Reservoir Assessment Tool 2.0: Stakeholder driven improvements to Satellite Remote Sensing based Reservoir Monitoring

Reservoir Assessment Tool 2.0: Stakeholder driven improvements to Satellite Remote Sensing based Reservoir Monitoring


ACKNOWLEDGEMENTS
First and foremost, I would like to thank Prof. Faisal Hossain for his constant guidance and support throughout the project. From identifying the problem in the first place through dialogue with stakeholders, to co-designing the proposed solutionsthis work wouldn't have been possible without his wisdom and deep insights into the problems of managing water in a data-scarce region.
I would also like to thank Prof. Erkan Istanbulluoglu for helping me get comfortable with hydrology and thinking about watershed modeling, which was especially crucial because of my non-hydrology background.
I express my gratitude to the Civil and Environmental Engineering Department, University of Washington, for providing me with the resources, community, opportunities, and a clean and safe working environment to complete this work.
I would like to especially thank Dr. Nishan Biswas, who laid the foundation for the work on which this dissertation builds on. I am thankful to the recently graduated members of the Sustainability, Satellites, Water and Environment (SASWE) Research group for helping me traverse graduate school as an international student during the height of the pandemic, and the current members for their valuable critiques of the work itself. This work would not be possible without the friends I made along the wayand the shared experiences I had, through IM games, Bainbridge Island visit, hikes in the PNW, lunch at Rainier vista, trips to affordable cafes, and biryani/game nights. I am grateful to have found a community who made Seattle feel like a second home and had always supported me through thick and thin.
Finally, I would like to express my sincere gratitude for my parents, who got onboard the difficult decision to let their only child travel halfway across the world to study hydrology with satellites. I am extremely grateful for their unwavering support, love, and belief in me.

Chapter 1. INTRODUCTION
Humans have significantly changed how the world's rivers flow by building dams and reservoirs at an unprecedented rate in the past couple of decades. Estimates suggest that about 1/6th of the total annual river flow is stored in reservoirs (Hanasaki et al., 2006;Mulligan et al., 2020). These reservoirs account for 57% of the Earth's total surface water storage variability (Cooley et al., 2021) that also have a characteristically different storage change pattern as compared to natural lakes (Ryan et al., 2020). This has significantly changed the seasonal streamflow characteristics of major rivers around the world (Cooley et al., 2021;Lehner et al., 2011;Zhou et al., 2016). For instance, due to regulation by existing reservoirs, the wet-season flow of the Mekong River is estimated to have reduced by 31%, while the dry-season flow has increased by 35% Mekong River Commission, 2019.
At a global scale, 48% of freshwater in rivers are impacted by anthropogenic regulation due to reservoirs, which is projected to increase to about 93% in the future (Grill et al., 2015). The effects of this regulation are very apparent in the basin-wide disturbance to biodiversity (Barbarossa et al., 2020), sediment and nutrient flux to the oceans (Li et al., 2021), channel morphology and land-use-land-cover (Fernandes et al., 2020). Information on reservoir operations is therefore crucial to understand how streamflow regulation drives human alteration of the natural aquatic environment and water resources. Developing this understanding requires prediction of the dynamic condition of the reservoir, such as inflow, outflow, active storage, storage change, water level and evaporative losses. Such data are collected or measured at the specific dam site in developed nations. However, barring a few exceptions, most data are not made publicly accessible or available in developing nations, due to confidentiality requirements on grounds of national security (Plengsaeng et al., 2014) or lack of in-situ measurement infrastructure (Bernauer and Böhmelt, 2020;Bonnema and Hossain, 2017). Thus, the overwhelmingly opaque nature of dam operations at most locations worldwide poses a significant challenge to tracking the dynamic nature of human alteration of flow by dams. Consequently, this hampers building a more complete understanding of the dam's impact on the natural-socioeconomic systems that are dependent on the river basins. This also hinders our ability to improve existing resource management tools in an increasingly dammed and human-impacted river basin in the future.
In dammed river basins, an increasing amount of a river's downstream streamflow is affected by upstream regulation. It is therefore crucial for these downstream reservoir operators to understand how the upstream reservoirs are operated, so that improved operating strategies can be adapted for irrigation, hydropower, and flood/drought management, while also maintaining environmental flow requirements (Bonnema et al., 2016). This naturally calls for cooperation and a need for open data-sharing among all dam operating and stakeholder agencies in a basin.
However, this is difficult to establish and maintain in transboundary river basins (Bakker, 2009).
Effective data sharing between countries requires a combination of various favorable factors, both technological and political (Balthrop & Hossain, 2010;Eldardiry & Hossain, 2021b). These include, but are not limited to, streamlined data-sharing mechanisms, transboundary water treaties for data sharing that facilitate reservoir operations and a willingness for all riparian countries to adopt a shared vision of development. Even in basins that are within a single country, such cooperation and data sharing can be a challenge. A prime example of which is the Cauvery basin dispute between the Indian states of Karnataka and Tamil Nadu (Khandekar & Srinivasan, 2021).
In this increasingly regulated and impounded river basin scenario of the near future, satellite remote sensing-based approaches can help fill the current gaps in data-availability and data access on reservoir operations and streamflow regulation. For instance, Global Reservoirs and Lakes Monitor (G-REALM; https://ipad.fas.usda.gov/cropexplorer/global_reservoir/) is one such satellite observation enabled product that reports the water level of select lakes and reservoirs based on satellite radar altimetry (Birkett et al., 2018). While such information on the water level of lakes and reservoirs can help understand the state of these water storage systems, additional information such as storage change, outflow from reservoirs, and reservoir operation patterns that are critical to understanding the human impact, are not derived. A similar and more regionalized system for monitoring reservoirs is the recent Mekong Dam Monitor created by the Stimson Center (https://www.stimson.org/project/mekong-dam-monitor/). While the breadth of information provided by Mekong Dam Monitor is extensive, to the best of our knowledge, the system is not open-source where all the source code and inputs are made publicly available to ensure reproducibility and scalability to other regions of the world.
In response to current limitations of the existing state of the art on reservoir monitoring systems, Biswas et al. (2021) introduced an open-source and reproducible tool based on satellite observations hydrological modeling-based called the Reservoir Assessment Tool (RAT; https://depts.washington.edu/saswe/rat_beta/). It is currently operational for a total of 1598 reservoirs around the world with scripts and data processing methodologies made fully public.
Since it is the first version (1.0) of such a global and reproducible tool, we refer to it as RAT 1.0 in the rest of the manuscript. RAT 1.0 reports the daily modeled reservoir inflow, monthly surface area variation, storage change and outflow, and the inferred rule curve for South America, Africa, South and Southeast Asia. The open-source nature of the project encourages stakeholders to freely use the framework and add features needed by stakeholders. For example, RAT 1.0 framework has been successfully used to develop a RAT for simultaneous management of water quantity and water quality (RAT-WQ 2 ) by several stakeholder agencies, such as by the Center for Water Commission, 2021). The seasonal cycle of the basin is driven by a tropical monsoonal climatic system, with the basin receiving most of the precipitation during the period of May to October . The Mekong River has been subject to increasing flow regulation in the recent past by the construction of reservoirs, especially due to an increasing demand for food, energy, and water for sustaining the rapidly growing economies of the region. The effects of this increased regulation are also very apparent in its current flow conditions, characterized by reduced flow during wet seasons and higher flows during dry seasons (MRC, 2019(MRC, , 2021, and a decreased sediment load (Kondolf et al., 2018;MRC, 2019). The Mekong River Basin also houses the Tonle Sap Lake (TSL) system, which experiences one-of-its-kind hydrodynamic phenomenon of annual flow reversal due to the monsoonal flood pulse. In addition to providing significant economic, cultural and nutritional value to the inhabitants of the TSL region, it also supports a unique ecosystem and fisheries (Arias et al., 2012). Further dampening of the flood pulse due to upstream flow regulation can lead to cessation of the flow reversal phenomenon, leading to significant disruption to the ecosystem that it supports (Pokhrel et al., 2018).
The biodiversity supported by the Mekong River is currently categorized as relatively high

Chapter 2. DATASETS AND METHODS
This chapter describes the data and methods used in this study to pursue the research question articulated in the previous chapter. The datasets used were satellite observations from various satellite sensors, which are described below. The methods center around estimation of the fluxesinflow, storage change and evaporationfrom the reservoirs, using a combination of hydrological modeling and satellite based storage change estimations.

DATA
Publicly available data from a large array of currently operational remote sensing satellites, such as Landsat 8, Sentinel 1 and 2 make it possible to measure the reservoir surface area changes at a high temporal frequency. Highly accurate water level measurements at a 10-day temporal resolution can be made using Jason-3. This study uses an ensemble of satellite sensors -Synthetic Aperture Radar (SAR) based Sentinel-1, Multispectral imaging-based Landsat-8 and Sentinel-2 A/B, and Radar altimetry-based Jason-3 -to estimate the reservoir's state at a higher temporal frequency. A detailed discussion on the surface area estimation can be found in Section 2.5. The reservoir surface area and water level elevation were estimated using an ensemble of sensors summarized in Table 1. The in-situ area-elevation curve was also obtained from the Electricity Generating Authority The dynamic state of the reservoir is modelled using the modeled fluxes to the reservoir and observed change in reservoir. At each reservoir, an assumption of conservation of mass is made - Where, is the outflow from the reservoir, is the inflow to the reservoir, is the evaporation from the reservoir, and Δ is the storage change of the reservoir (Figure 2). Ground seepage and precipitation on the reservoir are assumed to be negligible components of the reservoir mass balance. The Inflow is modeled using the Variable Infiltration Capacity -5 (VIC 5) model (Hamman et al., 2018), but it can be swapped with any other hydrologic model. It must be noted here that the reported outflow from the reservoir in equation 1 is a combination of released water from the reservoir and all the other losses, such as lateral diversions, groundwater seepage losses and consumptive uses. Even though it would be much more meaningful for stakeholders to model the actual downstream releases from the dam via the spillway and penstocks, due to the current limitations of data sources, partitioning the outflow into its constituent fluxes such as lateral outflow, spilled flow and penstock flow, is not trivial. This limitation of RAT or mass balancebased any reservoir modeling component has to be considered by stakeholders when using the outflow estimated by the RAT framework. The storage change (Δ ) of the reservoir is estimated by assuming a trapezoidal bathymetry of the reservoir.
Here, Δ refers to the total volumetric storage change, and −1 refer to the observed surface areas from satellite remote sensing. The ℎ and ℎ −1 are the water levels associated with the surface areas for time and − 1 (Figure 2). The ℎ and values, estimated from satellite data, are related to each other through the Area Elevation Curve (AEC) relationship. This allows for storage change estimations using either two successive satellite-based surface area observations, or two successive satellite altimeter-based water level elevation estimates. This allows RAT 2.0 to operate in two modes, one using only surface area estimates, referred to as RAT 2.0 (TMS-OS), and another using only water level estimates, referred to as RAT 2.0 (Altimetry) in this thesis. A discussion on both the modes of operation are described in section 2.5 and section 2.6 respectively.
The AEC relationship of each reservoir was calculated using the SRTM 1-Arc Second (30 m resolution) Global Digital Elevation Model (DEM) (Earth Resources Observation And Science (EROS) Center, 2017). The SRTM DEM was generated using SAR technology, which is limited in its ability to penetrate water surfaces to obtain the bathymetric elevation information. This results in a constant elevation over water bodies, which represents the water surface elevation during the SRTM's acquisition time over that region. This issue was dealt with using the AEC generation methodology described in (Biswas et al., 2021), which extrapolates the AEC below the water surface using a fitted power law curve. Given the modular nature of RAT 2.0, topographical survey-based AEC was also used when available to understand the impact of AEC quality on the prediction of reservoir outflow.

HYDROLOGICAL MODELING: VIC-5 AND METSIM
The RAT 2.0 framework is designed to be hydrological model agnostic, and hence users can use their preferred hydrologic model if needed. The current hydrological model used in RAT 2.0 during development is VIC 5 (Hamman et al., 2021). This model has a number of improvements compared to VIC 4.2.d (Hamman and Nijssen, 2016) which was used in RAT 1.0. A discussion on the improvements can be found in Hamman et al. (2018).
The meteorological simulation and forcing disaggregation are decoupled from the hydrological modeling routines in VIC 5, and are packaged as a separate model -MetSim model (Bennett et al., 2020). MetSim was used to perform the meteorological simulation at a 6-hour timestep using (i) the daily minimum and maximum temperatures, (ii) daily precipitation, and (iii) wind speed, as inputs. These meteorological forcings are then disaggregated to -(i) average air temperature, (ii) total precipitation, (iii) pressure, (iv) incoming shortwave radiation, (v) incoming longwave radiation, (vi) vapor pressure, and (vii) wind speed.
Using the disaggregated forcings obtained from MetSim, VIC 5 was run at 0.0625° grid resolution for the Mekong River Basin. The model is run in parallel computing mode, exploiting the native support for the Message Passing Interface (MPI) standard, resulting in drastic improvements to performance. The VIC Routing model (Lohmann et al., 1998)  over RAT 1.0 also underscore that inflow regulation has a negligible effect in these reservoirs, and RAT 2.0 is still able to produce more accurate predictions. However, we recognize that to improve the global utility of RAT 2.0, upstream regulation will have to be actively modeled, which we hope to carry out in a future iteration of improvement of the RAT framework.

EVAPORATION -PENMAN EQUATION
In contrast to RAT 1.0, where the modeled Evaporation from the VIC hydrological model was used, the evaporation in RAT 2.0 is modeled explicitly using the Penman Equation. Also known as the combination equation, the Penman equation for free water surface (Penman, 1948;Van Bavel, 1966) is defined as follows - Where, is the evaporation; Δ is the slope of the saturation vapor pressuretemperature relation; is the net incoming radiation; is the ground heat flux, which can be assumed to be 0 for daily calculations; is the mass transfer coefficient, see (Dingman, 2015, p. 259;Harbeck, 1962), is the psychrometric constant; is the water density; is the latent heat of vaporization; is the wind speed; * and are the saturation and actual vapor pressure for the air temperature.

HIGH FREQUENCY SURFACE AREA ESTIMATION -A TIERED MULTI-SENSOR (TMS) APPROACH
High temporal resolution of the storage change and outflow estimates is crucial for stakeholders to understand weekly to sub-weekly reservoir operations of upstream reservoirs. RAT 1.0 quantified the reservoir surface area at a monthly resolution by mosaicking Landsat scenes.
The advantage of this approach is that mosaicking of multiple Landsat scenes helps address the effect of cloud cover that can cause a loss of accuracy in the reservoir surface area estimation. The disadvantage of this approach, however, is the inability to quantify any sub-monthly reservoir operations.
Hence, to address the need for modeling sub-monthly reservoir operations, a novel multisensor surface area estimation algorithm with a tiered self-correction process was developed. This method takes advantage of the strengths of both optical and SAR sensors in a teamwork fashion.
Optical sensors can very accurately estimate the water area (Cordeiro et al., 2021), but they can be highly limited by the presence of clouds. They also can be uncertain around dendritic reservoir shorelines if the perimeter is significantly larger than the nominal surface area (Biswas et al., 2021). On the other hand, SAR can penetrate clouds and is hence not affected by the presence of clouds. However, SAR backscatter threshold-based water extent estimation methods are highly sensitive to the choice of the chosen threshold, and have a tendency to underestimate the surface areas in inundated vegetation (Ahmad et al., 2020). This method combines the high accuracy of optical sensors with the cloud penetrating property of SAR sensors. This method is referred to as TMS-OS (Tiered Multi-Sensor approach -Optical, SAR, pronounced as Teams-OS) in the rest of the manuscript.
The first module processes Sentinel-2 and Landsat-8 scenes in GEE to obtain the reservoir surface area using recommendations and findings from Cordeiro et al., (2021). The Region of Interest (ROI) was first delineated by taking a buffer around the reservoir boundary. The buffer distance was taken as 500 m by default, but is user configurable. The Sentinel-2 L2A -Surface Reflectance overlapping the ROI was scaled, and clouds were masked using the Scenes Classification Map (SCM) band in the preprocessing stage. Landsat-8 L2 Collection 2, Tier 1 -Surface Reflectance scenes overlapping the ROI were also identified, and clouds were masked using its pixel quality bitmask band (QA_PIXEL). The scenes with more than 90% cloud cover over the ROI were filtered out due to extreme cloud cover conditions, and no further processing was performed.
The cascade simple k-means clustering algorithm (Google, 2022) was used to perform unsupervised classification on a subset of pixels sampled from cloud-masked scenes. The best value for k was chosen based on the Calinski and Harabasz, (1974) (Cordeiro et al., 2021).
After classifying the unmasked pixels as water or non-water, the Zhao and Gao, (2018) method of classifying masked pixels as water/non-water based on historical water probability (Pekel et al., 2016) was employed. The water classification and cloud mask correction components were both implemented in GEE, resulting in minimal on-premise computing requirements. The resulting reservoir water area time-series of 16 and 5-day temporal resolutions from Landsat-8 and Sentinel-2, respectively, were combined to obtain a time-series of 1-5-day frequency ( ).
Even after employing skillful unsupervised clustering method for classifying water and correcting for cloud-cover related artifacts, the time-series can still have unrealistic drop or gain in surface area estimates. This can happen for several reasons, especially owing to complete automation of the processing pipelinesuch as sub-optimal choice of K by the Calinski and Harabasz, (1974) criterion, or inability of the Zhao and Gao, (2018) algorithm to completely filter out cloud related artifacts. Since SAR can penetrate clouds, a reasonable assumption was made that Sentinel-1 will be able to accurately quantify the trends of reservoir surface area correctly as increasing or decreasing even if there is quantitative mis-match from ground observations. Hence, further filtering and a novel trend-based correction using Sentinel-1 (SAR) was applied to the optically derived surface area time-series.
In the first tier of filtering, the areas reported by optical sensors are compared against Sentinel- since the time-series can contain outlier values due to challenging scenarios as described above which can skew the estimated median, a further filtering step was performed using a standard deviation ( ) threshold. Using a trial-and-error method, a threshold value of ±2 was found to be skillful in filtering out these outliers, by discarding deviation values outside the ±2 range. Finally, the apparent bias ( ) was calculated as the median deviation of this filtered distribution. This filtration step and calculation of is illustrated pictorially in Figure 3 using dummy data.
The normalized deviations were then calculated as: . = − . Data points were filtered if the normalized deviation was greater than a defined threshold value. Using a trial-and-error method, a threshold value of ~5% of the nominal surface area of the reservoir was found to be skillful at filtering out physically unrealistic surface areas. The normalized deviations were then calculated as: . = − . Data points were filtered if the normalized deviation was greater than a defined threshold value. Using a trial-and-error method, a threshold value of ~5% of the nominal surface area of the reservoir was found to be skillful at filtering out physically unrealistic surface areas.
A second tier of filtering was applied to the dataset by comparing the reservoir water area change trends estimated by optical and SAR sensors. Data points were filtered out if the difference in trends was greater than a defined threshold value. Using a trial-and-error method, a threshold value of ~5% of the nominal surface area of the reservoir was found to work well as the threshold value in weeding out unrealistic drops in . The trend is defined as follows: Where, 1 and 2 are dates of observation of optical sensors (1-5-day resolution) and * ( 1) and * ( 2) are the interpolated areas on the respective dates by the SAR Sentinel-1 sensor.
Similarly, the trends in were also obtained.
Finally, a trend-based correction step was applied to fill in the filtered data points using the apparent trend in surface area change estimated by the SAR sensor, . This step assumes that, even if are biased, the bias term will get cancelled out while calculating . For each point that was filtered out in the tiered filtering steps, the data was filled by estimating the area using as follows - Where, 2 is the date of observation of a data point that was filtered out, and 1 is the previous date of observation.

WATER LEVEL -ALTIMETRY
In addition to estimating the reservoir surface area using optical and SAR sensors, the reservoir water level was also separately estimated using the Jason-3 radar altimeter. Radar altimetry-based water level estimates are highly accurate, however, their applicability in regional to global reservoir mapping is limited by either the spatial coverage or the temporal resolution of the sensors. Due to the sparse spatial coverage of the Jason-3 altimeter, we validated the use of altimeter data over two reservoirs as a proof of concept to assess how RAT 2.0 performs using altimetry derived reservoir elevation data alone.
The water level (ℎ) of reservoirs were estimated using the Okeowo et al., (2017), an automated clustering based method for obtaining in-land water surface elevations from Jason-3 observations.
Although the Okeowo et al., (2017) method is quite robust for most cases as it uses a k-means clustering technique, there will always be challenging cases that warrant a final statistical sanity check and adjustment. For example, when the altimeter ground track has a short length over the reservoir or is near the shoreline, the reported height may be erroneous, especially if the reservoir shoreline contracts substantially due to reservoir drawdown or during a drought. These outliers appear as sudden jumps in the estimated reservoir water level, and were filtered using a standard deviation threshold-based method.
For an observation, the standard deviation was calculated in a rolling window of 4 preceding (left standard deviation -) and 4 following (right standard deviation -) observations. A filtration condition was then defined, such that, if both the left and right standard deviations were higher than 2 standard deviations, then the data points were filtered out. The threshold value of 2 standard deviations was chosen based on a trial-and-error basis such that most of the outliers were filtered out, and the authors recommend selecting a threshold value specific to each reservoir based on the intersection of the ground track with the reservoir. The rationale of the filtration step is as followsif a data point is more than 2 standard deviations as compared to both previous and next observations, then it is very likely that the data point does not reflect an actual rise or decrease in the water level, and hence can be filtered out, and vice versa. The filtered-out points were then filled by linearly interpolating the neighboring observations. The AEC was then used to obtain the corresponding surface area of the reservoir and the storage change was then estimated using equation 2.

OUTFLOW FORECASTING
In addition to providing nowcast and historical (hindcast) estimates of outflow, the functionality to forecast outflow for a 15-day lead time was also added to RAT 2.0. The forecasted meteorological conditionsminimum and maximum temperature, u-and v-components of wind speed, and precipitationobtained from the GFS dataset were used to force the VIC hydrological model to estimate the forecasted inflow to the reservoirs. The forecasted evaporation was also estimated using equation 3 with forecasted meteorological conditions obtained from the GFS dataset.
The RAT 2.0 forecasting framework provides two options of rule-curved based estimation of outflow for a 15-day lead time -(1) satellite derived rule curve from RAT 1.0, (2) user defined rule curve or time-varying target storage. The rule curve in RAT 1.0 is derived based on averaging the historical storage change patterns over a multi-year period for each month, and then inferring the most likely storage target ( ) that the reservoir has tried to achieve, as a fraction of the maximum storage ( ). The rule curves are then interpolated to a 5-day frequency using linear interpolation. The interpolation is performed to stay consistent with the pentad frequency of hindcasted ∆S using the TMS-OS approach. The expected storage change is then estimated from the interpolated derived rule curve as follows - Where Δ is the expected storage change as a percentage of maximum storage based on rule curve for lead time, . / corresponds to the storage as a fraction of maximum storage at lead time , and 0 / corresponds to the storage as a fraction of for the current timestep. The ST can be obtained either from the inferred rule curve (option 1) or from a user-defined target storage at . The can be obtained either from in-situ data, or from dam metadata repositories, such as the GRanD database (Lehner et al., 2011). This expected storage change was used in the mass balance equation 1 to estimate the forecasted outflow.
Where , and Δ are the outflow, inflow and storage change at a future time . The evaporation was ignored due to the low contribution of the evaporation in dictating the outflow for the Mekong region. Since reservoir operations can highly depend on the on-site decisions made by the dam operator, antecedent conditions, and the need to meet storage and distribution targets, RAT 2.0 also provides the functionality to directly input the expected storage change for a given lead time and visualize the expected outflow for the same lead time.
2.8 COMPUTATIONAL CONFIGURATION OF RAT 2.0 The RAT 2.0 was developed and tested in a CentOS 7 Linux environment, setup on a highperformance server. The server is configured with 2× 8 core Intel Xeon Gold processors, each having 2 threads per code, equating to a total of 32 computing units. MetSim and the VIC hydrological model were both configured to make use of the 32 computing units for parallel computation. The main memory of the system totals to 192 GB, allowing for a high ceiling for incore processing.
A data-processing module for pre-processing raw satellite data-products and transforming them into model inputs was developed. This module makes use of the high memory ceiling for performing data transformations as in-core operations to significantly improve the performance.
While this implementation has high memory requirements, the implementation can be easily changed to an out-of-core processing schema according to constraints.
In its current implementation, the RAT-Mekong model takes about 3 hours to run for the The remote sensing module offloads the processing to Google Earth Engine's (GEE) cloud computing infrastructure, and hence has minimal local processing requirements.
Currently, the RAT model is set up to run daily at a 3-day lag, to offset any data provider side delays and their inherent latency. For example, many satellite datasets have a 1-2 day latency, sometimes even longer with data outages. If there were no delays or latency associated with input data, then RAT 2.0 can theoretically produce outputs within 24 hours of the most recent satellite observation. We anticipate such a scenario to be a reality as IT capabilities improve and latency decreases for satellite missions and data providers.

Chapter 3. RESULTS AND DISCUSSION
This chapter describes the results of applying the methods discussed in the previous chapter. The relative contribution of evaporation, as compared to inflow, in determining the outflow is assessed below. Moreover, high frequency surface area estimates of the TMS-OS algorithm is also presented, along with comparisons of ∆S and Outflow estimations with RAT 2.0 (Altimetry) and RAT 1.0 (Monthly) is performed.

RESERVOIR EVAPORATION
The evaporative losses were found to be very nominal for the South East Asian reservoirs, with flux magnitudes less than 10% of inflow for 30 out of 36 assessed reservoirs. The catchment areas of the remaining 6 reservoirs are relatively small, which generate low inflow volumes.
Moreover, out of 6 other reservoirs, 5 are primarily used for hydroelectricity. Operators usually keep the reservoir close to full to maximize hydroelectricity production, which explains the higher evaporation compared to the natural inflow. In general, the evaporation does not seem to be a major control on the water balance of the reservoirs of South East Asian River basins, which is consistent with previous study of Bonnema et al., (2016). However, in other river systems across the globe, evaporative losses may be significant for relatively drier river systems, such as in the Nile (Eldardiry & Hossain, 2019) and the Tigris-Euphrates. Since the RAT 2.0 framework is designed for global applications, the evaporative losses are still taken into account for the hindcast and nowcast of the outflow from reservoirs, even though the evaporative losses do not play a major role in the currently studied Mekong Region. This will ensure that the RAT 2.0 framework is able to model reservoir dynamics in parts of the world, such as the Nile basin, Tigris-Eupgrates, and the Middle-East, where evaporation from lakes and reservoirs plays a more dominant role (Zhao et al., 2022).

HIGH-FREQUENCY SURFACE AREA ESTIMATION -TMS-OS
The surface areas using TMS-OS were obtained for the period of 2019-2021 (3 years) due to the overlap of data between all the datasets -Sentinel-2, Landsat-8 and Sentinel-1. Figure 7(a) shows the 1-5 day surface area time-series obtained using the optical sensors employing the methodology described in Figure 4(a). At this step, rapid drops and rises in the estimated surface areas can be noted. These drops in surface areas do not reflect actual changes in reservoir surface area, rather, they are artifacts of the Cascade k-means clustering algorithm and limitations of the optical sensor to discriminate water from land. Since the Calinski and Harabasz, (1974) criterion is used to select the number of clusters, a sub-optimal value for k can be selected by the criterion in case of challenging scenes. Such challenging scenes include, but are not limited to, cloud pixels that do not get identified accurately by the cloud masking algorithms, presence of high sediment load or algal conditions, or issues such as shadows in the scene due to terrain or clouds. In such cases, the clustering algorithm can select a higher value of k, that, even though has the lowest Calinski-Harabasz index value, but the chosen cluster of pixels as water pixels, does not represent all the water pixels in the scene. Choosing a sub-optimal value of k is a disadvantage of automatic selection of the number of clusters. However, these errors get identified, filtered and corrected in the next steps of the algorithm.

STORAGE CHANGE
The modeled storage changes derived from RAT 2.0-TMS-OS, RAT 2.0-Altimetry and RAT 1.0 were compared against in-situ observed storage change. The in-situ storage change was defined as follows - Where 1 and 2 are dates of consecutive satellite observation, and, ( 1) and ( 2) are the corresponding in-situ storages. The ∆S, hence has a unit of [ 3 ], representing the amount of storage change between consecutive satellite observations. The comparison metrics are summarized in Table 3. The estimated storage change values using Jason 3 altimeter elevation data that were further corrected based on methodology described in section 2.6, were compared against the observed insitu storage change. The comparison metrics are summarized in Table 3. Figure 9 shows the derived water level after filtering out outliers, and the estimated storage change values overlaid on the observed in-situ storage change. With the highest correlation, the altimeter-based technique was found to be the most skillful in modeling the storage change, albeit at a 10-day frequency. Figure 9. Comparison of modeled storage change using altimeter measurements of reservoir water level with observed storage change for Lam Pao, Thailand.
Since the comparisons were performed at the temporal resolution of the respective models, the statistics corresponding to RAT 2.0-TMS-OS quantify the performance at a 1-5 day frequency, RAT 2.0-Altimetry at a 10 day frequency, and RAT 1.0 at a 30 day frequency. This underlines the inherent improvement brought forth by RAT 2.0, that is, observing the reservoir dynamics at a pentad frequency or less using optical and SAR, and at 10-days using altimetry, as opposed to the monthly frequency of RAT 1.0. Even at this frequency, RAT 2.0 performs better for every metric.

OUTFLOW
The outflow estimates obtained using Equation 1 were compared with the observed outflows for the three validation reservoirs. Additionally, the outflow estimated by RAT 2.0 using the TMS-OS approach was compared against the outflows estimated by RAT 1.0 which are summarized in Table 4. The metrics were obtained for 1-5 day temporal resolution for RAT 2.0, and at a monthly frequency for RAT 1.0. Even at this higher temporal frequency, RAT 2.0 is able to improve the flux estimations as compared to RAT 1.0. 3.5 OUTFLOW FORECASTING  Additionally, the second method of outflow forecasting, using ∆S defined by the stakeholders, is presented in Figure 11. The user has to first select the reservoir of interest using the "res" dropdown menu. Once chosen, the user is presented with a table showing the "nowcast" estimates for inflow, ∆S, outflow and a column describing if the estimate is a "nowcast" or "forecast". The user is also presented with three rows of inputs, corresponding to the next 3 anticipated satellite observation dates, where the expected storage change as a percentage of the maximum storage of the reservoir can be passed. The plot showing the ∆S and outflow will then update, based on the choice of ∆S, with the forecasted outflows highlighted as orange circles. Figure 11. Forecasted outflow using user-defined ∆S, as a percentage of maximum Storage of the reservoir.

CONCLUSIONS
Regulation of rivers, and the issues associated with it, such as environmental degradation, conflicts among nations, and change in the river geomorphology are negatively affecting an increasing portion of the world. With a near-global coverage, publicly accessible methods and open-source code, and a record of reservoir operation dynamics, RAT 1.0 (Biswas et al., 2021) provided information on reservoir operation dynamics for understanding these increasingly regulated river systems.
Building on RAT 1.0, we made several improvements, based on stakeholder feedback, to develop a more skillful version 2.0 of the RAT framework. While these improvements were demonstrated using the Mekong River basin as an example, the newly developed framework can apply anywhere where reservoirs need to be monitored routinely. Using techniques such as, advanced unsupervised classification-based water area estimation, an ensemble of sensors with complementing strengths, and a tiered filtering and correction approach, we were able to estimate changes in the reservoir dynamics with higher accuracies than RAT 1.0. This method, called the TMS-OS, was found to be highly skillful in quantifying the storage change, with a higher average correlation, and lower RMSE and MAE values as compared to RAT 1.0. This bodes well for the Surface Water and Ocean Topography (SWOT) mission (https://swot.jpl.nasa.gov/) which is planned for launch in 2022. The SWOT satellite mission is a joint mission of the National Aeronautics and Space Administration (NASA)) and Centre National d'Etudes Spatiales (CNES), with contributions from the Canada Space Agency and the United Kingdom Space Agency (Biancamaria et al., 2016). SWOT is expected to monitor reservoir elevation and extent at a global scale using its wide-swath capability (Lee et al., 2010) and thereby complement altimeters that have limited but accurate spatial sampling. The SWOT mission is therefore expected to further improve our ability to predict the dynamic state of reservoirs.
In general, RAT 2.0, with its use of multiple sensors and tiered methodologies, performed better than RAT 1.0 in characterizing the outflow from reservoirs. The ability of visualize the forecasted outflow, based on forecasted inflow and forecasted storage change, either from historical reservoir operation patterns, or expectation of dam operators based on human understanding, was also introduced in RAT 2.0.
Even though significant strides were made in quantification of the reservoir operation dynamics, there is always room for improvement. The effect of upstream reservoir operations on the downstream reservoirs remains to be taken into account and studied at the basin level. With the use of a greater suite of altimeters and sensors, combined with machine learning to automatically fill in inter-sensor gaps or data outages, we believe it may be possible to predict reservoir states consistently at the daily resolution, leading to higher accuracy in outflow estimation. Such future improvements can further help democratize the access to information on operations of dams towards more sustainable and equitable management of water resources.