.0.5 was used to develop the hydrodynamic model. The main advantage of this model is improved stability due to the use of an implicit finite volume approach, which is more robust than the finite element or finite difference method (the earlier version of HEC-RAS). It allows a larger simulation time step compared to explicit methods allowing a higher value of Courant–Friedrichs–Lewy (CFL) condition. Diffusion wave or full momentum 2D equations are used in this model. The simulation grid cells have stage–storage relationships derived from terrain or DEM dataset, allowing larger computational cells without loss of terrain details. The hydraulic geometry was created as a 2D flow area of the floodplain. Manning's n was assigned according to the landcover type following the suggestions from the HEC-RAS River Analysis System 2D Modeling User's Manual. The hydrological model-generated calibrated inflow into the floodplains was defined as the upstream boundary condition. The downstream boundary condition of the model was the observed water level during the calibration–validation of the framework. Average precipitation over the floodplain was also included as a boundary condition into the model. The domain of the hydrodynamic model along with unsteady boundary locations are shown in Figure 3.

Using the optimized selection of model configurations and the error correction strategy, the framework was made functional for real-time operations. In the operational setting, latest GPM-IMERG nowcast and GFS forecast precipitation data were downloaded at every model time step. GFS forecast precipitation was resampled to 0.1° by 0.1° to match the resolution of GPM-IMERG precipitation. The SWAT hydrological model was simulated and the inflow at different locations calculated. These inflows were then used in the HEC-RAS model as boundary flow condition. For the forecast period, different downstream boundary conditions were tested (e.g., model simulated discharge, rating curve, normal depth) and it was found that the hydrodynamic model performed best when hydrological model simulated discharge was used as the downstream boundary condition.

The HEC-RAS model was then simulated for every forecast time step using a restart option to conserve computational time. The computed water heights and inundation maps were then corrected using the error correction technique mentioned in the section ‘Bias correction of simulated results’ and Equations (1) and (2). Finally, all the simulated maps and time series forecasts were pushed to the web-based user-interface at http://depts.washington.edu/saswe/flashflood. On this web-interface, users can view the real-time forecast precipitation, discharge, water heights, and inundation for up to 5-day lead-time. In Figure 4, the methodology for the operational system is summarized as a flowchart.

In this section, all the findings during calibration–validation and operationalization are discussed. The stations used during performance assessment of hydrological and hydrodynamic models including calibration and validation are shown in Figure 5.

The SWAT model was calibrated manually using rated discharge at six inflow boundary stations. These were the locations within the basin and on the boundary of the floodplain which contributes almost all of the inflows into the floodplain (shown in Figure 2). Calibrated parameters were selected from Neitsch et al. (2001) and their optimized values are reported in Table 1. These parameters are: (i) three surface water components (curve number and plant uptake and soil evaporation factor); (ii) two soil water factors (available soil water capacity and soil saturated hydraulic conductivity); and (iii) four groundwater parameters (groundwater delay, groundwater revap constant and re-evaporation threshold). Among these parameters, curve numbers were optimized by a multiplication factor of 1.231 from the original values and the others were replaced with the mentioned values.

Calibration–validation summary and statistics of the base hydrologic model are presented in Table 2. As an example, Amalshid station's calibration and validation comparison and scatterplots are shown in Figure 6. It was observed that for all stations, the model matches well with observations, except for a few unusual peaks. The unusual peaks and deviation from observations occurred mainly during the monsoonal flood season rather than the pre-monsoon flash flood season. The independent validation period yielded similar or improved results as the calibration period because of fewer numbers of extreme peaks in the flows. Naukaugaon station (shown in Figure 2) showed a very unrealistic mismatch with the simulated results. We believe this is an issue related to the rating curve used, which represents low to moderate flow, but is inappropriate for high flow. Using this calibrated model, the hydrodynamic model was simulated and validated in different locations within the floodplain.

Using the output of the calibrated hydrological model, the hydrodynamic model (HEC-RAS) was simulated over the corrected and reconditioned DEM for the floodplain region inside Bangladesh. Calibration of the hydrodynamic model was completed manually at different locations by optimizing Manning's n values. Optimized Manning's n values are shown in Table 3. In Figure 7, the comparison plot of simulated and observed water levels of Sunamganj and Sylhet stations are shown. Overall, the results of the hydrodynamic model are slightly more accurate in the calibration period compared to the validation period except at Sarighat station. We also observed that the simulated water level hydrographs at the stations lying within the floodplains (e.g., Sunamganj, Sheola, Sylhet, Sherpur) performed better than the stations located at the edge of the hydrodynamic model boundary (Amalshid, Manu-RB) (Table 4). Flows at the upstream boundary locations of the floodplains were more hydrologically dominated, which might explain the reason for lower hydrodynamic prediction accuracy downstream.

Using the calibrated and validated modeling framework, GPM IMERG and GSMaP NRT precipitation were used to simulate water heights at four different locations, as reported in Table 5. We found that GSMaP-simulated water heights underestimated the measured WL in all stations. We selected GPM IMERG precipitation as the nowcast precipitation product. For further improvements of the results of the base modeling framework simulation, we proposed an error correction method described in Equations (1) and (2) and then simulated the framework in forecast mode. In this case, GFS forecast precipitation along with nowcast GPM-IMERG precipitation was used to simulate SWAT and HEC-RAS models and generated forecasted inundation maps.

The HEC-RAS model was optimized using different spatiotemporal combinations. Various simulation time steps (5 minutes, 10 minutes, 15 minutes, and 30 minutes) were used along with different spatial grid resolutions (250 m, 500 m, and 1,000 m) to perform the unsteady simulation of the model for the pre-monsoon season of 2017. The comparisons of simulation time for different combinations are shown in Figure 8. From Figure 8, we observed the unnoticeable sensitivity of the framework to the simulation time step. A likely reason is that final results are accumulated over the hourly and daily scale, which minimizes the error associated with simulation time step at the minute scale. On the contrary, the simulation grid size greatly influenced the computational time of the model. It took less than 5 minutes to simulate the model at 1,000 m grid resolution (for any time step), whereas computation time was more than 1 hour for 250 m resolution. Root mean square error (RMSE) and correlation coefficient comparison of different simulation grid sizes are also shown in Figures 9 and 10, respectively. In the case of the finest spatial grid size (e.g., 250 m), the hydrodynamic model was found to be unstable, and there was a large variation in the simulated water stage in some stations. Sensitivity of simulated results was also dependent on the location of the grid cells. Some of the stations are located on river banks that lie on the boundary between two adjacent cells, which can have highly varying modeled water stage (e.g., Nakuagaon, Amalshid station for 250 m spatial resolution). On the other hand, the 1,000 m grid resolution yielded very unrealistic results at most of the stations. The river channels in the floodplain area are not large enough to be covered by a single coarser simulation grid cell. Sudden elevation drops from neighboring grids at 1,000 m resolution gives rise to uncertainty. Based on the required simulation time and obtained accuracy for different grid sizes, the 250 m simulation grid size was chosen as the optimal grid resolution. Using optimized simulation time step and spatial grid resolution based on year 2017 simulation, historical simulation of the framework was also performed for the pre-monsoon period of years 2016 and 2018.

In both years of validation (2016 and 2018) of the HEC-RAS model, all stations yielded better results compared to the year 2017. In Figure 11(a) and 11(b), simulated forecasts for different lead-times with the observed data are shown. The RMSE and correlation coefficient for the period of March–May are shown in Figure 11(c) and 11(d).

Another plot is shown in Figure 12 for the average of model performance vs lead-time to illustrate the framework's predictive skill. Here, the average RMSE and correlation coefficient of the stations are taken for different lead-time and plotted against the lead-time. From the figure, it can be seen that average RMSE of the stations is less than 0.5 m for 3-day lead-time with a correlation coefficient of more than 0.75.

The proposed framework was designed to produce discharge/water level forecast as well as inundation forecasts. Sentinel-1 (SAR)-based imagery product was used to assess the model-generated forecasts. The assessment was performed in three modes: (1) historical analysis of the framework for the years 2016–2018, (2) single extreme episode-based analysis for March 2017 flood, and (3) performance in an operational setting. In the following sections, these modes of assessment are described.

Model-generated inundation forecasts of pre-monsoon seasons of 2016, 2017, and 2018 were compared with Sentinel-1 SAR imagery-derived inundation. Google Earth Engine (GEE) was used to extract inundation from Sentinel-1 gridded SAR data. A total of 69 Sentinel-1 images during the model simulation were considered. From the analysis, we found that the probability of detection (POD) of non-water and POD of water to be more than 60% for up to 5-day lead-time. Here, POD was expressed as a percentage so that the model correctly identified water and non-water features compared to the selected Sentinel imageries. In Figure 13, the POD comparison of water and non-water features for different lead-times is shown. In Figure 14, the spatial POD map of water and non-water for 1-day lead-time and 5-day lead-time is shown. Some issues were observed during the water detection and non-water feature detection by the model. For example, in some locations identified as non-water by Sentinel-1 SAR data, the model erroneously identified them as inundated. A possible reason for these errors is due to the age of the DEM used. SRTM was created more than 18 years ago and is not up to date with recent human developments and modifications of the floodplain in the region.

The proposed framework was also tested for a single event when a major flash flood occurred in the region. For the event of March 2017 flood, five Sentinel-1 images were considered separately to illustrate the performance during a single extreme event. Two images from before the flood event (21 March 2017 and 26 March 2017), one image during the flood event (31 March 2017), and two images from after the flood event (2 April 2017 and 7 April 2017) were considered.

We observed that model-simulated inundation overestimated the Sentinel-1-based water area. During the dry period, the model overestimated the SAR-based inundation. It is possible that Sentinel-1, which employs a C-band radar, may not be the best choice for inundation mapping in the Haor area where flooded vegetation is widespread. Also, the difference between the actual acquisition time of the imagery and the simulated inundation maps could be another potential reason. In Figure 15, POD of water and non-water comparison is shown for different imagery dates and lead-times.

Upon making the framework operational, there was one flood peak observed on 4 March 2019 that was assessed via the user interface at http://depts.washington.edu/saswe/flashflood. To quantify the accuracy of the system in an operational setting, the inundation forecast was compared with Sentinel-1 SAR-derived inundation. The SAR imagery acquisition date was 4 March 2019 11:47 a.m. UTC. For the purpose of comparison, simulated forecast inundation of 4 March 2019 12:00 p.m. UTC was used. The overall accuracy showed 60% probability of detection compared with reference even after 4-day lead-time. In Figure 16, the derived SAR imagery-based water inundation is shown along with the model-simulated inundation for comparison. For locations close to Khaliajuri and Markuli stations, SAR imagery indicated inundation which was accurately detected by the forecasting system. Recent structural measures were taken to improve flood management through submersible embankments in that region, which are not reflected in the SRTM DEM. This limitation in the DEM propagates to the modeling framework at those locations.

The performance of the framework was also examined against available fine resolution satellite imagery. Two planet imageries (15 May 2019 and 26 August 2019) with 10 m spatial resolution were downloaded and classified using the maximum likelihood method. They were compared with the framework-generated forecast inundation and the results are shown in Figure 17. In the left panel of Figure 17, a comparison map of the model-simulated inundation forecast of 3 days lead-time and planet imagery of 15 May 2019 is shown. From the left panel, it can be seen that the modeling framework was able to identify most of the water and non-water features. It is true that in some regions the framework failed to identify actual features. This happened because of the recent (post-2017) man-made topographical alterations after the SRTM DEM acquisition time in 2000. Framework-forecasted inundation at different lead-times was compared with the two planet imageries and the results are presented in the right panel of the figure. From the classification accuracy metrics (i.e., overall accuracy, omission error of water, and commission error of water) comparison, it can be seen that overall the accuracy of the framework was more than 70% for up to 5-day lead-time. Overall, the framework was able to identify water and non-water features with more than 70% overall accuracy, having less than 25% omission error and less than 30% commission error of water features.

The modeling framework was implemented in 2019 pre-monsoon season for experimental prototyping and then made fully operational in the BWDB stakeholder environment during the pre-monsoon season of 2020. The accuracy of the framework was analyzed and reported in Figure 18. It can be seen that the framework performance significantly improved during the year of 2020. This happened due to implementation of the Weather Research and Forecast (WRF) downscaling, which downscaled the GFS forecast precipitation to a higher spatial resolution. The live test of the framework demonstrated an accuracy of average RMSE of less than 0.5 m and correlation coefficient of more than 0.80 for up to 3-day lead-time. Based on feedback shared by BWDB (co-authors to this paper), such accuracy resulted in real-world decision-making and uptake of our forecasts for societal dissemination.

In this study, a flash flood early warning framework for northeastern Bangladesh was developed using both hydrologic and hydrodynamic components of the flash flood generation mechanism. First, the most appropriate combination of topographic and hydrometeorological observations was selected for the base modeling framework. Global numerical weather prediction model forecasts were directly forced into the framework without any spatial downscaling to reduce the computation time. Finally, we analyzed the historical performance to illustrate the applicability of the framework in an operational environment. Based on the performance of the forecasts in the reference stations, we found that the framework was able to produce a skillful flash flood early warning for up to 5-day lead-time.

In addition to the computation time and forecast accuracy considerations, some other key features of the framework considered during development were: (i) economic constraints and (ii) scalability of the framework. The framework was built using open-source models and publicly available datasets, which is essential for economically constrained countries. In this approach, ground-based monitoring was not required; thus the method can be easily applied in ungauged or transboundary basins with inaccessible areas. The concept can be implemented with minimized computational resources and limited internet bandwidth, resulting in greater applicability in developing regions. Finally, all of the datasets used in this study are of global coverage with near-real-time availability. This implies that the same concept can be implemented anywhere in the world with similar topographic and hydrometeorological conditions. Even in regions where other more expensive methods are feasible, the system introduced here can be used to quickly trigger decisions when resources are limited and time is of the essence.

This study represents a critical step towards more accurate and more accessible flash flood forecasting systems. Potential future improvements to the framework are as follows:

• Improvement of quality of NWP forecasts. This will lead to better results from hydrological and hydrodynamic modeling. WRF downscaling can be tested to illustrate the relative improvements from the existing methods.

• Explore different correction methods of QPF (i.e., climatology-based QPF correction by Sikder & Hossain 2019). This can also be implemented to improve quality without compromising the computation time.

• Develop a more representative and recent topographically surveyed DEM. A more recent elevation dataset that better captures up to date human developments and modifications of the floodplain could significantly improve the performance of the hydrodynamic model and lead to more accurate inundation forecasts.

• Ensemble forecasts in place of QPF can also be tested in this framework to produce forecast probability range instead of single time series.

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