Managing river basins and environmental systems in a sustainable manner is receiving growing attention from national water resources institutes, the United Nations, non-governmental-organizations, and international research institutes. The newly adopted Sustainable Development Goals (SDGs) prescribe key hydrological, environmental, and economical processes to be expressed in terms of performance indicators. Water accounting systems are currently under development to facilitate the mapping and description of these SDG indicators at river basin scale [1
]. Water availability, water consumption, utilizable water, and water withdrawals are key elements of such accounting processes, as well as the services and benefits rendered. At the global scale, 60% of ET is from green water (precipitation stored in soil moisture), the rest being withdrawals from blue water sources (rivers, reservoirs, lakes, and aquifers) [5
]. Eco-hydrological modelling tools have been developed to quantify a wide range of natural ecosystem services as well as human intervention derived from these significant volumes of water [6
]. A comparison of different hydrological models that are suitable for modelling hydrological ecosystem services was conducted in [9
], among them are the Soil Water Assessment Tool (SWAT) [10
], Variable Infiltration Capacity VIC [11
], Integrated Valuation of Ecosystem Services and Trade-offs INVEST [12
], and ARtificial Intelligence for Ecosystem Services ARIES [13
]. SWAT was indicated as a preferred tool in a rigorous review of the modelling of certain ecosystem services [14
] for simulation of provisioning and regulating services, because hydrological, flow dynamics, water quality, plant growth and nutrient loading processes are included in the model. SWAT was also recommended as the most suitable model for long-term simulations in watersheds dominated by agricultural land uses [15
], since its original design was to assess the impact of land management practices on water, sediments, and agricultural residues. SWAT model is also preferred in studies in ungauged basins [16
Classically, the SWAT model is calibrated using a few hydro-meteorological stations e.g., [18
]. Large uncertainties in observed stream flow data are common [23
], and that more sophisticated calibration method needed to be developed. SWAT Calibration and Uncertainty Programs (SWAT-CUP) [24
] was developed for automatically computing sensitive model parameters and calibrating SWAT by means of parameter optimization. Most SWAT-CUP applications are using Sequential Uncertainty Fitting (SUFI-2) algorithms and flow observations to define the best parameter set. Additional uncertainty bounds are computed according to the good parameter sets and typical SWAT parameters are calibrated by means of parameter optimization. The current paper investigates how SWAT can be set up for assessing ecosystem services in ungauged basins using remote sensing data and auto-calibration facilities.
Several review papers on remote sensing technology for hydrology, e.g., [25
] and water management, e.g., [27
] indicate that land cover, land use, precipitation, ET, soil moisture, snow cover, and water levels can be determined from spectral radiances measured remotely. Several open access databases on precipitation have recently been developed on the basis of remote sensing data; see [29
]. A simultaneous development took place on operationalizing remote sensing-based energy balance models to accurately determine and upscale ET from local heterogeneous watersheds (e.g., [31
]) to continental scale (e.g., [32
]). Extensive reviews of remote sensing-based approaches to derive ET were carried out earlier by [33
]. Remote sensing provides a great source of data to study vegetation indices and Leaf Area Index (LAI) from multi-spectral bands [37
Some hydrological studies utilized remote sensing data before—or a combination of remote sensing and in situ data—to calibrate hydrological models, e.g., [38
]. Earlier research demonstrated the capacity to calibrate SWAT with remotely sensed ET data, e.g., [43
] and LAI [48
]. Several studies in Vietnam integrated SWAT and remote sensing data, including [21
], on the impact of climate change on stream flow in Dakbla River Basin. The objective of this Vietnamese case study is to use quasi open access remote sensing data to improve SWAT modelling performance using the standard SUFI-2 functionalities, both as forced variable, i.e., precipitation or calibration dataset, i.e., ET and LAI. The innovation is that the standard calibration module is based on remote sensing data instead of classical discharge data, and that it incorporates soil and vegetation parameters of individual Hydrological Response Units (HRU). The anticipated result of such a calibration approach is a better quantification of the natural and anthropogenic eco-hydrological processes in ungauged basins, which is vital for reporting ecosystem services to governments and the United Nations. The novelty of this study is the application of SUFI-2 in the optimization of 15 input parameters of the soil vegetation processes using observations of actual water management processes, such as irrigation and conservation of water in wetlands in a local context. Such level of detail and reflection of real world interferences of mankind on the natural hydrological cycle can never be achieved from flow measurements, and opens better opportunities for simulation of local eco-hydrological processes that occur locally in ungauged basins, which can never be interpreted from bulk flow measurements, if there are any.
2. Study Area
The Day Basin is located between 19°55′ to 21°10′ N and 105°20′ to 106°25′ E. The Day Basin is a sub-basin of the transboundary Red River basin (See Figure 1
). The total area of the basin is nearly 6300 km2
. The highest elevation is 1256 m in the western part of the basin. The Day Basin comprises several river tributaries, among which the largest is the Day River, with a total length of approximately 250 km. The Day Basin has a high biodiversity, with abundant flora and fauna in the forested hills, freshwater aquatics, and wetland. The land use is also diversified, although agricultural land use is dominant (64%).
The Day Basin encompasses the capital city of Hanoi (population in 2015: 7.5 million inhabitants) in the northeast and several major economic centers located downstream, such as Nam Dinh (population: 1.8 million) and Ninh Binh (population: 0.9 million). Both the Red River and Day Basin have been exposed to various hydrological research activities before, e.g., [8
] on hydrological and water quality modelling. Using the rating curve suggested in [50
], discharge at two locations (i.e., Ninh Binh and Phu Ly) was reconstructed from the year 2000 up to 2013.
The annual total precipitation is around 1700 mm per year, and reference evapotranspiration (ETo
) is approximately 1100 mm per year. The climate in the Day Basin has a monsoonal character. The wet season lasts from May to September, and dry season from October to April. Precipitation can reach up to 450 mm per month in some parts of the basin, and as low as a few mm during January and February (Figure 2
). Precipitation is measured at nine stations across the basin, and is available up to 2013. These measurements will be used to validate the open access precipitation product based on satellite measurements.
The water withdrawals for irrigation in the Day Basin are rather difficult to assess because various pumping stations lift water from the Red River, and also many inlets divert water from the river gravitationally. This diffusive and unmetered inter-basin water withdrawal complicates the computation of the irrigation hydrology and the water accounts related to that. The irrigation supplies in SWAT will therefore be adjusted to reproduce an ET value that matches with ET estimates from satellites.
5. SWAT Parameters to Be Optimized
A summary of various sets of parameters used in model calibration of different processes, i.e., surface runoff, snow, plant growth etc. can be found in [54
]. A similar list of parameters was also suggested by [16
]. Calibration of ET is less common, because this dataset is usually not available. More recent studies, including [8
], demonstrated that SWAT can be calibrated against spatial ET data as well. Key parameters for the calibration of ET and their value range were derived from the SWAT user’s manual [53
]. Since the purpose of this study is to calibrate the model for ET and LAI, the model parameters to be optimized were divided into two groups: ET and LAI. This grouping indicates which parameters are most sensitive to ET and LAI, and thus, suitable for optimizing the model performance for these two processes. The list is long because one parameter might control more than one process, e.g., the available water capacity of soil layers affects the generation of surface runoff, but simultaneously determines the amount of water that is evaporated (ET) and percolated to the underground. SUFI-2 automatically optimizes the selected parameters within their predefined range, hence this aspect of parameter selection is following the default guidelines of the tool.
Six parameters ESCO, EPCO, REVAPMN, SOL_K, SOL_AWC, and SOL_BD have been selected to optimize the ET simulations. The baseflow recession constant ALPHA_BF and SCS runoff curve number (CN2) were also included because of their influences on the surface–subsurface hydrological processes, and thus, on the water availability for evapotranspiration. Seven other parameters are identified that influence the leaf area index development, see [53
]: BLAI, ALAI_MIN, DLAI, LAIMX1, LAIMX2, FRGRW1, and FRGRW2. SUFI-2 thus optimizes 15 model parameters for different sub-basins and HRUs. Since SWAT-CUP is unable to read spatial data, time series of ET and LAI was derived and used as input into the model. While ET process is more dependent on the local climate and land cover properties, LAI is strictly driven by the plant type, and hence, the calibration should be done at sub-basin, land cover, and HRU levels. The utilization of high resolution spatial datasets as input into HRU analysis in SWAT (i.e., land use, DEM, soil maps) ensures fair distribution of HRUs, and hence, spatially distributed ET and LAI across the watershed.
In the case of conventional calibration using discharge, two river discharge points (i.e., Phu Ly and Ninh Binh) were used (500 simulations for both calibration and validation) and the acquired result was used to compare against the output from calibration technique using remotely sensed data.
In the first step of calibration technique using remote sensing data, average ET and LAI for 119 sub-basins and all land-cover groups was extracted from ensemble ET and Moderate Resolution Imaging Spectroradiometer (MODIS) and used as observed data in SWAT-CUP. Derived time series of ET and LAI from each sub-basin and land cover type was used to calibrate 15 parameters in each sub-basin and land cover group individually (500 simulations each run for calibration and validation). This first estimation of ET and LAI ensures a good agreement in term of average ET, LAI, and subsequently, water balance. In the second batch of calibration, a number of HRUs from each sub-basin representing all land use classes and ensuring a total coverage of at least 50% the sub-basin’s area were selected in the calibration (500 simulations each run for calibration and validation). The calibration result yields a set of soil and plant parameters which varies from sub-basins and land cover groups, and that reflects the local eco-hydrological processes, and is therefore highly valuable. This unprecedented spatial variability from SUFI-2 optimization cannot be obtained from soil surveys and soil maps, which makes this investigation interesting. The classical SUFI-2 calibration will give the same parameter value to all sub-basins and HRUs encompassed in a certain drainage area, and hence, the new system of calibration provides much more insight into the local characteristic of the soil and vegetation system.
6. Results and Discussions
The simulation covered the period 2000 to 2013 using 3 years of initialization (2000 to 2002) as modelling warm up period. Because remotely sensed data had a different temporal coverage—2003 to 2012 for ET and 2005 to 2011 for LAI—SWAT was calibrated from 2003 to 2007 for ET, and validated from 2008 to 2012. LAI was calibrated from 2005 to 2007, and validated from 2008 to 2011.
The monthly simulation for the entire Day Basin, as one bulk system, was presented in Figure 7
. The simulated values related satisfactorily to the observed ensemble ET values. The ET peak value from remote sensing did not exceed 140 mm/month, while the modelled ET was as high as 160 mm/month. For remote sensing-based calibration, simulated ET in SWAT is underestimated by 5 to 10 percent in the irrigated agricultural land downstream, while the forest land showed good correlation. NSE yielded from 0.61 (calibration) to 0.65 (validation). R2
for calibration and validation was 0.71, while Kling Gupta Efficiency (KGE) ranged from 0.80 to 0.83, respectively. The high values from KGE were due to the fact that low values were well represented for bias and correlation, as compared to NSE’s reputation on overemphasizing peak values. By contrast, traditional discharge-based calibration performed statistically more inferior with NSE, which were only at −0.01 and −0.20, and KGE at 0.46 and 0.47, while R2
was of significant values 0.77 and 0.76, for calibration and validation, respectively. Discharge-based calibration results also encountered a consistent underestimation of ET temporally as compared to remote sensing-based method. Another observation in both cases is that the lower ET values simulated during winter were always lower than the observed values from remote sensing, even though this difference was much lower when using ET and LAI as calibration inputs. This could be related to the dry period in which SWAT computes water stress due to a lack of soil moisture. This could suggest that the storage capacity of the soil in reality is higher, or it could also be related to lower vertical water fluxes between the top soil, sub-soil, and the unconfined shallow aquifer, or the sensitivity of vegetation to soil moisture. A very low reference ET0
during winter could also be an explanation. SUFI-2 ensured, however, that the spatial patterns match rather well. Some local differences occur unavoidably due to the inaccuracy of the various mathematical expressions used to compute a complex hydrological process such as ET. Remote sensing ET values reflect more the real world conditions as they are based on observations [33
]. The agreement between SWAT and remote sensing data was expressed by means of the correlation coefficient and the bias. Using spatial dataset in the calibration yields resulted in much higher agreement, and reveals that both the ET formulations in SWAT as well as the SUFI-2 optimization techniques are adequate. The same conclusion was drawn earlier by other researchers that validate SWAT on the basis of ET data series.
Further to the evapotranspiration simulations of SWAT, the vegetation response to water can be evaluated by means of a comparison of the simulated LAI (see Figure 8
) in two cases: (a) remote sensing-driven, and (b) discharge-driven calibration. The remote sensing-based calibration yielded acceptable NSE ranging from 0.50 for calibration and 0.57 for validation. R2
for calibration and validation was 0.51 and 0.59, while Kling Gupta Efficiency (KGE) ranged from 0.59 to 0.75. Conventional discharge-based calibration performed poorly, in term of statistics, with NSE of only 0.35 and 0.43, KGE of −0.34 and 0.03, while R2
was higher at 0.67 and 0.77 for calibration and validation, respectively. LAI values in discharge-based calibration were consistently underestimated compared with observation from MODIS. With regard to remote sensing-based calibration, the timing of the green cover development seems acceptable. The peak LAI values during Spring 2005 and 2006 are not accurate. This may be due to constancy of the LAI related calibration parameters for all the different simulation years. It would be better to make the maximum LAI parameter variable for each year, to enable it to better respond to years with weather anomalies. A good description of LAI evolution will improve the timing of rice emergence, which in turn, affects the irrigation and transpiration processes.
represents the accumulated ET during the calibration (2003 to 2007) and validation (2008 to 2012) period for remote sensing-based calibration. The total simulated ET is 5855 mm, compared with 5727 mm for the ensemble ET during the calibration, hence a difference of 128 mm (2.2%). The model performs consistently during the validation period with 5713 mm (simulated ET), as compared to 6015 mm (ensemble ET), leading to a difference of 302 mm (5.3%). In very general terms, the set of ET related equations in SWAT has a limited capacity to mimic the complex processes of soil evaporation, plant interception, and plant transpiration that occur in reality, due to the dynamic meteorological and hydrological processes. The results indicate that SUFI-2 succeeded in generating a close to reality ET in both calibration and validation of ET, both spatially and temporally.
shows, for every HRU, the distributed ET from the SWAT model and ensemble ET over the period 2003 to 2012. These graphs showed a spatial coherence between the two datasets.
To verify simulated streamflow from the remote sensing-driven calibration, the relationship between monthly river flow simulations and flow measurements at the Phu Ly and Ninh Binh stations was assessed, as shown in Figure 11
. The discharges in several cross sections was constructed using the rating curve of measured flows and water levels during restricted periods, measured using acoustic Doppler current profilers [51
The agreement between simulated river flow and station discharge measurements was expressed through R2
, ranging from 0.71 (Phu Ly) to 0.78 (Ninh Binh), and Nash–Sutcliffe from 0.55 to 0.63, respectively. Considering that river discharge was computed from measured water level [51
], the agreement between simulated river flow and station discharge measurements is good. Measured river flow data were not used in the calibration process, which shows that a good simulation of ET facilitates the prediction of stream flow. Note that the flow in these rivers is far from being natural, due to all the irrigation practices that are occurring. Measured river flow data were not used in the calibration process, which shows that a good simulation of ET will make it possible to calculate flows directly (streamflow and ET being the two largest components of the water balance) without having to optimize flow in the calibration process. Hence, this shows that flow data can be replaced by ET to optimize SWAT model performance, and that the ET is reflecting actual anthropogenic processes of the water cycle.
To make the fit between remotely sensed precipitation and ET feasible, extra water has to be supplied to the land use class irrigated land in certain months. The total amount of irrigation water supply is computed to be 1.934 billion m3
/y, and this amount of water is thus withdrawn from the Red River through various ungauged inlet points. This number can now be estimated with more precision because ET is known [16
]. The main water intake for irrigation is between January and March (during the Winter–Spring paddy rice) with a maximum amount of 98 mm during February. During the Summer–Autumn paddy rice season, the water intake is concentrated in July with an amount of 113 mm per month. The total storage change ∆S for the entire Day Basin indicates the difference between all inflow and outflow terms. ∆S during 2003 to 2013 for unsaturated and saturated zones are 11.7 and 9.6 mm/y, respectively, due to the fact that there is water locally stored in lakes, streams, and also moving within the saturated layers to the deep aquifer. This is a realistic number, and confirms that the water balance of the total system is rather accurately simulated.
The total water balance for the basin in average rainfall, a dry (2007), and a wet (2008) year, is shown in Table 1
. While the average annual rainfall was 1710 mm/y, it reached 2121 mm/y in 2008 and 1496 mm/y in 2007. The ET in 2007 and 2008, was 985 and 1010 mm/y, respectively, compared to 958 mm/y on average. This quasi-constancy of ET is noticed more often in other studies. The regulating role of soil water storage in the vadose zone, and the lower evaporative demand in wet years and higher demand during dry years, are some major causing factors for restriction of ET dynamics. Due to the impact of rainfall variation to overall water budget, the change in soil water content (∆SW) for 2007 was indeed negative (∆SW= −27 mm/y) while for 2008, it was positive (ΔSW = +80 mm/y). The final calibrated range of 15 parameters is expressed in Table 2
The availability of spatially distributed precipitation, ET and LAI gridded data from open access—or partially open access—earth observation data platforms, makes it feasible to calibrate soil and vegetation process parameters of eco-hydrological models, also when rivers are ungauged and the water distribution system is complex. In this study, four individual ET models were averaged linearly to match the simulations of ET from SWAT. No in situ measurements were available to verify the performance of individual ET models. From inspection of the water balance, a simple linear average value gave best results as compared to streamflow and Kc crop coefficients, described in the international literature. Nevertheless, it is required to undertake more studies on the ensemble ET product to provide further progressive insights on averaging of individual estimates.
Secondly, this paper demonstrates the capabilities of SWAT and the auto-calibration SUFI-2 to render bio-physical processes in data scarce basins in a distributed manner. A total of 15 essential bio-physical parameters of the unsaturated zone and exchange processes seem to be adequately calibrated in SWAT for each of the 7907 HRUs. Otherwise, there would not have been such a good agreement with the spatio-temporal variability of the remote sensing parameters, as in the case of conventional discharge-driven calibration. Furthermore, this level of spatial detail cannot be obtained from soil maps. The hydrological formulations in SWAT are thus adequate for simulating eco-hydrological processes. In the near future, remote sensing data on soil moisture, net primary production, and water quality will become available as well, which will undoubtedly further enrich the options to calibrate additional SWAT model parameters.
The approach proposed in this study using SWAT-CUP and SUFI-2 will improve the facilitation and standardization of calibration process for basins with scant field data. By optimizing evapotranspiration and photosynthesis for Hydrological Response Unit, swift estimates of surface runoff, erosion, groundwater recharge, baseflow, storage changes, withdrawals, and carbon assimilation can be determined and used to quantify ecosystems services. The availability of the system parameters will allow future predictions of the basin water cycle in response to external factors, such as climate and land-use changes, and computing scenarios for green growth, i.e., conservation plans, reduction of greenhouse gas emissions, etc.