Remote sensing data have been used in water resource studies to estimate rainfall, evapotranspiration (ET), soil moisture [1
], runoff, floods, droughts, snow water equivalent [2
], and irrigation [3
]. Apart from providing crucial information at the global scale, remote sensing can also be used as a proxy to the observed data in data-scarce regions. The water resources’ spatial and temporal variations are not confined to country boundaries. Therefore, it requires a broader consensus approach covering social, environmental, technical, and economic aspects for efficient management [4
]. The United Nations has identified water usage by different sectors as the basis for the complexities and the principle for the allocation of the international river waters [5
]. Despite a few studies that used hydrological and hydraulic models to understand dam characteristics, it is difficult to develop plans for the operation and maintenance of reservoirs without any real monitoring and modeling [6
Since the extent of water use is related to the availability of storage in reservoirs, estimating the change in storage over a seasonal or annual scale is critical. Moreover, reservoir storage estimation is dependent on many unknown factors, including diversion, inflow, and day-to-day operations. The complexities exacerbate for transboundary rivers because many stakeholders are involved across many countries. Cascading of dams in the main stem and tributaries as well as unmindful development within the catchment can further complicate conditions. Therefore, quantifying diversion from the reservoirs will be crucial to assess food and water security concerns of the agricultural regions.
Advances in remote sensing techniques can potentially aid in estimating crop water requirements or diversion for irrigation using the storage change estimations and inflow and outflow simulations from the hydrological model. However, accurate estimates of storage and diversion require a well-calibrated hydrological model reflecting the basin in changing conditions [11
]. Concisely, with the use of remote sensing data and good hydrological model, storage, inflow, and outflow can be estimated for the quantification of water diversion from reservoirs.
The Mekong River basin (MRB) is one of the largest transboundary basins in the world, shared by China, Myanmar, Thailand, Lao PDR, Cambodia, and Vietnam. The livelihood of more than sixty million people in the Lower Mekong Basin (LMB) is dependent on the Mekong River for agricultural activities—cultivation of rice, vegetables, livestock, fruit and nuts, and fisheries [12
]. The intensity and highly predictable occurrence time of the peak are the defining characteristics of the large tropical monsoonal river [14
], leading to a twentyfold increase in discharge between August and September. The MRB has high hydropower potential (~53,000 MW in the main stem and 35,000 MW from tributaries) [15
] and is equipped with approximately 42 dams to harness the hydropower production [16
]. There are plans to build 16 main stem and 110 tributary dams by 2030 [17
]. The underdeveloped condition of the MRB in terms of river impoundments [19
] and the sensitivity of the dam operation information are the fundamental cause of the lack of archived data sets in the region [20
]. Most of the studies are performed through hydrological models, with remote sensing data merely used as forcing inputs for simulations [8
Remote sensing plays a major role in characterizing the water bodies in remote areas where accessibility is an issue. The water level of the reservoir can be determined using the spatial extent estimated by the optical (Landsat 5 TM, Landsat 7 ETM+SLC-Off, Landsat 8 OLI-TIRS, and ASTER images) and synthetic aperture radar images (Cosmo SkyMed and TerraSAR-X) [26
]. A combination of remote sensing satellite data products such as moderate resolution imaging spectroradiometer (MODIS) and altimeters can be used for the global monitoring of large reservoirs using unsupervised classification approach [28
]. Similarly, variations in volumes of Lake Nasser and Toshka were estimated using the Landsat images and radar altimetry from TOPEX/Poseidon, Jason 1 and 2, Geostat, and Envisat [29
]. The Landsat images were also helpful in the estimation of storages in small reservoirs in the Preto River Basin, Brazil [30
]. Remote sensing was effective in estimating the reservoir volume located in the high topographic relief area in the Himalayan region in India [31
]. In addition to monitoring the volume variation, remote sensing aids in establishing guidelines for reservoir operations. The rule curve of the Fengman reservoir (China) was derived using the normalized difference water index (NDWI) from the Landsat data [32
]. The satellite imageries and water level data set from four different satellite altimetry databases—Global Reservoir and Lake Monitor (GRLM), River-Lake Hydrology (RLH), Hydroweb, and ICESat-GLAS level 2 global land surface altimetry data (ICESat-GLAS)—are available to estimate the water volume changes in lake and reservoirs [33
]. Bonnema and Hossain [36
] have used satellite observations to explore the artificial reservoirs’ operating pattern in the LMB—with two parameters affecting streamflow, residence time and flow alteration—and reported the range between 0.09 and 4.04 years and 11% and 130% of its natural variability, respectively. The satellite-based approach to assess the dam operation has been mainly confined to the United States, European, and African reservoirs.
Water diversion from the reservoirs for agriculture, domestic, and industrial purposes is also not explicitly inferred from these studies. Although the total annual runoff from the MRB is projected to increase by 21% by 2030 during the dry season and significantly during the wet season, the flow impoundment due to existing and planned dams is expected to affect the quantity and time of water availability to downstream regions [37
]. However, water partitioning between different sectors (hydropower, irrigation, domestic, industrial, navigation, and environmental) is ambiguous since there is no clear idea of demand versus availability for a given season considering the dynamics of changes in land use, basin development, and climate. Alternatively, empirical relationships can be useful to represent the dynamics of the area, water depth, and discharge at each of the reservoir locations.
Considering the global applicability of remote sensing data sets, it is ideal for any methodology to be scalable to other locations so that a standard procedure can be developed to assess water management of multipurpose reservoirs. This will help assess the need for robust observational data for water levels from a range of reservoirs to scale (repeat cycle or space/time resolution). Water diversion for irrigation, domestic, and industrial purposes will also enhance because of a large number of dams and will alter downstream flows. Therefore, it is important to assess the dams’ cumulative impacts as it can directly alter the flows, storage, and diversions.
To meet the increasing water demand due to the growing population, the amount of water required is also expected to increase for food grain production. The nexus among food, energy, and water is also evident from the large water diversion for irrigation purposes, which may potentially hamper hydropower generation as the reservoir waterhead that governs the efficiency of electricity production. Therefore, the proposed approach in this study can help in the sustainable management of the reservoirs in the context of food-energy-water nexus. Specifically, once the understanding of water resources in the basin is established, it can help divide the resources between diversion and storage, which in turn provides the guidelines for hydropower production and river basin management.
The objectives of this study were to (a) develop a methodology to estimate dam characteristics such as storage variation, inflow, and outflow from the reservoirs using remotely sensed observations; (b) analyze the effect of dams on the downstream region through estimating change in the flow and diversion from the reservoirs; and (c) characterize dam operation patterns defined by the rule curves. To investigate these scientific tasks, we utilized remote sensing products to compute storage variation in the reservoirs by combining the simulated inflows from the hydrological model to quantify the outflow from the dams and by developing rule curves based on the total reservoir storage. Furthermore, water diversions from the reservoirs were estimated using ET from MODIS as a proxy to crop water demand. The altimetry data set from GRLM and Hydroweb databases and the satellite imagery from Landsat 8 and Sentinel-2 and the variable infiltration capacity (VIC) model were collectively employed, in addition to the observed reservoir data for the dams, where available, in the basin.
The satellite imageries show seasonal expansion and contraction of the surface area of the Lam Pao, Sirindhorn, and Ubol Ratana reservoirs (Figure A1
). The surface area’s fluctuation followed the monsoon season, resulting in shrinkage during the premonsoon season (May) and swelling at the end of the season (November). The lesser variation in the surface area of the Sirindhorn (79 km2
), as compared with Lam Pao (154 km2
), was due to dam’s smaller catchment area. The minimum surface area remained constant possibly due to the management of the reservoir. The large variation in the surface area was caused by the high flows contributed by the large catchment area of the Ubol Ratana dam. The reservoirs’ maximum surface areas simulated by the satellite imageries were 239.6 km2
, 283.6 km2
, and 369.2 km2
for Lam Pao, Sirindhorn, and Ubol Ratana, respectively, with a difference of 0.17%, 1.55%, and 11.05% as compared with the observed surface area.
The results obtained using the remote sensing data were in agreement with the published literature. The simulated reservoir storage closely represents the observed storage with 15.73% relative root mean square error (RRMSE, R2
= 0.82) for Lam Pao and 3.65% RRMSE (R2
= 0.98) for Sirindhorn. A similar approach was adopted by Bonnema and Hossain [36
] for the storage estimation of reservoirs. However, Bonnema and Hossain [36
] used LandSat8 for the estimation of surface area, while this study uses Sentinel2 along with the LansSat8 images. The altimeter data was also used for the water level extraction in this study in contrast to the trapezoidal approximation using the Shuttle Radar and Topography Mission (SRTM) and the digital elevation map (DEM) to calculate total storage variation of the reservoirs. Landsat 8 images derived NDWI were used for the surface area calculation, while the water level was obtained from Shuttle Radar Topography Mission digital elevation model (SRTM DEM) and Jason-2 satellite altimetry mission for Sirindhorn and Oroville reservoir, respectively. Comparison of the simulated and observed storage volume agreed to within 20% of RRMSE. NDWI was also used for estimation of the surface area of large lakes and reservoirs in the Yangtze River Basin using the MODIS surface reflectance data. The storage simulated by the area based power model demonstrated 15.25% and 33.68% root mean square deviation (RMSD) with respect to reported capacities of lakes and reservoirs, respectively [67
The surface area and storage of the five largest reservoirs in the United States (Lakes Mead, Powell, Sakakawea, Oahe, and Fort Peck Reservoir) was estimated using the satellite-based data products. The estimates were highly correlated with observations with correlation coefficient (r) values between 0.92 and 0.99 for storage and 0.94–0.99 for the surface area. Moreover, the correlation coefficient (r) value varied from 0.08 to 0.97 for the reservoirs located in Iraq, Kyrgyzstan, Africa, and South America [28
]. Pipitone at al. [26
] obtained the R2
value for the surface area estimation of the Magazzolo reservoir as 0.97 and 0.95 from the unsupervised classification and visual matching approach, respectively. The general relationship between the measured volumes and remotely sensed surface areas of small reservoirs in the Preto River Basin showed a high correlation with R2
value equal to 0.83 [30
A well-calibrated hydrological model typically has an NSE value greater than 0.75, and the values of the NSEs between the observed and simulated flows for the gage station locations suggested a “very good” classification status [68
]. Streamflows varied substantially, with peak monthly streamflows gradually increasing from Chiang Saen in the upstream region to Kratie in the lower portion of the basin, ranging from 4000 m3
/s to more than 40,000 m3
/s. These comparisons confirmed that the VIC model could capture the magnitude and variations in streamflows throughout the LMB. The values of the NSE for the dam inflow were also greater than 0.65, the comparison can be considered as “good”. The decline in the NSE and R2
values for the dams as compared to the gauge station was partly due to low streamflow magnitudes. In other words, flows in the tributaries were less and the discrepancies between observed and simulated flows were, therefore, higher in those sections relative to the main stem gauge locations. Peak inflows to Lam Pao and Sirindhorn were estimated as 450 m3
/s, whereas it was 1000 m3
/s for Ubol Ratana, which constituted about 5.6% and 12.5% of the peak flow relative to the farthest upstream gauge station (Chiang Saen). The correlation between the observed and simulated streamflow at gage station locations for calibration and evaluation periods is almost identical to the result obtained by Bonnema and Hossain [36
]. However, Bonnema and Hossain [36
] used the Global Summary of the Day (GSOD) by National Climate Data Center (NCDC) meteorological data as forcings for the VIC model, while this study used the gridded Global Meteorological Forcing Dataset (GMFD). This study also evaluated the calibration of the VIC model at dam locations in addition to the gage station on the mainstem utilized by Bonnema and Hossain [36
]. The efficiency of the VIC model was simulated and the stream was similar to the NSE and r values obtained by Tatsumi and Yamashiki [24
The inflow to the Ubol Ratana reservoir was affected partly by the Chulabhorn Dam, situated on the upstream of the same river network. Since the catchment area and storage capacity of the Chulabhorn reservoir were relatively insignificant, the dam’s effect on the flows was considered minimal. Moreover, the high NSE value between the observed and simulated inflows to the reservoirs suggested that the resultant flow entering the reservoirs was accounted for comprehensively. The cascading effect of the upstream dams and other anthropogenic effects over the catchment area altering the natural streamflow were assimilated in the resultant inflow to the reservoirs. However, the compounding effect of these dams, as considered in this study, suggested that the series of dams on the tributaries and main stem of the river system, with minimum information on the management of these systems, could still be accounted systematically to provide reasonable estimates of inflows to other dams in the downstream sections of the Lower Mekong River.
Operating rule curves generally served as a guiding principle for reservoir storage, regulating the flow and maintaining the operations of a dam. The simulated rule curve closely resembled the current ones for the reservoirs. Kumphon [69
] also estimated the similar rule using the multi-objective genetic algorithm approach with the range 350.8–622.2, 116.6–880.8, and 139.0–1315.3 million m3
for Lam Pao, Sirindhorn, and Ubol Ratana reservoir, respectively. Similar rule curves for Sirindhorn and Ubol Ratana were also derived by a water circulation model that incorporates paddy water management, including reservoir operation and water allocation schemes [70
]. Since the cumulative impacts of the inflow and outflow from the reservoirs were quantified implicitly from the total storage changes, the rule curve was useful to estimate the discharge from the reservoirs based on the inflow to the reservoirs. The monthly range of the reservoirs’ rule curve was primarily governed by the magnitude of the inflow. Since the monsoon season (June–November) generated a large volume of inflow to the reservoir, the spread between the lower and upper limits of the rule curve was large enough to accommodate the uncertainty associated with the fluctuation in the catchment response during extreme events. Since the inflow during the dry period was low in magnitude and fluctuation, the uncertainty in the rule curve was also less. The low reservoir storage during the dry period helped accommodate the high inflows during the monsoon period.
Suitable conditions for the application of the approach: The approach is highly favorable for the derivation of the information of the large reservoirs located in remote locations or with restricted data. However, the efficiency of the technique is restricted for the small reservoirs with surface area less than 1 km2
. This is the requirement for minimizing the edge effect when delineating the waterbody based on the 3 kernel × 3 kernel at 30 m resolution. It will facilitate the precise edge delineation of the water body. We might not be able to define the size of the water-based on the data we have and the analysis performed in this study. Based on the data and the analysis performed in this study, it is difficult to define the size of the waterbody. Moreover, the storage of the reservoir also requires the path of the altimeter to pass through the reservoir. Therefore, the eligibility of the reservoir for storage estimation is dependent on two factors; the size of the reservoir, and the location of reservoir on altimeter path. The size of the reservoir alone does not ensure the suitability of it for storage estimation, as the altimeter path may or may not pass through the reservoir. However, the accurate estimation of the storage is dependent on the precise quantification of the change in inundated area of the reservoir. The precise estimation of the reservoir surface area is dependent on the high temporal and spatial resolution of the satellite images. Hence, relative to coarse resolution satellite products, LandSat8 and Sentinel2 enable us to compute the area of large reservoirs with minimal error along the edges of the waterbody. In addition, the remotely sensed altimeters are beneficial for the reservoirs falling within the satellite path. The altimetry data accuracy is also dependent on the altimetric range (distance between antenna and target), satellite orbit information, the geophysical range corrections, and target size [28
]. The hydrological characteristics of the dam catchment play an important role in determining the inflow to the reservoirs. Based on the catchment properties of the Lam Pao, Sirindhorn, and Ubol Ratana dams, the approach can be used for the watersheds with minimal groundwater influence on the streamflow. Besides, the accuracy of the streamflow simulation is also governed by the high-resolution stream network. Furthermore, the approach is applicable in the river basins where the flow alterations due to the anthropogenic activities are proportionate to the basins studied here. Lastly, the results of the approach can be validated for the reservoirs that have weekly storage data for at least a year.
Remotely sensed data can play a major role in providing vital information over vast regions that are not easily accessible to collect in situ data. In this study, we have utilized the Landsat8 (2013–2018), Sentinel-2 (2013–2018), GRLM (2008–2018), Hydroweb (2008–2018), and MODIS ET (2008–2016) data sets with the VIC model to complement the nonavailable dam operation data in describing the flow variation due to the dam on downstream regions. The major findings of this investigation are summarized below.
Variations in the reservoirs’ total storage were characterized exclusively using remote sensing data from satellite imageries (LandSat8 and Sentinel2) to estimate the surface area and altimetry data sets (ERS-1, T/P, ERS-2, GFO, Jason-1, Envisat, Jason-2, and SARAL/AltiKa) for defining the water level change. Simulated total storage conditions closely agreed with observed storages with NSE and R2 values of 0.90 or higher. The reservoir’s total storage was minimum during May–June and then increased gradually to reach the optimal values by the monsoon termination (October–November).
The VIC model simulated the streamflow that compared well with the observed streamflow across seven gauge stations in the Mekong River basin, with an NSE of more than 0.85 and R2 of more than 0.90 peak values from ~8000 to ~50,000 m3/s. The inflow to each of the reservoirs was also accurately simulated by the VIC model (NSE > 0.64 and R2 > 0.67) and combined with the storage variation to estimate the outflow with an NSE greater than 0.65.
The monthly comparison of the inflow to and outflow from the reservoirs showed a general reduction in outflows. The seasonal cycle of the inflow and outflow from the reservoirs exhibited the lesser outflow during the wet period (June–November) and more outflow during the dry period (December–May) than inflow. Lam Pao had an average decrease of 54% in the outflow during the wet period and a 94% increase during the dry period; these values were 70% and 36%, respectively, for Sirindhorn. Ubol Ratana had also a decrease of 60% in the outflow during the wet period without much change in flows during the dry period.
Diversions played an important role in the downstream flow decrease. Seasonal diversions matched closely with the monthly demand for ET. However, diversion during the wet season was dampened as crop water requirements were meted out with augmentation from precipitation.
Using the rule curve analysis, the total storage was minimum during June–July, and it increased by November and then decreased for the remainder of the year. The average storage volumes had minimum (maximum) values for Lam Pao, Sirindhorn, and Ubol Ratana as 400 (1025) million m3, 850 (1630) million m3, and 700 (1750) million m3, respectively.
The approach is highly suitable for the analysis of large reservoirs falling within the path of the altimetry satellite passes. However, the approach proposed in this study is suitable for reservoirs that have surface areas more than 0.01 km2. The watershed where negligible groundwater-surface water interaction and similar land cover change conditions prevail is also critical for the implementation of this analysis. In addition, high resolution stream network and bi-weekly storage information for at least a year are required for better results.
By combining different satellite remote sensing data sets with a hydrology model, this study highlighted the feasibility of generating continuous time series of storage variation, water level fluctuation, reservoir inflow and outflow, and rule curves to assess the operation and maintenance of large reservoirs and their impacts on the downstream hydrology in the LMB. Since the LMB region was home for millions of people and was being rapidly developed to harness hydropower, the approach proposed in this study could ultimately help evaluate the feedback among food, energy, and water systems.