Enhanced flood risk with 1.5 °C global warming in the Ganges–Brahmaputra–Meghna basin

Simulations were used from the HAPPI project: atmosphere-only climate simulations, forced by sea-surface temperatures (SSTs), sea-ice concentration (SIC) and green-house gas concentrations. SSTs and SIC were from the OSTIA observational dataset [26] for current day (Hist) simulations (2006–2015), and SSTs from the Coupled Model Intercomparison Project Phase 5 (CMIP5 [27]) output were used to estimate the future scenarios corresponding to 1.5 °C and 2 °C global warming above pre-industrial conditions, at the end of the 21st century [10]. Large ensembles were produced by running simulations with different initial condition perturbations. Seven models were used for this analysis (table S2), and most of the models had around 100 simulations or more for each of the scenarios (table S3).

Simulations using the CESM-CAM5 coupled climate model were designed using specific GHG concentration pathways, to stabilize temperatures at 1.5 °C and 2 °C global warming above pre-industrial conditions by 2100 [11]. These simulations cover 2006–2100, and are continuation runs of 11 CESM-CAM5 historical simulations (1920–2005) [28] run as per the CMIP5 design [27]. For current day climate, the 2006–2015 decade from the 2 °C simulations was analyzed, to match the time period of the HAPPI current day simulations. The 2090–2099 decade was analyzed for the future scenarios.

Area averages over the GBM river basin were calculated from climate model outputs. The basin definitions were identified based on the HydroBasins dataset [29]. For each model, years from different ensemble members were pooled for analysis, resulting in a large number of years representing the historical, 1.5 °C and 2 °C worlds in each model (10 years per simulation multiplied by number of ensemble members as per table S3). Values such as ensemble means were calculated across the distributions of years and ensemble members. Observational datasets were used for model evaluation (see table S4).

We primarily analyzed the yearly maximum of monthly rainfall (RXmonthly). This is because the characteristic duration of rainfall event resulting in the largest flooding events in the GBM is on the order of a month or longer, so this variable was chosen as a proxy for the change in river discharge in the GBM (see section 2.5). We also looked at yearly average precipitation and yearly maximum of 5 day mean rainfall (RXx5day), which is a standard climate change index defined by the Expert Team on Climate Change Detection and Indices [30, 31]. RXx5day represents extreme rainfall connected to flooding in small catchments or tributaries. We validated the seasonal cycle of precipitation and monsoon winds in the region for each of the models.

The climate models have precipitation biases, (compare 'OBS/Reanalysis' with 'Hist' in figures 1(a) and (b)), with the majority of models over-predicting the peak rainfall. Some of the models also tend to simulate an early monsoon onset compared to observed (figure 2). We additionally note that the observational datasets have limitations. There are differences between precipitation observation datasets (figures 1(a) and (b)), and it has also been suggested that high-altitude precipitation (such as over the Himalayas), may be significantly underestimated in observations, due to poor coverage of stations and underestimation of solid precipitation [32, 33].

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Our projection of future changes in flooding is based on precipitation changes in the climate models, rather than the absolute conditions in those models. The response of precipitation to climate change predicted by the models is based on physical mechanisms, which may represent the relative increase or decrease in precipitation even in the presence of biases. We concentrate our analysis on the basin scale, to reduce the influence of small scale effects, which the models have difficulty representing. Bias-correction, can be applied to remove biases in the mean (and variability), but should leave the relative change in precipitation unchanged. However, this is not always the case for changes in extremes [34], so bias-correction methods need to be applied with caution. We analyze scaling factors based on the relative change in precipitation, and do not apply bias correction for this study.

As the Asian monsoon drives the majority of precipitation over the GBM region, it is important that the climate models represent the monsoon circulation. Separate analysis including most of the HAPPI models used here has been made for the Asian monsoon precipitation [35, 36] and specifically for monsoon onset and length [37]. These studies determined that the HAPPI models capture the Asian monsoon circulation sufficiently to investigate future changes in precipitation. Reference [35] showed an increase in both intensity and frequency of extreme precipitation in the region, and increases in particularly damaging persistent rainfall extremes in northern India. An evaluation of the large scale 850hPa winds for each model, and its change between scenarios, is shown in figure 3. The models have varying biases in the monsoon winds, although MIROC5, with the highest precipitation, has a notable strengthening of the monsoon winds relative to the ERA-Interim reanalysis. The patterns of change in the monsoon circulation vary between between 1.5 °C—'Hist', and 2 °C–1.5 °C, and vary between models, which is consistent with [36], who additionally concludes that the precipitation change is dominated by the thermodynamic response and changes related to circulation are more uncertain.

Zoom In Zoom Out Reset image size

Flood hazard was estimated by the use of the Bristol global flood model [38]. In this modeling framework, calculations of flood extent are performed with an implementation of the well-known LISFLOOD-FP flood inundation model [39]. LISFLOOD-FP is a hydraulic model, solving the 2-dimensional shallow water equations. This configuration of the model uses a recently published bare-earth version of the Shuttle Radar Topography Mission (SRTM) global elevation database (MERIT DEM [40]), and global river and catchment hydrography from HydroSHEDS [29] to determine catchment areas and channel locations. It applies a Regional Flood Frequency Analyses (FFA) [41] using global data for river discharge (Global Runoff Data Centre (GRDC) dataset) and rainfall. In this approach, river hydrographs for locations not included in the GRDC dataset were estimated based on distributions from rivers with similar characteristics. For this study, the FFA based on global data was adjusted using gauge data from the Ganges and Brahmaputra rivers to represent the local discharge more accurately.

A limitation of the MERIT DEM used by the hydraulic model, is that it does not include flood defenses. For example the western banks of the Brahmaputra are protected by embankments from Chilmari to Sirajganj, so flood extents simulated by the model will differ with observations in these areas.

This model has previously undergone extensive validation for catchments in the UK and Canada [38], and in the USA [42]. They found that the model performance approaches the skill expected by models built with high quality local data, and that the model performs better for wet regions and rivers with larger catchments [42]. The improved MERIT DEM dataset used in this study may also improve the model performance (compared to SRTM topography used in [38] and [42]). We evaluate the performance of the model along a stretch of the Brahmaputra river in section 3.2.

For this paper, new simulations were performed over the region 21–31N, 84–94E. The hydraulic model simulated flood inundation at 30 arc second (∼900 m at the equator) resolution, which was downscaled to the MERIT DEM at 3 arc second resolution. The model simulates fluvial flooding in catchments above ∼50 km². For this study, flood hazard for return periods of 1 in 5, 1 in 20 and 1 in 100 years were calculated. This covers a large portion of the GBM basin, including the whole of Bangladesh. Firstly a 'baseline' simulation was calculated using observed distributions of current river discharge and rainfall, and secondly simulations were run with scaled river discharge to represent changes due to global warming. Changes in flood area and flood depth, between the baseline and future scenarios, were analyzed over the region and additionally over the sub-basins Ganges, Brahmaputra and Meghna (regions shown in figure S1 is available online at stacks.iop.org/ERL/14/074031/mmedia).

To determine scaling factors based on the global warming scenarios, the change in RXmonthly was used as a proxy for the change in river discharge. This was chosen as we are interested in the peak flows, and in the GBM system the highest flood waters due to the monsoon rains build up over a period of at least a month. This metric represents the change in precipitation driving flooding events, however we note that this does not take into account, evaporation or catchment hydrology, which may cause discharge to scale differently to the change in precipitation. This may also vary by return period, for example table 3 in [20], however these numbers have large uncertainty ranges (e.g. figure 5 in [22]). Furthermore, a comparison of precipitation and runoff for models used in this study, show consistent changes between these two variables (figure S3). As discharge results from the accumulation of runoff, this gives us confidence in our proxy as an approximation of future change. The advantage of this simple approach is that it does not rely on potentially problematic bias correction methods, or hydrological modeling that may be under-constrained due to sparse observations and hence introduce a greater level of uncertainty.

We determined the percentage change in RXmonthly, between the current day climate and 1.5 °C and 2 °C worlds, averaged over the GBM basin. The ensemble mean was calculated separately for each of the 8 models considered (7 HAPPI models and CESM). A weighted average taking into account the sampling uncertainty and model spread was used to produce a best estimate (referred to as the Multi-Model Summary, see text S3). The best estimates were 7.0% ± 3.6% and 10.7% ± 4.7%. The best (medium) estimates along with the upper (high) and lower bound were used to scale the baseline/present day precipitation and river discharge in the hydraulic model for the future scenarios.

On the 16th of August, the Brahmaputra recorded a record high level of 20.84 m at Bahadurabad. However, an analysis of river discharge at this station shows that despite being a record river level for this particular gauging site, the discharge for this event (78 500 m³ s⁻¹, measured on the 17th of August) was only a 1 in 5 year return period flow (95% confidence interval 3–9 years).

Other than the uncertainty in discharge measurements, the discrepancy between the return periods for river level and discharge at Bahadurabad is most likely explained by very local changes in the river channels from sedimentation and construction of flood embankments. The relationship between discharge and river height is a local effect around the gauge as a result of erosion and deposition as mobile sediment waves move through the system. However at the reach scale that we consider here these local variations will cancel out and our overall estimates of the impact of increasing flows on inundation extent will be reasonable ones. The discharge return period for this event is used for the following comparison, as discharge is used to drive the hydraulic model.

Because ground based measurements of flood extent do not exist at the resolution of the flood model, we compared flood extent from the model with estimates from two satellite products: (1) Copernicus Sentinel-1 Synthetic Aperture Radar (SAR) data and (2) the Joint Research Centre Global Surface Water (GSW) dataset [46], which is produced from Landsat imagery. The Sentinel data at ∼10 m resolution was processed and water bodies were detected based on the backscatter amplitude (supplementary text S1). SAR products can penetrate clouds, so the Sentinel-1 data can give a snapshot of inundation extent. The GSW dataset provides a flood recurrence product at ∼25 m, based on many images over the 1984–2015 period.

The 1 in 5 year modeled hazard was compared over a region downstream of the Bahadurabad gauge (89–90E 24.4–25.2N, figure S2). We compared this against a single Sentinel-1 image from 22 August 2017. This is not exactly like-for-like, as the flood model uses a 1 in 5 year discharge everywhere, whereas the actual flow in different river segments will be greater or lower depending on local conditions. We also compared against flood recurrence greater than 20% (5 years) in the GSW dataset.

Detection of flood extent in satellite data is not exact, both false positives and missed detection are possible. For example, in SAR data, smooth surfaces such as roads or mud flats may in some cases be classified as water, water roughened by wind may be classified as land, and flooded areas covered by vegetation may not be detected due to backscatter effects. Topographical shadow effects may cause false positives as well. Some of these effects, such as topographical shadows, can be corrected by the image processing, but this is still not a perfect representation of the ground situation. The GSW dataset will also not detect bodies of water obscured by vegetation, and persistent cloud cover may cause short duration flooding events to be missed.

Figures 4(a)–(b) compares model flooded regions with satellite data, at the satellite's resolution. The flood model captures a large proportion of the complicated braided river structure and flood plain. Figure 4(a) shows the modeled fluvial flood extent against the Sentinel data for 2017. As only ∼7% of the catchment areas reside in Bangladesh, we expect the fluvial flooding to drive the majority of the flooding. On the flood plain to the east of the river, the modeled fluvial flooding under-predicts the Sentinel flood extent (figure 4(a)). Figure 4(b) shows the modeled fluvial flooding against the GSW flooding. The model shows better agreement with the GSW dataset than the Sentinel image. This may be either to do with the nature of the different instrumentation and processing picking up different types of flooding or because the the GSW data is a recurrence product which more closely represents the 1 in 5 year maximum extent as the model does, compared to the Sentinel data which is a snapshot of flooding on 22 August 2017.

Zoom In Zoom Out Reset image size

In the region to the west of the river, the model is over-predicting the observed event. This is probably because the western bank of the Brahmaputra river is protected by an embankment that is not represented in the MERIT DEM and therefore not in the model. In addition, the hydraulic model simulates a 1 in 5 discharge in all river segments in the domain, so this comparison is less valid in other parts of our domain which may have a different return period for 2017.

Metrics representing this comparison were calculated over this region. Using the Critical Success Index (see [42], supplementary text S2) the model fluvial flood extent has a score of 0.47 for GSW and 0.36 for Sentinel. This shows lower performance than shown in [42] and [38]. As matching the higher resolution satellite data is a very difficult test we also calculated the absolute fractional error between data aggregated to larger scales (figure 4(d)), with errors of 0.22 (Sentinel) and 0.16 (GSW) at 100 m, reducing to 0.13 and 0.15 respectively at 5 km.

Given this < 15% error in fractional flooded area at 5 km, we conclude that the model is fit for the the sub-basin scale relative change analysis which we perform here. The hydraulic model is a physically based model, and is mass and momentum conserving. So the evaluation of the 1 in 5 year event gives us confidence that the relevant physics and the topography is well represented by the model and a greater return period of river flow will result in a realistic flood inundation for the 1 in 20 and 1 in 100 year events.

Changes to GBM precipitation are shown for RXmonthly in figure 1(c). All except for one of the models shows a wetter climate in the future scenarios, and in the 2 °C simulations, greater than two-thirds of years are wetter than the mean in the current climate, for all except two of the models. ECHAM6-3-LR shows a small drying to 1.5 °C, but an increase between 1.5 °C and 2 °C. Similar but slightly higher changes are seen for RXx5day (figure S4) and lower changes for yearly precipitation (figure S5), showing a greater wetting change for shorter duration events. Figure 1(c) gives estimates for the ensemble mean change, which in general may differ from the changes of higher return period events. However, in this study, changes for different return periods are consistent within the sampling uncertainty and there is no systematic trend of higher or lower scaling of precipitation at higher return periods (figure S6). We additionally note that using the ensemble mean may result in underestimating the uncertainty as the sampling uncertainty increases for higher return periods.

In all of the models apart from CAM4-2degree, the 1.5 °C—'Hist' change is greater than the 2 °C–1.5 °C change. This is partly because the global warming from historical to 1.5 °C is greater (∼0.6 °C) than between 1.5 °C and 2 °C (0.5 °C). Comparing the RXmonthly change per degree of warming (figure S7), half of the models still show a greater change for 1.5 °C—'Hist', however the other models show no-change or a smaller change, so there are nonlinear changes between scenarios which vary between the models. These differences may be due to the removal of suppressive effect of rainfall between 'Hist' and 1.5 °C, and varying representations of aerosols in the model. The patterns of circulation changes also differ between 1.5 °C—'Hist' and 2 °C–1.5 °C in all of the models (figure 3), so different mechanisms may be dominating in the different models.

Flood hazard maps were produced using the hydraulic model for the baseline period and low, medium and high estimates for the 1.5 °C and 2 °C scenarios. The changes in flooded area over the region are shown in figures 5(a) and (b) for the medium estimate of the 1.5 °C scenario for 1 in 5 and 1 in 100 year flood hazards. A zoomed in section shows the area around Dhaka. The 1 in 5 year hazard has a much smaller flood extent than the 1 in 100 year hazard, but crucially the change in additional area flooded between 1.5 °C and the baseline (red regions) is considerably larger for the 1 in 5 year hazard than the 1 in 100 year hazard, highlighting the importance of changes in these frequently occurring events. The changes to the depth of flood waters for the same simulations are shown in figure 5(c) and (d) with large areas increasing flood depth by 20 cm and smaller areas increasing by 50 cm or more. The medium 2 °C scenario shows incrementally larger changes compared to the 1.5 °C scenario, in both flood area and flood depth (figure S8).

Zoom In Zoom Out Reset image size

The changes in flood extent were aggregated over three sub-regions representing the Ganges, Brahmaputra and Meghna basins. These are shown in figure 6 for each of the scenarios and 1 in 5, 1 in 20 and 1 in 100 year flood hazard. The more frequent (less severe) flood hazards, show greater percentage increases in flood extent than the more extreme flooding events. This may be expected by the nature of river valleys as lower, flatter areas will flood in the less severe events. However, when the flood waters reach steeper areas in more extreme events, the relative increase in area will be less for a given change in water level. The relative change in the 1 in 5 year flood area (figure 6, blue and red bars) is greater than the corresponding relative precipitation change (blue and red shading). There are small differences between the basins, but the relative change in flood area consistently decreases for the 1 in 20 and 1 in 100 year floods. The 1 in 100 year flood area in the Ganges and Brahmaputra rivers show a smaller change than the relative precipitation change.

Zoom In Zoom Out Reset image size

Figure 6 shows the percentage area relative to the baseline flood increase, however the absolute change in area does not show a consistent trend with the return period of the event (figure S9). We also note that the 1 in 100 year events experience a greater change in flood depth, so these particularly extreme events may become more destructive due to higher flood waters, even without the flooded area increasing dramatically.