The VMD covers an area of approximately 40 500 km2 in the south of Vietnam, where the Mekong River discharges into the South China Sea through a number of estuary branches. The landscape is dominated by flat floodplains formed by deposited river sediments. Floodplain sedimentation is estimated to approximately 9.5 mmyr-1 on average (Manh et al., 2013). The properties of the deposits and the associated nutrients in combination with the tropical climate form ideal conditions for high agricultural productivity. Official statistics indicate that over 64 % of the delta (2.6 million ha) is used for agriculture, with rice as the dominant crop (three-quarters of the total cultivation land), followed by orchard farms and sugar cane (GSO, 2015). Covering only 12 % of the total land area of Vietnam, the delta contributes 52 % to the national food production and over 80 % to the Vietnamese rice export (GSO, 2015).

From July to December, high discharge of the Mekong River triggered by the Asian monsoons cause a large-scale inundation in the delta. Our study area, referred to as the delta flood-prone region in Vietnam, comprises 2 million hectares of nine provinces. The area is commonly divided into two ecological regions on the basis of inundation depth, named “deep inundation” (above 1.5 m) and “shallow inundation” (below 1.5 m). The deep-inundation areas (marked in yellow in Fig. 1) encompass the two most important floodplains in the delta, i.e. the Plain of Reeds (PoR), and the Long Xuyen Quadrangle (LXQ).

Flooding in the VMD is characterized by slowly rising and receding rates, with a mean value of 5–10 cmday-1. The flood hydrograph at Tan Chau and Chau Doc (blue points 1 and 2 in Fig. 1) usually has two peaks. The first peak normally falls from mid-July to mid-August. The second, often higher peak, arrives from September to October. Floodwater from Cambodia enters the VMD via three main routes. The mainstream branches of the Mekong, i.e. Mekong and Bassac rivers, convey 90 % of the total flood volume. The remaining 10 % are the transboundary overland flow from Cambodian lowlands to the PoR east of the Mekong River, and the LXQ west of the Bassac River (Hung et al., 2012; Tri, 2012). Besides inundation caused by high river discharge, tidal floods occur in the vicinity of rivers/canals in the coastal areas, characterized by short but repeated durations following the high tides of the South China Sea and the Gulf of Thailand (Apel et al., 2016).

Synthetic flood events were estimated for station Kratie (Fig. 1, left panel) with 10-, 20-, 50-, and 100-year return periods, referred to as T10, T20, T50, and T100. This station is commonly defined as the upstream entrance of the MD and is used as upper boundary of the hydraulic inundation model of the MD. The estimation of flood events is based on Dung et al. (2015). The authors developed and tested different bivariate copula-based statistical models on extreme values, using annual maximum discharge Qmax and flood volume FV. Both variables are important for the characterization of the long-lasting annual floods in the MD. From different models that were tested, the Gumbel–Hougaard copula was selected as most suitable, with log-normal distributions describing the marginals of Qmax and FV. The outcomes of the mentioned study is the very first publication on flood frequency analysis for the MD, considering both peak discharge and flood volume. We refer readers to the original paper for a detailed description.

Four pairs of peak discharge (Qmax) and volume at Kratie (FV), corresponding to T10, T20, T50 and T100 floods were selected from the bivariate copula model. The most probable pairs were selected from the Qmax/FV pairs with equal joint probabilities corresponding to the return periods specified above. A full probabilistic analysis using a large number of Qmax/FV pairs with equal joint probabilities was not performed due to the high computational demand of the large-scale hydraulic model (on average, 2–3 h are required for one simulation of the whole flood season June–November on a PC installed with Intel i7-CPU 3.0 GHz, 16 GB RAM). The selected Qmax and FV values range from 56 500 to 66 000 m3s-1 and from 459 to 525 km3, from the different return periods.

Damages to agriculture crops are highly dependent on the time of occurrence of flooding (Penning-Rowsell et al., 2003; Förster et al., 2008; Klaus et al., 2016). To account for the timing, each of the four Qmax/FV pairs were scaled to synthetic flood hydrographs covering the whole flood season from 1 June to 30 November using four typical hydrograph shapes (shpi, i=1–4), representing different flood patterns and, consequently, different possible damages. These shapes were adapted from Dung et al. (2015) (see Fig. 4b). The shape shp3 has the highest possibility of occurrence (p=0.34). The flood in 2000 closely followed this shape, with a minor alteration as the first peak arrived some days earlier. The other three hydrograph shapes have equal probabilities of occurrence (p=0.21–0.23). The shape shp2 has an early flood peak, while shp1 shows a late peak. The disastrous flood in 2011 resembled shp1.

This procedure results in 16 synthetic discharge time series at Kratie. They serve as the upper boundary condition for the flood propagation model. In each simulation, the lower boundaries (i.e. tidal levels), dyke scenarios and operation schemes of flood control structures were fixed as recorded in 2011, i.e. the most recent damaging flood. The scenarios are denoted using the return period and the hydrograph shape. For example, scenario T100shp3 corresponds to the 100-year return period of Qmax/FV and the hydrograph shp3.

To transform the discharge series into spatially distributed inundation water levels and associated timing in the VMD, a quasi-2-D, large-scale hydraulic model was used. The model domain covers the entire MD, including the VMD and Cambodia lowlands and the Tonle Sap Lake. It uses Kratie as the upper boundary condition and the tidal level monitoring gauges along the South China Sea and the Gulf of Thailand as downstream boundaries.

The model was initially developed by Dung et al. (2011), using the 1-D river model modelling package MIKE11-HD developed by Danish Hydraulic Institute (DHI). The hydrodynamic module (HD) provides the full dynamic solution of the 1-D Saint-Venant equations. The solution is based on an implicit finite difference scheme developed by Abbott and Ionescu (1967). Floodplain inundation (2-D flow) was presented by the 1-D model through wide cross sections for the Cambodian part of the model domain, which is appropriate for the comparatively low anthropogenic impacts on the channel network, and natural inundation dynamics of in this part of the MD. In the Vietnamese part of the delta, flood compartments were represented by virtual canals and control structures presenting the dykes. The original model was later refined and updated by Manh et al. (2014) and Triet et al. (2017) (see Fig. 5). The model was calibrated and validated with gauged data and maximum inundation extents derived from satellite data for a number of flood events, including the moderate flood of 2009, the low flood of 2010, and the extreme floods in 2011 and 2000. For a detailed description of the 2-D model see the original paper by Dung et al. (2011).

The hydraulic simulations provided discharge and water level time series at the model calculation nodes. Inverse distance weighting (IDW) was applied to interpolate between nodes for complete spatial coverage. Inundation extent and depth were then obtained by intersecting the water levels with the lidar-based DEM.

Based on the rice cropping system and cropping calendar, three gridded inundation maps were produced for each simulation scenario. The first map, labelled “July”, was generated using the maximum water level of 61 days from the beginning of simulation from 1 June to 31 July. It was applied to calculate damage to the summer–autumn crop (SAC) in both shallow- and deep-submergence regions. The second map (August) was based on the maximum water level from 1 June to 15 September in order to calculate damage to the autumn–winter crop (AWC) in the shallow-submergence region. For the AWC in the deep-submergence region, we used the map defined by the maximum annual submergence (annual) within 1 September to 30 November. The inundation grid cells of these maps were classified into three inundation depth classes to assign a stage-damage curve (Fig. 4c) to each grid cell. Exposed areas of rice crops were calculated by intersecting the inundation maps with the two land-use maps presenting the SAC and the AWC, as summarized in Table 1.

All these procedures were performed using Python scripts (Python version 2.7) and the Python-supported module of ArcGIS (ArcGIS 10.4).

Flood damage (D) was calculated on a pixel basis following equation (Eq. 2) and aggregated to damage per province and to the whole study area. D=Y×MP×A1⋅RD1+A2⋅RD2, where D is the total monetary damage (in USD). Y and MP are the average rice yield and market prices taken from official statistical data for 2011: Y=5.0 ton per hectare, and MP=USD280 per ton. A1 and A2 are the total exposed areas classified as partial and full losses. RD1 and RD2 are the relative damage factors specified on the basis of the damage curves adopted from Dutta et al. (2003) (see Fig. 4c). The percentage of damage (i.e. in terms of yield reduction) depends on the duration of contact with floodwater (in days) and inundation depth, which was classified into two groups, i.e. below 0.5 m and above 0.5 m. Because flooding in the VMD is characterized with a long duration of submergence of 2–5 months (Toan, 2014), loss factor RD2 was set to 1 for the areas where inundation depth is above 0.5 m (A2), while the partial loss factor RD1 was set to 0.5, which is the maximum damage of this class in the damage functions, for areas with inundation depth below 0.5 m (A1).

We calculated the EADs for each of the four hydrograph shapes following Apel et al. (2016). EAD is defined as the product of probability of exceedance of a given flood event and its damage: EAD=∑i=1nΔPi⋅Di¯. ΔPi and Di¯ are calculated as follows: Di¯=12(Di+Di+1)ΔPi=pi+1-pi, where ΔP is the increment of annual probability of exceedance =Δ(1-p), with p as the annual probability of non-exceedance. In this work p=0.90, 0.95, 0.98 and 0.99 according to the selected return periods; D is the calculated damage induced by the given event; i is the numerator of the probability levels considered, and n is the number of probability levels.

The average estimated annual damage (EAD‾) of the four hydrograph shapes was computed as the weighed sum of the EAD values, with the probability of occurrence of the hydrograph (Pr) as weights (Eq. 6). The average crop risk indicator was computed by dividing the average EAD by the total annual rice plantation area. EAD‾=∑i=14Pr(shp=shpi)EADi

In the final step, we investigated how flood risk will change in two land-use scenarios. Triet et al. (2017) proved that the construction of high-dyke areas in the northern delta provinces An Giang and Dong Thap increased the flood hazard in the centre of the delta. Thus, the first scenario considers the opening of the sluice gates in the high-dyke areas in these two provinces to introduce floodwater to the paddy fields during the main flood period, September–October. In response to this change in the flood management the farming system also changes: farmlands with triple-season cropping are converted to double-season cropping, i.e. no cultivation of the AWC. The second scenario considers an expansion of high dykes, i.e. an increase in the height of the existing low dykes, in these two provinces to enlarge the area with triple rice crop production. This scenario follows the development scenario, the so-called “food production scenario”, proposed in the Mekong Delta Plan (Deltares, 2013). Dyke height was increased using information of dyke elevation from neighbouring compartments and maximum water level of the historical flood in 2000, which was chosen as the design event for flood control infrastructures in the delta.

The flood propagation model simulated the 16 synthetic floods for these two dyke scenarios, while the lower boundary conditions were preserved as in 2011. Figure 6 exemplarily illustrates the simulated inundation extent for the 10-year flood for three scenarios (current situation, expansion and removal of high dykes).

The damage estimation was validated by comparing the estimated damage for the flood in 2011 with official damage data. The exposed cropping area was overestimated by 18 %, i.e. 32 500 ha in comparison with a reported area of 27 000 ha (Tinh, 2012). We estimated rice crop losses of USD 42.7 million. This number is equivalent to 81 % of the reported agricultural damages from the National Steering Committee for Flood and Storm Prevention and Control (USD 52.8 million) (MRC, 2011). Flood damages to other agriculture crops and facilities, e.g. farmhouses, which were included in the reported damages (lumped into a single value categorized as agriculture losses), were not yet incorporated in the presented damage estimation. Considering that paddy rice is the predominant crop in the delta, it is very likely to share a large part of the reported losses. Paddy fields derived from the land-use LC2014 raster account for 72 % of whole agriculture land within the focus area of this study (deep- and shallow-inundation areas). Assuming a linear distribution of damages in the lumped official reported damages with land-use proportion, it can be reasoned that the simulated damages are in the range of the reported. However, it has to be acknowledged that spatial distribution and market prices of different crops are likely to be important for the damage estimation. In any case, although not the ideal piece of information, the reported agriculture losses were the only available data with which to evaluate our rice crop damage calculation.

A large share of the overestimated exposed area of rice crops can be attributed to the simulated inundation extents, although efforts have been made to update, refine, and calibrate the model (Dung et al., 2011; Manh et al., 2014; Triet et al., 2017). The main source of uncertainty stems from the interpolation of 1-D model results to a 2-D raster, which could not be reduced even by the high-resolution lidar DEM. Triet et al. (2017) reported a flood area index (FAI) of 0.64 for the comparison of modelled and observed inundated areas for the whole VMD. The FAI was computed by dividing the sets of pixels presenting the intersection of observed and simulated inundation with the set of pixels presenting the union of observed and simulated (Eq. 3 in Aronica et al., 2002). This value increased to 0.74 only if the flood-prone area of the VMD was considered. According to Aronica et al. (2002), who suggested that a FAI higher than 0.7 is considered acceptable for an inundation simulation model, it can be concluded that the performance of the inundation model for the VMD is acceptable for the flood-prone area of the VMD, where the bulk of flood damages occur.

A small share of the overestimation of the 2011 flood might stem from the land-use data. Considering the rapid expansion of the triple-season rice areas in the delta (Le et al., 2018), it can be expected that the used land-use product of 2014 overestimates the spatial coverage of triple-season rice paddies, possibly resulting in an overestimation of the flood damage. Also, land-use data have a resolution of 250×250 m; therefore the majority of inland canals (width 10–30 m) were likely classified as rice pixels (see Fig. 3). Considered the channel density in the delta of 14 mha-1 (Hung et al., 2012), not separating these inland water pixels might contribute with a small share to the overestimation. Additionally, other important factors were not considered, such as dyke failures and local flood management measures, i.e. early harvesting of rice crops despite not being 100 % ripe or local raising of dyke segments with sandbags.

Some of these error sources might be resolved by further refining the model. For instance, the land-use data set can be improved by considering the inland canals when the crop areas are extracted. Enhancing model performance is, however, not that straightforward considering the huge amount of data required for the large-scale model domain. Despite these deficiencies, the flood damage assessment proposed in this study can produce reliable results, particularly when the typically large errors in flood damage estimation are taken as a reference (e.g. Schröter et al., 2014). Thus, the proposed method is judged to be appropriate to estimate flood hazard and risk to rice cropping in the VMD.

The flood event time series at Kratie were transformed into four hazard indicators: maximum water level, date of occurrence, inundation extent, and depth. The simulated annual maximum water levels (AMWLs) at locations with water level gauges in the VMD are summarized in Table 2. Figure 7 presents the date of occurrence (DO) of the AMWLs at these points. The chosen locations encompass nine gauges in the mainstream Mekong (i.e. Mekong River and Bassac River), and seven inland gauges in the two most important floodplains, i.e. LXQ (Long Xuyen Quadrangle) and PoR (Plain of Reeds). Inundated areas were calculated for the three periods specified in Table 1 and aggregated to the flood-prone area of the VMD on the basis of the four hydrograph shapes (Fig. 8).

The simulation results show that the AMWLs in the VMD vary substantially depending on flood magnitude (TF) and hydrograph shapes. We estimated a relative change of 10 %–20 % (40–60 cm) in simulated AMWL for an event with T100 compared to T10 in the deep-submergence region (DSR), and 4 %–8 % (5–10 cm) in the shallow-submergence region (SSR). The minor increase in the SSR can be attributed to the strong tidal influence at these stations. Hung et al. (2012) and Triet et al. (2017) reported a tidal influence of 70 %–80 % to the river flow and water level at Can Tho and My Thuan during the flood season. Towards the northern part of the VMD the tidal influence reduces to below 2 % at Tan Chau and Chau Doc at the border with Cambodia (Hung et al., 2012).

The hydrograph shape also influences the AMWL: higher AMWL were obtained for events with shp1 and shp3. The shape shp3 resulted in higher AMWL in the DSR, while shp1 yielded higher AMWL in the SSR of the delta. In contrast to this, shp2 and shp4 caused lower water levels. We found a difference in AMWL ranges from 9 cm at SSR to 45 cm at the DSR induced by the shapes shp2 and shp4 compared to the other two hydrograph shapes. These results are explained by the date of Qmax at Kratie: shp1 and shp3 have the second and higher peak in mid-September (shp3) or mid-October (shp1), about 1 month later than the other FP. When this peak is routed to the VMD, it already meets partly filled floodplain compartments, and coincides with the period of highest tidal levels in the year (October–November). Therefore, the highest AMWL is caused by the hydrodynamic interaction between the upstream and downstream boundaries, and preceding inundation dynamics. Our simulation thus provides numerical evidence to confirm the statement in Tri (2012) and Triet et al. (2017) that the superposition of river flood peaks with high tide periods results in substantial backwater effects and higher water levels up to the border with Cambodia.

The date of occurrence (DO) of the AMWL in the VMD is less sensitive to changes in flood peak discharge and shape of hydrograph than the actual AMWL (see Fig. 7). The DO can be divided into two groups. The first group is composed of stations with prevailing tidal influence, e.g. Can Tho, My Thuan, and Vi Thanh. For this group AMWL occurs in late October, similarly to the period of maximum tidal levels in 2011, the downstream boundary conditions of the flood propagation model. The second group contains gauges further north or gauges far from the main rivers, where the tidal influence is largely reduced. These gauges have the DO in the first half of October.

Figure 8 illustrates the frequency distribution of maximum inundation extent for three periods: July, August, and the whole year. This indicator varies strongly depending on the hydrograph shapes. A 10-year flood event with hydrograph shape shp1 results in the same maximum extent as the 100-year flood with shp4. This result proves the necessity to incorporate the temporal evolution of flood events into flood hazard and risk assessments in the MD. Our estimation of inundated areas from the T10 to the T100 event changed from 5 % to 10 % of the size of the flood-prone region of the VMD (2.0 million ha) in July, from 30 % to 50 % in August, and from 60 % to 75 % for the annual maximum extent. This means that even for the 100-year flood, 25 %–40 % of the flood-prone region was cut off from inundation by the implementation of high dykes, initiated after the flood in 2000.

Exposed areas and flood damage (D) to rice crop were calculated on a pixel basis and then aggregated to the eight provinces located in the study area (Fig. 9 and Table 3). The average damage ranged from USD 39.0 million for a 10-year flood to USD 75.0 million for a 100-year flood. These numbers account for 0.23 %–0.45 % of the total gross domestic product (GDP) of the eight provinces in 2011. Since such assessments are not available for neighbouring deltas in South Asia, e.g. Chao Phraya in Thailand or Irrawaddy in Myanmar, a comparison of these deltas is not possible. Our worst scenario, i.e. T100shp3, resulted in damages of USD 115.7 million. This value is less than half of the reported overall damage of the flood in 2000 (USD 250 million; reported in MRC (2012), scaled to USD 500 million with 2011 price levels by Chinh et al., 2016), which is considered a 20-year flood in the MD (Le et al., 2007). Although the damage figures from the event in 2000 were the overall damages, of which agricultural damages were an unknown part, this indicates that a large reduction in flood losses can be linked to the flood management and adaptation measures being implemented in the VMD following the Decision No. 99/1996 from the Government of Vietnam (The Government of Viet Nam, 1996). Note that the plan was initiated in 1996, but was implemented to a large extent after the flood in 2000 had occurred.

The hydrograph shape has a substantial effect on damage to rice crops in the VMD. The shape shp3 resulted in a high amount of damage, roughly 1.5 times higher than the average damage. Flood hydrograph shp2 closely matched the average damage (85 %–90 %), while shp1 and shp4 resulted in approximately 60 %–70 % of the average damage (Fig. 9b). These findings support our hazard and risk assessment approach and point to the relevance of the temporal evolution of the flood event for damage estimation. For example, the total flood damage from T10shp3 (USD 57.8 million) was about 10 % higher than the T100shp1 event (USD 51.0 million).

The results provide evidence that rice cropping in the VMD is most vulnerable to flooding stemming from the early flood pulse in August–September when the total plantation area is expected to be at its maximum. During this period damage occurs to the second crop (SAC) in both DSR and SSR, and the AWC in the SSR. This finding is in line with the extremely high damages during the flood in 2000, which had a first peak at Tan Chau (point 1 in Fig. 1) on 2 August, i.e. 1 month earlier than usual (Xo et al., 2015). In contrast, damage from late flood events, i.e. shp1 or shp4, are limited to the AWC in the DSR due to failure of dyke structure, or flood levels overtopping the current dyke height, similar to the flood damages recorded in 2011.

Aggregating the mean D values per province showed that the two northern provinces, An Giang and Dong Thap, accounted for about two-thirds of the total D of the VMD, with flood damage in An Giang being about 1.5 times higher than in Dong Thap. This difference increases up to a factor of 3 in the case of late flood peaks (Table 3). This can be explained by the larger areas with triple rice crops in An Giang. This third crop in the DSR is particularly vulnerable to inundation in October–November (see Fig. 3b). Damages for the other two provinces in the DSR also showed remarkable differences, with ∼16 % of the total flood damage calculated for Long An located in the Plain of Reeds. This is 10–15 times higher than the share of Kien Giang (D in Kien Giang is less than 2 % of the total D), although the rice cropping area is larger in Kien Giang (Bouvet and Le Toan, 2011; Nguyen et al., 2015).

The remaining SSR provinces accounted for about a quarter of the total D, which can be linked to their smaller rice cultivation areas in combination with their flood control measures. Our calculated flood damage to rice crop for Can Tho is, on average, in the range USD 2.9–6.3 million from the 10-year to the 100-year flood. These figures amount to about 60 % of the estimated urban damages of Can Tho (USD 5.0-9.7 million) for events with similar magnitudes (Chinh et al., 2017). Since Can Tho has the highest urban/rural area ratio compared to the other delta provinces, it is likely that agriculture losses have an equal or higher share in flood damages compared to urban losses in the other provinces of the delta.

For the current land use, the EAD for rice cropping in the whole delta amounts to USD 4.5 million (see Fig. 10b), with an average crop risk of USD 1.0–4.6 for each unit of land [ha] (see Fig. 11a). The highest risk was calculated for the provinces located in the deep-submergence region (DSR), except for Kien Giang, where the EAD was very low.

An expansion of areas with triple rice cropping in An Giang and Dong Thap, as defined in the expansion development scenario, would increase the exposure to flooding. Figure 10b shows that this expansion of AWC would triple the EAD to above USD 15.0 million. The expansion of triple rice cropping also means that the current low-dyke system must be raised to a higher design level in order to support the cultivation of AWC during September–November. This, in turn, would lead to higher inundation hazards and risks in downstream provinces (Triet et al., 2017). Figure 11c shows both effects: a substantial increase in EAD in Dong Thap and An Giang, and a slight increase in EAD in the downstream provinces resulting from a higher-inundation hazard caused by the high-dyke development.

The second development scenario – introducing floodwater to the paddy fields after the SAC and no cultivation of AWC in An Giang and Dong Thap – resulted in smaller D and EAD values, as expected. Abandoning AWC cultivation would result in an EAD decrease of about 40 % (Fig. 10b). The majority of these changes in both scenarios stems from changes in An Giang and Dong Thap. The other provinces account for 5 %–15 % of the changes in EAD only.

The full flood control measures (high dykes) supporting the expansion of AWC areas in An Giang and Dong Thap after the flood in 2000 have been continuously debated. Triet et al. (2017) revealed that the change from low-dyke to high-dyke systems in this upstream part of the delta increased the inundation hazard in downstream areas, e.g. by an increase of 9–13 cm in AMWL at Can Tho and My Thuan. Howie (2005) and Käkönen (2008) challenged the claim that farmers could have greater benefit by being able to add another harvest to the cropping system, because of more investment cost for mineral fertilizers to counter the losses of natural fertilization by deposited sediment (Manh et al., 2014), together with negative social and environmental consequences. The profitability of triple rice farming was reported to reduce from initially 57 % to 6 % after 15 years compared to double rice counterparts due to higher production costs (Tran et al., 2018). These arguments and findings might make the expansion of AWC less attractive.

On the other hand, rice cultivation areas in the southern part of the delta are likely to decrease due to increased salinity intrusion following higher sea levels (Smajgl et al., 2015; Hak et al., 2016) and land subsidence (Laura et al., 2014; Minderhoud et al., 2017). Expansion of AWC in the northern part of the delta is an option for countering such losses and ensuring food security. Our results could support an evaluation of the costs and benefits of further high-dyke development and triple rice cropping expansion and thus provide important information for future flood management and land-use planning in the delta. Additionally, the damage and risks maps can serve as a basis for flood management. They could support the agricultural insurance, initiated in 2011 by the decision of the Prime Minister of Vietnam. In this programme the insurance premium depends on the rice yield only, and the spatial pattern of the flood hazard is not considered (The Government of Viet Nam, 2011).

One of the major sources of uncertainties in our estimation of D and EAD is associated with the inundation maps. The process of interpolating from 1-D water levels to 2-D inundation raster inherits uncertainties from the hydraulic model and from the DEM used for interpolation (Brandt, 2016). A full 2-D modelling approach might enhance the quality of flood inundation mapping, but this comes at a cost: the set-up of a full 2-D hydraulic model on that scale is challenging because of the high density of man-made channels and hydraulic infrastructure, which need to be implemented in the model with high accuracy to substantially improve the simulations. Additionally, model runtime becomes critical in detailed large-scale 2-D hydraulic simulations within a risk assessment requiring a large number of model simulations. Large-scale approaches building on coarse-resolution modelling with sub-grid parameterization (Sampson et al., 2015) still cannot provide sufficient accuracy to properly map the hydraulic dynamics in such a complex system with flat topography, where details matter a lot. However, refining this approach in high resolution, including all relevant hydraulic structures, and implementation in a highly parallelized environment (e.g. on GPUs) could provide a viable path for reducing the uncertainties in hydraulic modelling of the MD.

Another uncertainty source is the land-use maps used to quantify flood exposure of rice crops. The land-use raster was produced using satellite data from 2014; thus it is somewhat outdated considering the dynamics of agricultural land-use change in the MD. The area of triple-season rice has very likely increased from 2014 to the present. An updated and higher-resolution land-use data set would certainly provide more up-to-date results.

Another uncertainty source is the limited number of return periods used to calculate EAD. Ward et al. (2011) showed that the number and choice of the selected return periods can introduce a significant bias in the EAD estimates. They also pointed out the importance of considering damages to frequently occurring low damage floods, in line with the findings of Merz et al. (2009). However, as the agricultural system in the MD is well adapted to frequently occurring floods (in fact these floods are the basis for the current practice of paddy rice cropping), this effect is likely not as important as in the European studies listed above.

Our work does not consider flood damage to other agriculture crops (e.g. orchard farms) or land-use types (e.g. shrimp farms). Although these production types are smaller than rice cropping in terms of area, they generate much higher economic value per cultivation unit. Therefore, it is highly recommended to include these crops in future studies on agricultural flood risk in the VMD. To facilitate such assessments, efforts need to be made to collect data on crop and aquaculture type and area, plantation calendar, and their vulnerability to inundation. Such risk assessments could then be used for scenario planning by varying land-use types and cropping patterns or changing boundary conditions by climate change and upstream developments. In addition, dedicated efforts for validation and development of crop damage curves for the VMD based on recorded flood damages at the plot scale would increase the credibility of the presented risk analysis. Our study is validated against large scale, aggregated damage data only, and it is open to ensuring that small-scale variations, which could be important for local adaptation measures, are represented sufficiently well.