The Mekong is a transboundary river flowing across Southwest China and Southeast Asia (Fig. a). The river originates from the Tibetan Plateau at an altitude of about 5200 m a.s.l. and flows in a northwest–southeast direction through six countries (China, Myanmar, Laos, Thailand, Cambodia, and Vietnam) before pouring into the East Vietnam Sea. The Mekong drains an area of 795 000 km2 with an average annual discharge of approximately 475 km3. Its upper portion is 2140 km long and drains an area of 176 400 km2. The high mountains and low valleys characterizing the Lancang River basin contribute to the spatial variability of precipitation, whose annual average varies from 750 to 1025 mm across the basin. Precipitation is also unevenly distributed across the year, with two distinct dry (December to May) and wet (June to November) seasons. The streamflow reflects a similar seasonal pattern . Although the drainage area of the Lancang River accounts for about 22 % of the total catchment area, the Lancang contributes only to 16 % of the average annual discharge of the whole Mekong River .

The advantageous topography and abundant water availability make the Lancang River basin an ideal spot for the hydropower industry . The first dam on the mainstream of the Lancang (Manwan) began its operations in 1992, followed by Dachaoshan in 2003 and Jinghong in 2008. The two largest dams (Xiaowan and Nuozhadu) became operational in 2009 and 2013, respectively. And since 2016, at least one dam has joined the Lancang's reservoir system every year. Overall, this rapid transformation of the basin resulted in a system comprising 11 operational and 1 planned dam (Fig. b).

The design of the cascade reservoir system reflects the topographic characteristics of the basin. Specifically, the presence of narrow valleys with steep sides required the construction of high dams (see Fig.  and the list of design specifications in the Supplement Table S1). In turn, this resulted in reservoirs with large storage capacity relative to inflow, steep banks, and long and horizontally narrow shapes. The total storage capacity is 42 170 Mm3, about 55 % of the average annual discharge at Chiang Saen gauging station, the first downstream station with publicly available data (Fig. ). These reservoirs form a long and complex cascade system, so it is only by studying it in its entirety that we can understand how storage operating patterns have evolved over the past decade.

In this study, we focus on the 10 largest operational reservoirs (each with a volume larger than 100 Mm3), all located on the main stem of the Lancang River. We select 2008–2020 as our study period because it includes the year of commission of most dams (8 out of 10), a choice that allows us to study their operations during the filling period as well as under regular operating conditions. Extending the temporal horizon to include the year of commission of the two remaining dams (Manwan and Dachaoshan, commissioned in 1992 and 2003) would complicate the analysis unnecessarily, since their aggregated storage capacity corresponds to only 2.14 % of the current total system capacity. For the aforementioned study period we gathered data on the digital elevation model (DEM), satellite imagery, and radar altimetry water level.

Recall that for 9, out of 10, reservoirs, the SRTM-DEM can provide full information on bathymetry (Sect. ). To estimate the E–A curve of these reservoirs, we first isolate the DEM data with the contour corresponding to maximum water level and dam crest line. The purpose of this step is to calculate the curve within the extent of the reservoir only and thus avoid errors due to the inclusion of surrounding areas. Then, we calculate the surface area corresponding to each 1 m elevation of the DEM. Specifically, with each elevation value (each meter) from the lowest elevation within the reservoir extent to the maximum water level, we count the number of pixels having a value equal to or smaller than that elevation value. This is because, when water reaches that elevation, the area corresponding to those pixels is inundated. Then, we multiply the number of pixels by the pixel size (30 m × 30 m) to get the water surface area. We finally fit a fifth-degree polynomial (degree determined by trial and error) to the data points so obtained. For the remaining reservoir, Manwan, we apply the same procedure but only to the portion above the water level recorded by the SRTM. To approximate the E–A curve below that water level, we fit a fifth-degree polynomial to the part above the water surface and then extend it below the water surface, as in and .

With the E–A curve at hand, we calculate the storage volume corresponding to each 1 m elevation of the DEM. This operation is carried out using the following trapezoidal approximation : 1Vi=∑j=l+1i(Aj+Aj-1)(Ej-Ej-1)/2, where Vi is the storage volume corresponding to the water level Ei and water surface area Ai, and l denotes the lowest elevation of the reservoir bathymetry (i.e., Al=0). The trapezoidal approximation is used here instead of a direct calculation from the DEM because the latter is not applicable to Manwan – while it is desirable to minimize the differences in data processing for all reservoirs. In addition, with the E–A curves validated by water level observations (from altimetry data) and water surface area (from Landsat images), we can confidently develop the E–S and A–S curves from the E–A curves using the trapezoidal approximation (recall we do not have observed storage data to validate the E–S and A–S curves estimated directly from the DEM). To strengthen the rationale for using the trapezoidal approximation, we compare the results obtained with the two methods. The differences in storage corresponding to each water level in the variation range are not more than 1 % (for Jinghong, Miaowei, Huangdeng, and Wunonglong) and 2 % (for Nuozhadu, Dachaoshan, Xiaowan, Gongguoqiao, and Dahuaqiao). The detailed comparisons for Nuozhadu and Xiaowan reservoirs can be found in Table S4 and Fig. S4. Finally, we use the data points on storage volume to fit the A–S and E–S curves. All aforementioned operations are carried out in Python 3.7 with the aid of the OSGeo library.

Water surface data can be inferred from Landsat images by classifying each pixel with either a single spectral band (e.g., near-infrared band) or a spectral index calculated from multiple bands (see Table S5 for a list of the most common indices). In general, the use of a single spectral band reduces the computational requirements , but spectral indices tend to provide more robust results . Whatever the method used, one key challenge with Landsat images stands in the presence of clouds, cloud shadow, and no-data pixels (for Landsat 7), which may lead to a misclassification of water pixels and the consequent underestimation of the water surface area. To handle this problem, we use a novel variant of the WSA estimation algorithm introduced by and , originally conceived to extract water surface area from the normalized difference vegetation index (NDVI) layer – which is included in the 250 m resolution global Terra MODIS vegetation indices (MOD13Q1), a level-3 MODIS product provided by NASA.

Like the modified version by , our algorithm consists of two main phases: mask creation and water classification improvement, illustrated in Fig.  with light blue and light green boxes. In the first phase, the cloudless images are processed together to create two products: the expanded mask and zone mask. The two masks are then used in the second phase, where the Landsat images are individually processed to obtain the water surface area corresponding to the collection time of each image. The major modifications with respect to the version by are the selection of cloudless images (Step 1.1) and identification of additional water zones (Step 2.5), two modifications needed to ensure that the algorithm performs well with Landsat images (instead of the NDVI layer of MOD13Q1). Further details for each phase and step are provided below.

(1.1) Selection of cloudless images. Cloudless images are the ones that do not contain clouds or contain very little clouds on the reservoir surface extent. For our application, we define a cloudless image as an image with less than 20 % of cloud cover on the maximum reservoir surface extent. To identify these images, we use the BQA band (the band of quality assessment), which contains the information on cloud pixels. As we shall see, working on a subset of cloudless Landsat images is necessary to preserve the quality of the frequency map and masks produced in the next steps. Note that the version by did not include this step because cloud effects are partially removed from the NDVI layer in MOD13Q1 . This is the result of selecting the best available pixel value (the low clouds and the highest NDVI value) from all daily acquisitions within a 16 d period.

(1.2) NDWI-based classification. To classify the water and non-water pixels, we use the normalized difference water index (NDWI) with a threshold value equal to 0. The choice of index and corresponding threshold is based on a preliminary analysis, in which we compared the performance of NDWI, NDVI, and MNDWI (modified normalized difference water index) for all 10 reservoirs. The results, reported in Figs. S5–S9 for 60 cloudless Landsat images for each reservoir, show that the NDWI-based classification matches the maximum water extent reported in the Maximum Water Extent dataset, developed by the European Commission’s Joint Research Centre . On the other hand, NDVI and MNDWI tend to provide less reliable results. As for the threshold value, 0.05 and 0.1 (for NDWI) tend to lead to an underestimation of the water pixels, since the total number of times a water pixel is correctly classified as water is less than 60. We also manually checked the obtained water layers with the true-color Landsat images for a number of images before making our decision of using NDWI. The NDWI layers calculated this way are subsequently used in Step 2.2.

(1.3) Frequency map creation. To create the frequency map, we first calculate the percentage of times in which a pixel is classified as water (based on its NDWI value) in all selected cloudless images. This operation is carried out for all pixels within the bounding box of the reservoir extent. Then, we create the frequency map by selecting the pixels with frequency larger than 0. This step is illustrated in Fig. a, b.

(1.4) Frequency map expansion. We expand the frequency map by buffering it with three additional pixels; in other words, we add three pixels around the peripheral water pixels (see Fig. a, b). The expansion is aimed to ensure that no possible water pixels are missed out. This 90 m buffer around the nominal shoreline is deemed sufficient for our case study, since reservoirs in the Lancang are located in steep terrains, where the storage is controlled by elevation more than area. The expanded frequency map is used in Step 2.2 to clip the NDWI layer; hereafter, we refer to it as the expanded mask.

(1.5) Zone mask creation. In the last step of Phase 1, we convert the frequency map into a 50-zone mask. As illustrated in Fig. c, the ith zone contains the pixels classified as water with a frequency greater than 2⋅(i-1) % and less than or equal to 2⋅i % (with i=1,…50). For example, Zone 1 contains the pixels classified as water from 0 % to 2 % of the time, while Zone 2 contains those classified as water from 2 % to 4 % of the time. At the end of this phase, we obtain the two inputs for the next phase, that is, the expanded mask and zone mask.

(2.1) NDWI calculation. Here, we calculate the NDWI index for the remaining Landsat images – with clouds, cloud shadow, and no-data pixels – and pass them to the next step in the form of a raster layer for each image. Note that the goal of this second phase is to improve the water surface classification of the images, so as to maximize the number of data points available for our study period, especially for the monsoon season when Landsat observations are heavily affected by clouds. Failing to improve the cloudy images can make the water surface area estimates inaccurate.

(2.2) Clipping the NDWI layer by the expanded mask. The NDWI raster layer obtained in Steps 1.2 and 2.1 is clipped by the expanded mask created in Step 1.4.

(2.3) The k-means-based classification of the water pixels. Because of the presence of clouds, and other disturbances, the of use of the same NDWI threshold (equal to 0) in all Landsat images may lead to overestimation or underestimation errors of the water surface area. To find NDWI thresholds for each Landsat image, we resort to k-means clustering. Specifically, we set k equal to 3 (a value found by trial and error) and apply k-means clustering to all pixels in the NDWI layer (Fig. a). Water pixels tend to fall into the cluster with the highest NDWI value because the NDWI of water pixels has a higher value than the one of non-water pixels. Results are verified by manually checking the classified water layer with true-color Landsat images.

(2.4) Water fraction calculation (by zone). The zone mask created in Step 1.5 is used here to divide the water extent layer (obtained in the previous step) into 50 zones. For the ith zone, we define the water fraction pi as follows: 2pi=ni/Ni,i=1,2,…,50, where pi represents the ratio between the number ni of pixels classified as water in zone i (with the NDWI-based k-means clustering) and the total number Ni of pixels in zone i (retrieved from the zone mask). The information provided by the water fraction of each zone is used in the next step to improve the water pixel classification.

(2.5) Identification of additional water zones. We improve the classification of water pixels by identifying the additional water zones based on their water fraction. To do so, we resort again to the k-means clustering algorithm. Moreover, because of the continuity of water extent (water expands from higher frequency to lower frequency zones), we also take into account the zone number (or frequency value). Then, we formulate a clustering problem in a two-dimensional space constituted by water fraction and zone number. We solve the clustering problem with a value of k equal to 2, found by trial and error. Figure b and c show two examples with k=2, while Fig. d reports an example for an unsuitable value of k. The lowest zone in the higher cluster (zone 14 in Fig. b and zone 31 in Fig. c) is the threshold above which zones are converted to water pixels. The reason for converting all pixels from the threshold zone onwards to water pixels is that the pixels in the same zone have the same (or very similar) inundation probability, and, at each observation time, they fall into one of two scenarios: (1) they are both non-water pixels, or (2) they are both water pixels (even when the water fraction of that zone is less than 100 %, due to cloud cover). Naturally, there can be a small error from the threshold zone. For example, Zone 14 in Fig. b contains pixels with 26 %–28 % inundation probability, but sometimes, the threshold of inundation probability is not exactly 26 % (e.g., 26.5 %, 27 %, …). Note that we could increase the performance by dividing the frequency map into a larger number of zones, but this would require a larger number of cloudless images. This step represents the second modification of the original WSA estimation algorithm, which uses a quality parameter not suitable for Landsat images – since the cloud effects are not mitigated, unlike the NDVI layer in MOD13Q1.

(2.6) Overlapping. Finally, the layer of additional water pixels is overlapped on the layer of water extent obtained in Step 2.3. The final output is the improved water classification for each image characterized by cloud cover or other disturbances. All the aforementioned operations are carried out in Python 3.7 with the aid of the OSGeo and SKLearn libraries.

Determining the filling strategy of a reservoir means deciding the rate with which the reservoir is filled and, therefore, the fraction of inflow that is retained on a periodic basis – monthly, in our case. The problem is formalized by the following mass balance equation: 3St=St-1+θ⋅Qt-Et, where St is the reservoir storage at time t, Qt the inflow volume in the interval (t-1,t], Et the evaporation loss in the interval (t-1,t], and θ a parameter varying between 0 and 1 and expressing the fraction of inflow volume retained by the reservoir. In our case, the goal is to determine the value of θ (in each month) for Nuozhadu and Xiaowan during the periods 2012–2013 and 2009–2010, respectively.

Observed inflow data are not available, so we resort to modeled ones. Specifically, we use daily inflow data simulated by VIC-Res , a variant of the variable infiltration capacity (VIC) model – a large-scale, semi-distributed hydrological model first developed by . Similarly to VIC, VIC-Res contains a rainfall–runoff module and a routing module. In the first module, the study region is organized as computational cells with a customizable cell size (0.0625∘ in our case), in which key hydrological processes (evapotranspiration, infiltration, baseflow, and runoff) are calculated as a function of hydro-meteorological forcings (precipitation, temperature, wind speed, etc.) and soil parameters. Then, the routing module routes the simulated runoff and baseflow throughout the river network using the linearized Saint-Venant equation . In VIC-Res, the routing process includes a detailed representation of water reservoir operations: for each reservoir in a given study site, the model calculates the mass balance and release, with the latter determined by operating rules or predefined release time series. VIC-Res has been tested on several sites, including the Lancang River basin (over the period 1996–2005). In particular, reported the results of model calibration with observed discharge at Chiang Saen station and model validation with observed discharge at Jiuzhou station, located right upstream of Xiaowan reservoir.

Here, we use the same VIC-Res model, which we force with rainfall and temperature (maximum and minimum) data retrieved from CHIRPS-2.0 and ERA5 dataset. Land use and cover data were obtained from the Global Land Cover Characterization dataset, while the soil data were extracted from the Harmonized World Soil Database. The monthly leaf area index and albedo were derived from the Terra MODIS satellite images, which provide changes in canopy and snow coverage over time. The flow direction map used by VIC-Res is based on the SRTM-DEM. Since considered the dams built before 2005 and used the rule curves proposed by , we slightly adapt the model to handle two challenges for our current study: (1) consider more reservoirs (all dams on the mainstream built until 2013) and (2) leverage the actual storage data retrieved from satellite data. To set up VIC-Res in our study, we therefore proceeded as follows. For each reservoir, we take data on inflow (simulated), storage (estimated from the satellite data), and evaporation (simulated using the estimated water surface area and evaporation rates calculated with the Penman equation). We then invert the mass balance equation to calculate the release, which is used as input to VIC-Res to simulate the storage dynamics of each reservoir. The process is repeated sequentially – starting with the most upstream dam – so as to ensure that the cascading impacts of dams are captured correctly. To ensure the reliability of this analysis, we extend the model validation at Chiang Saen for the filling period of Nuozhadu and Xiaowan (2009–2013); see Fig. S10. The comparison between modeled and simulated storage is reported in Fig. S11.

The availability of storage data also allows us to decipher the impact of dam operations on downstream (measured) discharge. To do that, we calculate two time-varying indicators of hydrological alteration (I1 and I2). I1 represents the fraction of the natural flow retained in the reservoir system for each month in which the system is storing water (i.e., when ΔSt=St-St-1>0). I2 represents the fraction of the actual flow released from the reservoir system for each month in which the system is releasing water (i.e., ΔSt<0). I1 and I2 are calculated as follows: 4I1,t=ΔStΔSt+Qt,5I2,t=ΔStQt, where Qt is the observed discharge volume downstream of the reservoir system (at Chiang Saen) in month t. Note that the denominator in Eq. () approximates the natural flow in month t (it is the sum of actual discharge and volume of water retained upstream in a given time interval).

The E–A curves of Nuozhadu and Xiaowan reservoirs are illustrated in Fig.  (panels a and d), where the blue circles represent the data points derived from the DEM, and the light blue lines are the fifth-degree polynomials fitted to them. Note that both curves correctly intersect the point identified by maximum water level and maximum water surface area, retrieved from . A similar evaluation is carried out for the A–S and E–S curves (Fig. , panels b, c, e, f) but this time using design specifications on full storage (A–S and E–S curves) and dead storage (E–S curves).

We carry out an additional validation of the E–A curves by comparing them against observations of water level and surface area obtained from radar altimetry data and Landsat imagery. These observations, illustrated in Fig. a and d by cyan diamonds, follow the curves identified through the DEM closely. Naturally, the cyan points are primarily concentrated between the dead and maximum water levels, which denote the normal range of operating conditions. As we shall see later, points below the dead water level correspond to the dam filling period.

The E–A, A–S, and E–S curves of the remaining eight reservoirs are reported in Figs. S12 and S13. Apart from the curves of Jinghong and Huangdeng, which are evaluated using the radar altimetry water level, the curves of the other reservoirs can only be evaluated by comparing them against the design specifications reported by . Such an evaluation is only partially successful, since we did not find a perfect match between curves and design specifications for Jinghong, Gongguoqiao, Miaowei, Dahuaqiao, and Wunonglong reservoirs. Considering that the procedure used to estimate the curves has been successfully employed in several studies , we suspect that the reason behind the mismatch may lie with the information on dam design specifications available to the public. In turn, this reinforces the need for research aimed to retrieve data on large-scale infrastructure in transboundary river basins. We also note that this source of uncertainty does not severely affect our study, since those five reservoirs account for a small fraction of the total system's storage (2.36 %, 0.74 %, 1.55 %, 0.69 %, and 0.64 %, respectively).

Recall that the WSA estimation algorithm builds on the idea of using cloudless images to create the expanded mask and zone mask, which are then employed to correct the classification of water pixels in images affected by clouds and other disturbances. In our case, such improvement is needed for 56 % of the 3004 Landsat images available for our study period (the number of usable images increases from 26 % to 82 %). As one might imagine, the classification correction is particularly important during the wet season, when cloud cover is more frequent – the number of usable images increases by 54 % of 1770 images (from 30 % to 84 %) in the dry season and 58 % of 1234 images (from 21 % to 79 %) in the wet season. The performance of the algorithm for each reservoir is summarized in Table S6.

The WSA time series of Nuozhadu and Xiaowan reservoirs are reported in Fig. . The first result to note is the stark change in the WSA values before (light blue points) and after (cyan points) the classification improvement. The time series of corrected WSA values also starts to reveal the reservoirs' operating patterns: the sharp increase beginning in 2012 (Nuozhadu) and 2009 (Xiaowan) denotes the starting point of the reservoir filling period, while the large, annual, fluctuations suggest the presence of a broad range of operating conditions – the maximum surface area is reached only at the end of the wet season, while the rest of the year seems to be used to fill in and empty the reservoirs. In Sect. , we will see how such variability translates into storage patterns.

To evaluate the results obtained with Landsat imagery, we leverage the radar altimetry water level data and E–A curves to obtain two independent WSA time series. As shown in Fig. , both modeling approaches provide very similar results. The same outcome can be seen in the WSA time series of Huangdeng and Jinghong reservoirs (see Fig. S14). We also provide a quantitative comparison of Landsat-derived and altimetry-converted WSA for all four reservoirs mentioned above (see Table S7). With this additional analysis we therefore serve two purposes: scrutinize the WSA values for the main reservoirs and empirically validate the approach based on Landsat imagery, the only one available for the remaining reservoirs.

Having established how the reservoirs in the Lancang River basin have been filled in and operated, we can finally explain their time-varying influence on the discharge measured at Chiang Saen (Sect. ). The graphical analysis of total storage and discharge (Fig. c) highlights the stark changes in the flow regime in response to the increase in upstream storage. The flow regime changed drastically in late 2013, when the filling of Xiaowan and Nuozhadu was completed. By discharging water during the dry season and retaining it in the wet season, the hydropower dams largely increase low flows and decrease high flows (Table S8). For example, the mean of the annual peak discharge decreased from 11 157 (1990–2008) to 6186 m3 s-1 (2013–2020) (-45%), while the mean of the annual lowest discharge grew from 638 to 1003 m3 s-1 (+57%). Similar figures are found for other statistics (Table S8). We can also note a macroscopic change in the seasonal discharge pattern, from ample annual fluctuations to more rapid flow changes. All these observations are confirmed by the wavelet analysis reported in Fig. S17.

As shown in Fig. b, the degree of flow alteration at Chiang Saen caused by the Lancang's dams increased significantly over time with three distinct stages: the first stage (before Xiaowan reservoir began operating), the middle stage, and the last stage (after Nuozhadu reservoir began operating). That means the range of variability of I1 (in black color) and I2 (in red color) increased over time: [0, 0.04] and [-0.10, 0] in the first stage, [0, 0.20] and [-0.44, 0] in the second stage, and finally, [0, 0.50] and [-0.91, 0] in the last stage. With the number of reservoirs increasing rapidly in the last decade, the downstream discharge became increasingly controlled by dam operations.

By bringing the monthly precipitation anomalies (for the Lancang River basin) into the overall picture (Fig. a), we can better understand how dam operations partially contributed to downstream droughts and pluvials. A case in point is the drought in the period 2019–2020. The monthly precipitation anomalies show that, in the wet season of 2019, the Lancang River basin received less precipitation, especially in May and June (around 50 mm less than the average for those months). However, the values of I1 indicate that the reservoir system kept retaining part of the inflow during the central months of the year (up to about 46 % in October). Because of such a combination of meteorological drought and dam operating strategies, the downstream area underwent a critical dry period, with Chiang Saen gauging station recording extremely low flows during the summer months . The release of water during the subsequent dry season only partially alleviated the effect of the ongoing drought, since the negative precipitation anomaly persisted until mid-2020. Importantly, the 2019–2020 data suggest that the dam operating strategy was not largely affected by the meteorological conditions: the Lancang dams currently store about 46 % of the estimated natural flow during the wet season (regardless of the monsoon's intensity) and then discharge it during the dry one, controlling up to 89 % of the dry-season flow – a pattern that has emerged since Xiaowan and Nuozhadu became fully operational. These results also highlight the importance of emergency releases from the upstream reservoirs, such as those that were implemented in 2016 . In regard to those emergency releases, it should be noted that the 2016 drought had much smaller magnitude than the 2019–2020 one and that the drought occurred during the first half of the year, when the reservoir system was releasing water following its normal operations. In sum, the availability of (inferred) storage data can help us put droughts, emergency releases, and pluvials into a broader perspective. Yet, such analyses should ideally be complemented by more significant data-sharing efforts among all riparian countries, a point that was recently reinforced by .