Application of ENSO and Drought Indices for Water Level Reconstruction and Prediction : A Case Study in the Lower Mekong River Estuary

Water level monitoring is important for understanding the global hydrological cycle. Remotely-sensed indices that capture localized instantaneous responses have been extensively explored for water level reconstruction during the past two decades. However, the potential usage of the Palmer’s Drought Severity Index (PDSI) and El Niño Southern Oscillation (ENSO) indices for water level reconstruction and prediction has not been explored. This paper examines the relationship between observed water level and PDSI based on a soil-moisture water balance model and three ENSO indices for the lower Mekong River estuary on a monthly temporal scale. We found that the time-lagged information between the standardized water level and the ENSO indices that enabled us to reconstruct the water level using the ENSO indices. The influence of strong ENSO events on the water level can help capture the hydrological extremes during the period. As a result, PDSI-based water level reconstruction can be further improved with the assistance of ENSO information (called ENSO-assisted PDSI) during ENSO events. The water level reconstructed from the PDSI and ENSO indices (and that of remote sensing) compared to observed water level shows a correlation coefficient of around 0.95 (and <0.90), with an RMS error ranging from 0.23 to 0.42 m (and 0.40 to 0.79 m) and an NSE around 0.90 (and <0.81), respectively. An external assessment also displayed similar results. This indicates that the usage of ENSO information could lead to a potential improvement in water level reconstruction and prediction for river basins affected by the ENSO phenomenon and hydrological extremes.


Introduction
Water resources, including fresh water supply, hydrological extremes and agricultural irrigation, is critical for human sustainability in the near future [1,2]. With the increasing demand for water resources, it is more and more important to monitor the water volume in rivers, lakes, reservoirs and wetlands [3]. Traditionally, the water level (or stage) is one of the basic hydraulic variables measured along river reaches and estuary mouths. It is a useful information for understanding local water variability. It is also essential for monitoring hydrological extremes (including floods and droughts) that cause agricultural and economic losses [4]. Therefore, continuous, accurate, and timely water level monitoring for an early alert regarding hydrological extremes is essential [5].

The Geographic Setting of the Mekong River Basin
The Mekong River Basin, one of the largest river basins of the world, represents the largest basin in Southeast Asia, draining an area of 795,000 km 2 with a total length of 4800 km. It is inhabited by 65 million people [35]. The three-rivers region (TRR), located on the southeastern Qinghai-Tibet Plateau, is its headwater source, with water flowing through parts of southern China, Burma, Laos, Thailand, Cambodia and Vietnam before discharging into the South China Sea. The Mekong River is divided into the Tonle Sap river, the Bassac river and the lower Mekong at Phnom Penh, Cambodia. Before reaching the ocean, the main river is subdivided into several branches. The lower Mekong River Basin is relatively flat (Figure 1). The hydrological regime of the Mekong River Basin is mainly controlled by the monsoon winds blowing northeast or southwest depending on the season. The southwest monsoon from the ocean corresponds to the rainy season occurring between May and October each year, with a peak in September [35]. Therefore, the water level of Mekong basin begins rising and changing abruptly in May in the eastern and northern highland of the Mekong basin. The highest water level is attained in August or September upstream and as late as in October downstream. The rainy seasons contribute more than 85% of the annual rainfall [36]. Meanwhile, November to March of each year corresponds to dry seasons affected by the northeast monsoon from the mainland, which have less rainfall [35]. The water level of the Mekong basin is the lowest in April. Note that the water within the Mekong River mainly comes from the rainfall in its downstream watershed. The water volume changes depending on the rainfall caused by the monsoons.
In essence, various studies have demonstrated that significant interaction between the Asian monsoon and ENSO exists [37,38]. For instance, ENSO can weaken the East Asian summer monsoon over northern China and Japan [39][40][41]. The South Asian summer monsoon is positively correlated with the tropical Pacific trade winds, which can affect ENSO events [42]. A close relationship between the monsoon and discharge in the lower Mekong basin indicates the presence of a teleconnection between the Mekong River discharge and ENSO [43]. Therefore, the reconstruction of water levels based on the ENSO information is possible in the lower Mekong River Basin, as shown in Section 4. The hydrological regime of the Mekong River Basin is mainly controlled by the monsoon winds blowing northeast or southwest depending on the season. The southwest monsoon from the ocean corresponds to the rainy season occurring between May and October each year, with a peak in September [35]. Therefore, the water level of Mekong basin begins rising and changing abruptly in May in the eastern and northern highland of the Mekong basin. The highest water level is attained in August or September upstream and as late as in October downstream. The rainy seasons contribute more than 85% of the annual rainfall [36]. Meanwhile, November to March of each year corresponds to dry seasons affected by the northeast monsoon from the mainland, which have less rainfall [35]. The water level of the Mekong basin is the lowest in April. Note that the water within the Mekong River mainly comes from the rainfall in its downstream watershed. The water volume changes depending on the rainfall caused by the monsoons.
In essence, various studies have demonstrated that significant interaction between the Asian monsoon and ENSO exists [37,38]. For instance, ENSO can weaken the East Asian summer monsoon over northern China and Japan [39][40][41]. The South Asian summer monsoon is positively correlated with the tropical Pacific trade winds, which can affect ENSO events [42]. A close relationship between the monsoon and discharge in the lower Mekong basin indicates the presence of a teleconnection between the Mekong River discharge and ENSO [43]. Therefore, the reconstruction of water levels based on the ENSO information is possible in the lower Mekong River Basin, as shown in Section 4.

Ground-Based, In-Situ Data and Remote Sensing Observations
Ground-based, in-situ data, at both the Vam Kenh and Dinh An stations, were chosen to correlate with the modeled indices and to conduct validation, respectively, in this study (Figure 1b) due to their geographic location near the estuary mouth and longer data time span, when compared to other stations (Figure 1a). These data are available on request at Mekong River Commission (MRC) (http://www.mrcmekong.org). Apparently, both station time series, with data spanning between 1 January 1992 and 31 December 2006, exhibit a strong similarity and share the same periodicities of 342 days, 176.6 days, and 1369 days, except with different amplitudes (Figure 2). A yearly peak usually occurs in September corresponding to severe flooding for the aforementioned southwest monsoon, and the lowest water level occurs in April, which is consistent with the general description in Section 2.1.

Ground-Based, In-Situ Data and Remote Sensing Observations
Ground-based, in-situ data, at both the Vam Kenh and Dinh An stations, were chosen to correlate with the modeled indices and to conduct validation, respectively, in this study (Figure 1b) due to their geographic location near the estuary mouth and longer data time span, when compared to other stations ( Figure 1a). These data are available on request at Mekong River Commission (MRC) (http://www.mrcmekong.org). Apparently, both station time series, with data spanning between 1 January 1992 and 31 December 2006, exhibit a strong similarity and share the same periodicities of 342 days, 176.6 days, and 1369 days, except with different amplitudes (Figure 2). A yearly peak usually occurs in September corresponding to severe flooding for the aforementioned southwest monsoon, and the lowest water level occurs in April, which is consistent with the general description in Section 2.1. Remote sensing products for the land surface derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) (National Aeronautics and Space Administration (NASA), Washington DC, USA) [44] were used, which are available from the Land Processes Distributed Active Archive Center (LP DAAC) managed by the NASA Earth Science Data and Information System (ESDIS) project (https://lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table). For the above data products, Normalized Difference Vegetation Index (NDVI) data from the MOD13C2 product and the land surface temperature (LST) data from the MOD11C3 product were chosen, serving as a baseline for comparison with the PDSI and ENSO indices. To be comparable with the time span of groundbased observations, remote sensing data were utilized up to the end of the year 2006.

Palmer Drought Severity Index (PDSI)
The Palmer Drought Severity Index (PDSI) [14], widely used in the United States [45], is a standardized index based on a soil-moisture water balance model that quantifies meteorological drought utilizing all available temperature and precipitation data to estimate relative dryness. It ranges between −10 (dry) and +10 (wet). The index has proven that it is strong at capturing long-term drought, while weak in recognizing short-term drought. The PDSI modeled data at 2.5° × 2.5° spatial resolution are readily available at http://www.cgd.ucar.edu/cas/catalog/climind/pdsi.html [22]. Remote sensing products for the land surface derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) (National Aeronautics and Space Administration (NASA), Washington DC, USA) [44] were used, which are available from the Land Processes Distributed Active Archive Center (LP DAAC) managed by the NASA Earth Science Data and Information System (ESDIS) project (https://lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table). For the above data products, Normalized Difference Vegetation Index (NDVI) data from the MOD13C2 product and the land surface temperature (LST) data from the MOD11C3 product were chosen, serving as a baseline for comparison with the PDSI and ENSO indices. To be comparable with the time span of ground-based observations, remote sensing data were utilized up to the end of the year 2006.

Palmer Drought Severity Index (PDSI)
The Palmer Drought Severity Index (PDSI) [14], widely used in the United States [45], is a standardized index based on a soil-moisture water balance model that quantifies meteorological drought utilizing all available temperature and precipitation data to estimate relative dryness. It ranges between −10 (dry) and +10 (wet). The index has proven that it is strong at capturing long-term drought, while weak in recognizing short-term drought. The PDSI modeled data at 2.5 • × 2.5 • spatial resolution are readily available at http://www.cgd.ucar.edu/cas/catalog/climind/pdsi.html [22].

El Niño-Southern Oscillation (ENSO) Indices
The El Niño-Southern Oscillation (ENSO) is an irregularly periodic climatic event that is caused by the interaction between the atmosphere and ocean over the tropical eastern Pacific Ocean.
It can manifest from sea-level atmospheric pressure differences and sea surface temperature (SST) anomalies [24]. El Niño and La Niña represents the warming and cooling phases of the ENSO events, respectively. Such events can lead to floods and droughts that occur every few years and last several months.
Several studies have been conducted to classify different types of ENSO events based on spatial location, occurrence period, and during which season it occurred within a year. For instance, for spatial locations, ENSO events could be divided into two types: one with a center in the central Pacific and the other in the eastern Pacific [46][47][48][49]. According to the occurrence period, two types have a dominant period of about two years (central Pacific) and ranging from two to four years (eastern Pacific), respectively [48]. On the basis of the occurrence season, Xu and Chan [50] classified it into spring-occurring and summer-occurring ENSO events. Combining these two classifications, the ENSO events that are located in the central Pacific are likely to occur in the summer and mainly have a two-year period (namely the central Pacific type), whereas these events situated in the eastern Pacific may start in spring and their dominant period varies from two to four years (namely the eastern Pacific type).
Numerous SST-based ENSO indices have been derived using different areas of the equatorial Pacific Ocean (e.g., Niño 3 (eastern tropical Pacific) region, Niño 4 (central tropical Pacific) region and Niño 3.4 (eastern-to-central Pacific)). Taking into consideration indices for the ENSO events (i.e., the central Pacific type and the eastern Pacific type) in the ocean, in the atmosphere, and coupled atmosphere-ocean systems, the index SST anomaly in Niño 3.4, the Southern Oscillation Index (SOI), and Multivariate ENSO Index (MEI) were used in this study. All ENSO indices were available from https://www.esrl.noaa.gov/psd/data/climateindices/list/. Figure 3 displays the ENSO events using SOI, SST anomalies in the eastern tropical Pacific Ocean area (5 • N-5 • S, 120 • W-170 • W, called Niño 3.4 region), and MEI. The ENSO event starts when both the El Niño and La Niña event occurs in a pair with various intensities and irregular periods and durations. In addition, the SST anomaly in the Niño 3.4 region and MEI have a high correlation with each other, while being negatively correlated with SOI.

El Niño-Southern Oscillation (ENSO) Indices
The El Niño-Southern Oscillation (ENSO) is an irregularly periodic climatic event that is caused by the interaction between the atmosphere and ocean over the tropical eastern Pacific Ocean. It can manifest from sea-level atmospheric pressure differences and sea surface temperature (SST) anomalies [24]. El Niño and La Niña represents the warming and cooling phases of the ENSO events, respectively. Such events can lead to floods and droughts that occur every few years and last several months.
Several studies have been conducted to classify different types of ENSO events based on spatial location, occurrence period, and during which season it occurred within a year. For instance, for spatial locations, ENSO events could be divided into two types: one with a center in the central Pacific and the other in the eastern Pacific [46][47][48][49]. According to the occurrence period, two types have a dominant period of about two years (central Pacific) and ranging from two to four years (eastern Pacific), respectively [48]. On the basis of the occurrence season, Xu and Chan [50] classified it into spring-occurring and summer-occurring ENSO events. Combining these two classifications, the ENSO events that are located in the central Pacific are likely to occur in the summer and mainly have a two-year period (namely the central Pacific type), whereas these events situated in the eastern Pacific may start in spring and their dominant period varies from two to four years (namely the eastern Pacific type).
Numerous SST-based ENSO indices have been derived using different areas of the equatorial Pacific Ocean (e.g., Niño 3 (eastern tropical Pacific) region, Niño 4 (central tropical Pacific) region and Niño 3.4 (eastern-to-central Pacific)). Taking into consideration indices for the ENSO events (i.e., the central Pacific type and the eastern Pacific type) in the ocean, in the atmosphere, and coupled atmosphere-ocean systems, the index SST anomaly in Niño 3.4, the Southern Oscillation Index (SOI), and Multivariate ENSO Index (MEI) were used in this study. All ENSO indices were available from https://www.esrl.noaa.gov/psd/data/climateindices/list/. Figure 3 displays the ENSO events using SOI, SST anomalies in the eastern tropical Pacific Ocean area (5° N-5° S, 120° W-170° W, called Niño 3.4 region), and MEI. The ENSO event starts when both the El Niño and La Niña event occurs in a pair with various intensities and irregular periods and durations. In addition, the SST anomaly in the Niño 3.4 region and MEI have a high correlation with each other, while being negatively correlated with SOI.  Based on the guidelines from the Climate Prediction Center of the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Prediction, the SST anomaly index for the Niño 3.4 region should meet two criteria for the identification of an El Niño (or La Niña) event: (i) the five-month running means of absolute SST anomaly values are larger than or equal to 0.5 • C and (ii) it persists for at least six months. From the intensity guidance (Table 1)

Correlative Analysis and Water Level Standardization Results
The NDVI and LST time series are consistent with the observed water level series (i.e., Vam Kenh station). Therefore, linear regression (i.e., a model fitting with an offset plus a slope) was directly applied to establish a correlative relationship between the observed water level and those remote sensing data (i.e., NDVI and LST) within the overlapping period for both the time series (i.e., from January 2000 to December 2006). Owing to less obvious and direct connections between the monthly PDSI and ENSO index time series and the observed water level time series, observed water levels were standardized to be consistent with the PDSI and ENSO indices before establishing their correlative relationship. This can be achieved by subtracting the monthly averaged water level from the monthly median values (called water level anomaly) for the entire 15-year data time span divided by standard deviation of the corresponding month ( Figure 4a). Based on the guidelines from the Climate Prediction Center of the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Prediction, the SST anomaly index for the Niño 3.4 region should meet two criteria for the identification of an El Niño (or La Niña) event: (i) the five-month running means of absolute SST anomaly values are larger than or equal to 0.5 °C and (ii) it persists for at least six months. From the intensity guidance (Table 1) and time series of SST anomalies between 1991 and 2012 (Figure 3b), two obvious El Niño events (1991)(1992)(1997)(1998) and one obvious La Niña event (1998)(1999)(2000)(2001) can be observed, which have a longer duration and stronger intensity within our water level study periods. Based on SST anomaly values and the criteria of ENSO events, the El Niño in 1991-1992 and in 1997-1998 ranged from June 1991 to June 1992, and from May 1997 to May 1998, respectively, whereas La Niña in 1998-2001 spanned from July 1998 to January 2001. These three events are expected to have had a strong influence on the Mekong water levels.

Correlative Analysis and Water Level Standardization Results
The NDVI and LST time series are consistent with the observed water level series (i.e., Vam Kenh station). Therefore, linear regression (i.e., a model fitting with an offset plus a slope) was directly applied to establish a correlative relationship between the observed water level and those remote sensing data (i.e., NDVI and LST) within the overlapping period for both the time series (i.e., from January 2000 to December 2006). Owing to less obvious and direct connections between the monthly PDSI and ENSO index time series and the observed water level time series, observed water levels were standardized to be consistent with the PDSI and ENSO indices before establishing their correlative relationship. This can be achieved by subtracting the monthly averaged water level from the monthly median values (called water level anomaly) for the entire 15-year data time span divided by standard deviation of the corresponding month ( Figure 4a).  To increase the ability to capture ENSO events and remove noise from the PDSI index, two extra preprocessing procedures were conducted to create an ENSO-assisted PDSI. First, we observed that the PDSI upstream had a stronger sensitivity in dryness conditions than that downstream. Therefore, a representative PDSI data time series point upstream in the Mekong River during ENSO events had to be selected owing to the lesser influence of ENSO events upstream than downstream. The PDSI gridded value very close to Vam Kenh station was not used to avoid potential contamination from the ocean. Second, smoothing and re-scaling processes had to be applied to PDSI data before a correlative analysis could be made. This can be achieved by a five-month moving averaging process with PDSI value re-scaling that matches the water level in a standardized form.
It can be revealed that the PDSI is strongly correlated to the standardized water level anomalies (Figure 4b). A time lag between negative SST and water level anomalies is observed (Figure 4c), because it takes several months for the water level in the Mekong delta to start its response to an ENSO event. Based on a detailed analysis of SST and water level anomalies, we observed that there is a six-month lag before year 1998 and a nine-month lag after January 1998. These lags suggest that there is a teleconnection between the ENSO indices and the water level in the Mekong delta. Many researchers have studied the teleconnections between ENSO events and hydrological variables, such as river discharge [52], precipitation [53], and water storage [49]. In fact, time lag analysis (e.g., [54]) is widely used to improve hydrological modeling.
In this study, the time-lagged El Niño information was utilized to improve our water level reconstruction during the El Niño events. In our study periods (January 1992 to December 2006), two obvious El Niño events (i.e., 1991-1992 El Niño and 1997-1998 El Niño) cause the six-month lag and nine-month lag, respectively. From Table 1, it can be seen that the intensity of the 1991-1992 El Niño beyond the medium level (i.e., SST > 1.06) lasted for six months, and that in 1997-1998 lasted for nine To increase the ability to capture ENSO events and remove noise from the PDSI index, two extra preprocessing procedures were conducted to create an ENSO-assisted PDSI. First, we observed that the PDSI upstream had a stronger sensitivity in dryness conditions than that downstream. Therefore, a representative PDSI data time series point upstream in the Mekong River during ENSO events had to be selected owing to the lesser influence of ENSO events upstream than downstream. The PDSI gridded value very close to Vam Kenh station was not used to avoid potential contamination from the ocean. Second, smoothing and re-scaling processes had to be applied to PDSI data before a correlative analysis could be made. This can be achieved by a five-month moving averaging process with PDSI value re-scaling that matches the water level in a standardized form.
It can be revealed that the PDSI is strongly correlated to the standardized water level anomalies (Figure 4b). A time lag between negative SST and water level anomalies is observed (Figure 4c), because it takes several months for the water level in the Mekong delta to start its response to an ENSO event. Based on a detailed analysis of SST and water level anomalies, we observed that there is a six-month lag before year 1998 and a nine-month lag after January 1998. These lags suggest that there is a teleconnection between the ENSO indices and the water level in the Mekong delta. Many researchers have studied the teleconnections between ENSO events and hydrological variables, such as river discharge [52], precipitation [53], and water storage [49]. In fact, time lag analysis (e.g., [54]) is widely used to improve hydrological modeling.
In this study, the time-lagged El Niño information was utilized to improve our water level reconstruction during the El Niño events. In our study periods (January 1992 to December 2006), two obvious El Niño events (i.e., 1991-1992 El Niño and 1997-1998 El Niño) cause the six-month lag and nine-month lag, respectively. From Table 1, it can be seen that the intensity of the 1991-1992 El Niño beyond the medium level (i.e., SST > 1.06) lasted for six months, and that in 1997-1998 lasted for nine months. This result is consistent with the two observed temporal lags from the analysis of SST and water level anomalies. As a result, when the SST is larger than 1.06, ENSO can affect the Mekong River downstream water level. In addition, the duration of the medium or stronger intensity of ENSO events corresponds perfectly with the time lag of water levels affected by ENSO. As a result, the whole time series can be divided into two sections (January 1992-December 1998 and January 1998-December 2006) based on the two obvious El Niño events. The six-month and nine-month shifts of the SST anomaly time series were applied to the first and the second section, respectively. Owing to the high correlation of other ENSO indices with SST, the same procedure was adopted for the SOI and MEI.
Applying the above-mentioned PDSI data preprocessing procedures to these time-shifted ENSO indices, all ENSO indices showed high agreement with the water level anomaly at the Vam Kenh station ( Figure 5). Compared to PDSI, the ENSO indices performed better. For different ENSO indices, the SST anomaly performed the best. As a result, the correlative relationship between the water level and PDSI and ENSO indices was established, which allows the conversion of the PDSI or ENSO indices back into the water level anomalies time series. In other words, the estimated anomalies were then added back to the seasonal monthly median values to reconstruct the water level. months. This result is consistent with the two observed temporal lags from the analysis of SST and water level anomalies. As a result, when the SST is larger than 1.06, ENSO can affect the Mekong River downstream water level. In addition, the duration of the medium or stronger intensity of ENSO events corresponds perfectly with the time lag of water levels affected by ENSO. As a result, the whole time series can be divided into two sections (January 1992-December 1998 and January 1998-December 2006) based on the two obvious El Niño events. The six-month and nine-month shifts of the SST anomaly time series were applied to the first and the second section, respectively. Owing to the high correlation of other ENSO indices with SST, the same procedure was adopted for the SOI and MEI. Applying the above-mentioned PDSI data preprocessing procedures to these time-shifted ENSO indices, all ENSO indices showed high agreement with the water level anomaly at the Vam Kenh station ( Figure 5). Compared to PDSI, the ENSO indices performed better. For different ENSO indices, the SST anomaly performed the best. As a result, the correlative relationship between the water level and PDSI and ENSO indices was established, which allows the conversion of the PDSI or ENSO indices back into the water level anomalies time series. In other words, the estimated anomalies were then added back to the seasonal monthly median values to reconstruct the water level.

Result Assessment Schemes
To conduct a comparison between observed water levels and estimates derived from the indices, the Pearson correlation coefficient (PCC), root mean square error (RMSE), and Nash-Sutcliffe model efficiency (NSE) coefficient were chosen to assess the performance of each index. PCC is a number that quantifies the linear relationship between two variables, which is defined as and RMSE is an accuracy estimate, which is given by The NSE was developed by Nash and Sutcliffe [55] and is used to evaluate hydrological models. It is calculated as where represents the observation, is the mean of , and is the estimation. The value of the NSE ranges from −∞ to one. An NSE value higher than zero indicates that the estimation matches well with the observation. The NSE value is negative when the accuracy of the estimation is lower than the accuracy of the observation. Overall, the closer the NSE value is to one, the more reliable the estimation.

Results and Discussion
Both an internal and an external assessment were computed to examine whether these indices (including remote sensing indices, PDSI and three kinds of ENSO indices) are suitable for water level reconstruction. To assess the internal correlation, the correlation between the observed Vam Kenh monthly water level and the remote sensing indices (NDVI and LST) were obtained during the corresponding data curve fitting. The corresponding curves were then used to predict the water level. It was found that both the NDVI-and LST-reconstructed water levels match well with the observed water level (Figure 6). However, the peak water levels between the NDVI-reconstructed water level and observed water level show relatively large differences in the years

Result Assessment Schemes
To conduct a comparison between observed water levels and estimates derived from the indices, the Pearson correlation coefficient (PCC), root mean square error (RMSE), and Nash-Sutcliffe model efficiency (NSE) coefficient were chosen to assess the performance of each index. PCC is a number that quantifies the linear relationship between two variables, which is defined as and RMSE is an accuracy estimate, which is given by The NSE was developed by Nash and Sutcliffe [55] and is used to evaluate hydrological models. It is calculated as where X o represents the observation, X o is the mean of X o , and X m is the estimation. The value of the NSE ranges from −∞ to one. An NSE value higher than zero indicates that the estimation matches well with the observation. The NSE value is negative when the accuracy of the estimation is lower than the accuracy of the observation. Overall, the closer the NSE value is to one, the more reliable the estimation.

Results and Discussion
Both an internal and an external assessment were computed to examine whether these indices (including remote sensing indices, PDSI and three kinds of ENSO indices) are suitable for water level reconstruction. To assess the internal correlation, the correlation between the observed Vam Kenh monthly water level and the remote sensing indices (NDVI and LST) were obtained during the corresponding data curve fitting. The corresponding curves were then used to predict the water level. It was found that both the NDVI-and LST-reconstructed water levels match well with the observed water level (Figure 6). However, the peak water levels between the NDVI-reconstructed water level and observed water level show relatively large differences in the years  On the contrary, the peaks of water level reconstruction derived from LST have smaller disparities with the observed water stage than that of NDVI in the time periods (from January 2000 to December 2004) influenced by the 1998-2001 La Niña, indicating a stronger ability of LST to capture ENSO's influence on the water level when compared to NDVI. Overall, the remote sensing indices can reconstruct the water stages of the lower Mekong basin, but shortcomings exist, particularly in the time period of ENSO events. The vegetation index (NDVI) and temperature (LST) record the localized instantaneous responses, which are the consequences of regional precipitation On the contrary, the peaks of water level reconstruction derived from LST have smaller disparities with the observed water stage than that of NDVI in the time periods (from January 2000 to December 2004) influenced by the 1998-2001 La Niña, indicating a stronger ability of LST to capture ENSO's influence on the water level when compared to NDVI. Overall, the remote sensing indices can reconstruct the water stages of the lower Mekong basin, but shortcomings exist, particularly in the time period of ENSO events. The vegetation index (NDVI) and temperature (LST) record the localized instantaneous responses, which are the consequences of regional precipitation and water storage exchange within a hydrological cycle, but not the cause. This is the reason why the water level reconstructions of remote sensing indices are not well represented. Figure 7a displays the reconstructed water stage based on the PDSI index; the observed water level matches almost perfectly in the study time span except for two periods: from January 1992 to December 1993 and from January 1998 to December 1999. These two periods are squarely close to 1991-1992 El Niño ranging from June 1991 to June 1992 and the 1997-1998 El Niño ranging from May 1997 to May 1998, respectively. Doubtlessly, the differences in the two periods are caused by the two El Niño events separately. The Mekong River Basin lies within the tropical and subtropical latitudes. The ENSO influence on the subtropical region (i.e., upstream) is relatively reduced when compared to the tropical region (i.e., downstream). As a result, the dryness conditions of the PDSI data time series in the upstream of the Mekong River Basin should be more sensitive than that of the downstream for the two above-mentioned mismatched periods. Hence, the PDSI data time series for the upstream Mekong River Basin for the two above-mentioned mismatched periods were chosen to replace the corresponding downstream PDSI. and water storage exchange within a hydrological cycle, but not the cause. This is the reason why the water level reconstructions of remote sensing indices are not well represented. Figure 7a displays the reconstructed water stage based on the PDSI index; the observed water level matches almost perfectly in the study time span except for two periods: from January 1992 to December 1993 and from January 1998 to December 1999. These two periods are squarely close to 1991-1992 El Niño ranging from June 1991 to June 1992 and the 1997-1998 El Niño ranging from May 1997 to May 1998, respectively. Doubtlessly, the differences in the two periods are caused by the two El Niño events separately. The Mekong River Basin lies within the tropical and subtropical latitudes. The ENSO influence on the subtropical region (i.e., upstream) is relatively reduced when compared to the tropical region (i.e., downstream). As a result, the dryness conditions of the PDSI data time series in the upstream of the Mekong River Basin should be more sensitive than that of the downstream for the two above-mentioned mismatched periods. Hence, the PDSI data time series for the upstream Mekong River Basin for the two above-mentioned mismatched periods were chosen to replace the corresponding downstream PDSI.  After replacement, the combined upstream and downstream PDSI (hereafter called ENSO-assisted PDSI) reconstructed water level is almost perfectly in agreement with the observed water level (Figure 7c). There are still slight differences: the peak of the ENSO-assisted PDSI reconstructed water level is a little bit lower than the observed water level peak as a whole. The differences may be attributed to the fact that the PDSI can only consider the moisture within 1-meter subsurface thickness [22], while ignoring the potential presence of water deeper than 1-meter. Overall, ENSO-assisted PDSI can reconstruct water stages very well to monitor droughts and floods, in spite of slight defects.
As ENSO events play an important role in the water level reconstruction of the above remote-sensing indices and PDSI, the water stages using the ENSO indices were reconstructed directly. Figure 8 shows that all ENSO indices performed well, but there are peak overestimations in the years 1999 and 2000. The ENSO-reconstructed results performed better during or after the 1991-1992 El Niño event than the results of or after the 1997-1998 El Niño event. This may be caused by the strong intensity of El Niño in 1997-1998, which had a considerable effect on the reconstruction of the water level. Besides the obvious overestimations in the years 1999 and 2000, other smaller peak disparities display a random pattern. In other words, the ENSO indices both overestimated or underestimated the water level in a random manner. This result may be due to the uneven distribution of rainfall affected by ENSO events. In addition, considering three types of ENSO indices, their degrees of peak overestimation in the years 1999 and 2000 are different. Compared to the SOI (for the atmosphere only), the water level overestimations are much larger when based on SST anomalies (for the ocean only) and MEI (for the coupled atmosphere-ocean). This indicates that the ocean is one of the significant factors causing the overestimation. Perhaps SST and MEI indices take into account some rainfall in the ocean region, which is outside our study region. Therefore, if we can find a way to remove the rainfall beyond our study region, the overestimations should be further mitigated. After replacement, the combined upstream and downstream PDSI (hereafter called ENSOassisted PDSI) reconstructed water level is almost perfectly in agreement with the observed water level (Figure 7c). There are still slight differences: the peak of the ENSO-assisted PDSI reconstructed water level is a little bit lower than the observed water level peak as a whole. The differences may be attributed to the fact that the PDSI can only consider the moisture within 1-meter subsurface thickness [22], while ignoring the potential presence of water deeper than 1-meter. Overall, ENSOassisted PDSI can reconstruct water stages very well to monitor droughts and floods, in spite of slight defects.
As ENSO events play an important role in the water level reconstruction of the above remotesensing indices and PDSI, the water stages using the ENSO indices were reconstructed directly. Figure 8 shows that all ENSO indices performed well, but there are peak overestimations in the years 1999 and 2000. The ENSO-reconstructed results performed better during or after the 1991-1992 El Niño event than the results of or after the 1997-1998 El Niño event. This may be caused by the strong intensity of El Niño in 1997-1998, which had a considerable effect on the reconstruction of the water level. Besides the obvious overestimations in the years 1999 and 2000, other smaller peak disparities display a random pattern. In other words, the ENSO indices both overestimated or underestimated the water level in a random manner. This result may be due to the uneven distribution of rainfall affected by ENSO events. In addition, considering three types of ENSO indices, their degrees of peak overestimation in the years 1999 and 2000 are different. Compared to the SOI (for the atmosphere only), the water level overestimations are much larger when based on SST anomalies (for the ocean only) and MEI (for the coupled atmosphere-ocean). This indicates that the ocean is one of the significant factors causing the overestimation. Perhaps SST and MEI indices take into account some rainfall in the ocean region, which is outside our study region. Therefore, if we can find a way to remove the rainfall beyond our study region, the overestimations should be further mitigated.  The observed water stage for all the above results comes from the Vam Kenh station in the Mekong delta. The same procedure was applied to the Dinh An station independently, so as to demonstrate the applicability of our method in this study. To internally evaluate the reconstruction of various indices at two stations, the above-mentioned three assessment metrics, namely PCC, RMSE, and NSE in Section 3, were chosen as the performance criteria. In general, the reconstructed water level from the ENSO-assisted PDSI and ENSO indices performed better than the remote sensing indices (i.e., NDVI and LST), no matter whether in terms of the PCC, RMSE, and NSE (Table 2). This indicates that the causal information, based on the PDSI and ENSO, improves water level reconstruction. For the same index, the result for Vam Kenh is better than that for Dinh An, to which the topography and the proximity to the ocean may be attributed. The observed water stage for all the above results comes from the Vam Kenh station in the Mekong delta. The same procedure was applied to the Dinh An station independently, so as to demonstrate the applicability of our method in this study. To internally evaluate the reconstruction of various indices at two stations, the above-mentioned three assessment metrics, namely PCC, RMSE, and NSE in Section 3, were chosen as the performance criteria. In general, the reconstructed water level from the ENSO-assisted PDSI and ENSO indices performed better than the remote sensing indices (i.e., NDVI and LST), no matter whether in terms of the PCC, RMSE, and NSE (Table 2). This indicates that the causal information, based on the PDSI and ENSO, improves water level reconstruction. For the same index, the result for Vam Kenh is better than that for Dinh An, to which the topography and the proximity to the ocean may be attributed. Among the three ENSO indices, the SST for ocean only performed best, the MEI for both ocean and atmosphere was second, and the SOI for the atmosphere only had the worst result. This implies that different ENSO indices may show different performances when projecting the hydrological conditions of a given basin, and the oceanic ENSO index based on SST is relatively effective in the Mekong basin.
In order to externally assess the suitability of our method, we based it on the constructed relationship between the water level from the Vam Kenh station and indices, and a scale factor simply obtained from the ratio of water discharge from two separated estuaries to predict the water level at the Dinh An station (hereafter called "Dinh An water level prediction" to distinguish it from "Dinh An water level reconstruction"). The result of the Dinh An water level prediction for the PDSI and ENSO-assisted water level prediction was slightly less accurate than the Dinh An water level reconstruction, while the prediction based on ENSO indices outperformed that of the reconstruction. However, the performance of the Dinh An water level prediction is still well represented (Table 2). Overall, our approach can not only reconstruct the water level at the gauged stations, but also predict the water level of an ungauged station based on information from another gauged station.

Conclusions
Contrary to the traditional usage of localized instantaneous measurement from remote sensing, we explore the feasibility of reconstructing the water level in the lower Mekong delta using PDSI, and three ENSO indices on a monthly temporal scale, for the reason that the causal information can serve as a precursor of localized instantaneous measurement. We found that the time lag (i.e., six months for 1991-1992 and nine months for 1997-1998) between standardized water levels and the PDSI and ENSO indices during the El Niño event is an important component for a better water level reconstruction, due to its ability to capture hydrological extremes.
Compared to remote sensing indices (i.e., NDVI and LST), PDSI-and ENSO-reconstructed water levels perform better. The combination of upstream and downstream PDSI (i.e., ENSO-assisted) and ENSO information obtained the best-reconstructed water level, which is fairly satisfactory with a correlation coefficient around 0.95, with an RMSE value ranging from 0.23 to 0.42 m and an NSE around 0.90. This compares to the reconstructions based on remote sensing with a correlation coefficient less than 0.90, with an RMSE value ranging from 0.40 to 0.79 m and an NSE less than 0.81, respectively. This result indicates that the potential usage of ENSO information could lead to an improvement in drought-index-based water level reconstruction and prediction. It is anticipated to be applicable to large basins affected by the ENSO phenomenon and hydrological extremes.
Notice that different ENSO indices may show different performance when projecting the hydrological conditions depending on the river basins, as El Niño poses a different magnitude of effect for different river basins. A holistic assessment for different river basins is yet necessary to further validate these results in the near future. Further improvement may lie in better temporal resolution of the PDSI and ENSO indices with good accuracy. To further extend our proposed method, an attempt can be made to use the environmental indices derived from climate model outputs for future water level projections. This can potentially help investigate whether the overestimation of potential evapotranspiration (PET) [56,57] has a substantial influence on water level reconstruction.