Introduction

Climate change, while a globally recognized challenge, manifests differently across regions and ecosystems, with some systems particularly vulnerable1. One critical aspect of climate change is its impact on glacial systems, as changes in meltwater volume and timing alter hydrological regimes, triggering cascading effects on adjacent regions. These shifts can exacerbate climate feedback loops, further intensifying environmental challenges. Since 1980, the Tibetan Plateau has warmed by 0.42 °C per decade, and precipitation has increased by 11 mm, both of which have been major drivers of lake expansion and accelerated river runoff2,3,4,5. This warming has caused rapid glacier and snowpack retreat, with glacier meltwater accounting for 9% of lake expansion between 1995 and 20206.

These dramatic changes have disrupted the balance of the Tibetan Plateau, also known as Earth’s “Third Pole” or “Asian Water Towers”, with a rapid shift from solid to liquid water reserves7. The destabilization of this system—manifested in the escalating risk of glacial lake failure, which threatens critical infrastructure8,9—has emerged as a significant scientific challenge. It is also a critical issue for the sustainable development of both the plateau and downstream regions9,10,11. The importance of these river systems is underscored by their profound impact on the billions of people and diverse ecosystems that depend on them across Asia, underscoring the urgency of understanding these changes12.

The hydrological reorganization events experienced on the cryospheric and alpine systems are dramatically reshaping landscapes and basin structures8,10, and have evolved into climate-driven cascading effects. In 2016, in the Yukon Territory of Canada, meltwater runoff from the Kaskawulsh Glacier shifted within a short period from a northward-flowing, closed-basin system to a southward-flowing, external drainage system, causing Kluane Lake and Slims River to lose the main inflow and altering the basin’s original hydrological regime8. On the Tibetan Plateau, similar reorganization processes are reshaping the regional hydrological pattern10,13,14. The Zonag and Yanhu Lake drainage basins are a typical example, located in the hinterland of the first World Natural Heritage Site on the Tibetan Plateau—the Hoh Xil region9,13 (Fig. 1a). This study documents and analyses the impact of the connection between the Zonag and Yanhu Lake inland water system (inland lake) and the Yangtze River outflow system (outflow river) on regional hydrology and ecosystems, driven by both natural failure and artificial drainage15,16. This study seeks to enhance our understanding of the interconnected impacts of climate change and hydrological evolution, providing insights into risk mitigation and sustainable management strategies for high-altitude regions.

Fig. 1: Study area.
figure 1

a The location of the Tibetan Plateau and the Zonag and Yanhu Lake drainage basins. b Zonag Lake and Kusai Lake, Haidingno’er Lake and Yanhu Lake were separate basins before the dam failure of Zonag Lake in 2011; the light blue areas indicate glaciers. c After the dam failure of Zonag Lake in 2011, the separate basins were merged into a single basin, and three waterways marked by dark blue lines were formed between the four lakes; an artificial drainage channel marked by red was constructed in 2019 between Yanhu Lake and the Qingshui River, which is marked by purple.

Full size image

Results and discussion

Hydrological reorganization history of the Zonag and Yanhu Lake drainage basins

Before 1995, Zonag Lake, Kusai Lake, Haidingno’er Lake, and Yanhu Lake existed as independent hydrological units (Fig. 1b). Observational data show that between 1996 and 2010, Zonag Lake expanded rapidly, with a direct consequence of this growth being the shoreline breach event in September 2011, triggered by a heavy precipitation10,17. The resulting flood runoff reshaped the regional hydrological pattern, creating an eastward flow pathway that connected Kusai Lake, Haidingno’er Lake, and Yanhu Lake, leading to expansion of the four-lake system18,19 (Fig. 1c). This event not only serves as a landmark example of climate change driving lake breaches on the plateau18,20, but also profoundly altered the regional hydrological landscape and river network structure, ultimately forming the Zonag Lake-Yanhu Lake basin water system10,21.

The combined effects of increased precipitation and glacier melt due to climate warming are the primary drivers behind the abnormal rise in Zonag Lake’s water level13,22. The chain reaction triggered by the breach event had particularly significant consequences in the Yanhu Lake basin: after 2011, the lake continued to expand22, and the risk of overflow posed a direct threat to critical infrastructure such as the Qinghai-Tibet Highway, Railway, and oil and gas pipelines, while also impacting the ecological security of the Yangtze Rivers’ northernmost source through hydrological connectivity22. Notably, this evolving risk highlights the vulnerability of the plateau’s endorheic basin hydrological system to climate change: when natural regulation mechanisms fail, local water disasters can trigger cascading amplification effects throughout the hydrological network.

To counteract these growing risks, the Yanhu Lake drainage diversion project, implemented in 2019, established an artificial waterway between Yanhu Lake and Qingshui River10. This engineering measure achieved two key objectives: it mitigated the risk of overflow and breach in Yanhu Lake, and it restored the hydrological connection between the Zonag and Yanhu Lake drainage basins and the northernmost source of the Yangtze River.

Lake water changes caused by hydrological reorganization

The elevation difference between Zonag Lake to the entrance of the artificial waterway that was built in 2019 of Yanhu Lake is approximately 284 m. This topographic gradient provides the fundamental geographic conditions for the hydrological connectivity of the four lakes in the basin (Fig. 2a). Lu et al.10 calculated the water areas of Zonag Lake, Kusai Lake, Haidingno’er Lake, Yanhu Lake (Supplementary Table 1) and the glacier areas within the basins (Supplementary Table 3) between 1986 and 2019. They found that the four lakes experienced two distinct phases of change: from 1986 to 2011 and from 2011 to 2019. After the dam failure in 2011, the water area of Zonag Lake declined dramatically, while the other three lakes expanded sharply. Although the dam failure caused short-term mutations, regression analysis showed that the slope of the area change trend of the four lakes was positive in both the 1986–2011 and 2012–2019 periods, indicating overall growth in each stage. In contrast, between 1986 and 2019, glacier area in the basin continued to shrink, while glacial meltwater increased year by year (Supplementary Table 3).

Fig. 2: Topographic profiles of streams in the Zonag and Yanhu Lake drainage basins, and hydrological status of the lakes (1986–2024).
figure 2

a Black arrows indicate the direction of runoff from the waterways connecting the four lakes; red lines indicate the digital elevation model (DEM) of the topographic profiles along the waterways. b–e Water area of Zonag Lake, Kusai Lake, Haidingno’er Lake, and Yanhu Lake from 1986 to 2024, with water area data split into three time periods: 1986–2011 (gray), 2012–2019 (red), and 2020–2024 (blue). f Total water area (square mark) and total water volume (triangular mark) in the Zonag and Yanhu Lake drainage basins from 1986 to 2024; the red dashed line indicates the trend.

Full size image

This study utilized satellite monitoring and model estimation to extend the calculation of water area and volume, glacier area and glacial meltwater for the years 2020–2024 (Supplementary Tables 1, 3). The results show that significant changes in the water area of each lake occurred after the drainage diversion project in 2019. From 2020 to 2024, although the water areas of Zonag and Kusai Lakes showed an increasing trend (Fig. 2b, c), the water areas of Haidingno’er Lake and Yanhu Lake continued to decline due to water discharge from the diversion project (Fig. 2d, e). Consequently, the total lake water area and water volume in the basin decreased, with average annual decline rates of −0.33% and −0.66%, respectively. The total water area decreased from 784.97 km² to 771.86 km², and the total water storage dropped from 15.12 billion m3 to 14.62 billion m3 (Fig. 2f). However, when viewed over the long-term changes from 1986 to 2024, the positive slopes in the regression analysis indicate that the total water area and volume of lakes within the basin have continued to increase (Fig. 2f), with average annual growth rates of 0.83% and 0.95%, respectively. Over the same period, glacier area has consistently decreased, with an average annual reduction of 0.33%, while glacier meltwater volume has shown a steady upward trend, with an average annual growth rate of 2.51%.

Correlation matrix analysis revealed the potential driving role of meteorological factors in the changes in the glacier-lake system under climate warming (Fig. 3). The high negative correlation between lake area and glacier area of r = −0.90 explained that glacier retreat is the direct driving factor of lake expansion. The r-value between temperature and glacial meltwater reached 0.96, which indicates that temperature is a key factor influencing glacial retreat and the increase in lake area and water volume. In addition, precipitation and lake area exhibited a positive correlation with r = 0.58. Although the correlation coefficient is weaker than that of temperature, it also indicates its influence on changes in lake water volume.

Fig. 3: Spearman correlation matrix.
figure 3

Correlation analysis between lake area, precipitation, temperature, glacier area, and glacial meltwater, gray indicates lower correlation, while blue indicates higher correlation.

Full size image

The impact of heavy precipitation on reorganized river-lake systems

Based on an understanding of the long-term evolutionary trends of the hydrological status of the Zonag and Yanhu Lake drainage basins, a further focus is placed on the short-term driving effects of extreme precipitation events on lake dynamics. In May 2024, the Zonag Lake region experienced heavy precipitation. The monthly precipitation reached 88.6 mm, far exceeding the average value of 20 mm for May between 1986 and 2023, with an increase of 195%. The annual average precipitation was as high as 543.1 mm, the highest for the same period in the past 38 years (Fig. 4c). This event caused the water level of Zonag Lake to rise from 4739.9 m on 22 May to 4740.3 m on 15 July, before declining sharply to 4739.6 m by 31 October (Fig. 4e). Overall, the water level has shown a steady downward trend after 2016 (Fig. 4d). The analysis results based on Landsat images further reveal the dynamic changes in water area of Zonag Lake. Between 4 June and 14 July 2024, the lake area expanded from 152.98 km² to 157.69 km², an increase of 4.71 km² (Fig. 4a, b), before contracting to 152.49 km² by 2 October 2024.

Fig. 4: Precipitation in the Zonag and Yanhu Lake drainage basins from 1986 to 2024 and the change of the water area of the Zonag Lake in 2024.
figure 4

a Water area of Zonag Lake before and after the enlargement of the outlet in 2024; gray indicates an increase in lake area, while red indicates a decrease. b Difference in lake water area change before and after the enlargement of Zonag Lake outlet; the gray bars represent increases in lake area, while the red bars indicate decreases. c The dark gray bars are the average monthly precipitation in May for each year from 1986 to 2024, and the light gray bars are the average annual precipitation from 1986 to 2024. d Water level (blue dot) of Zonag Lake with red trendline, 2016 to 2024. e Water level (blue dot) of Zonag Lake from June to October 2024.

Full size image

Under the continuous hydrodynamic changes, the outlet of Zonag Lake was progressively eroded and enlarged. Field survey results indicate (Fig. 5b) that from 24 May to 27 June 2024, the width of the outlet has expanded markedly. Through quantitative analysis of multitemporal satellite imagery from 2015, 2021, and 2024, the shoreline of Zonag Lake retreated by 121.8 m in 2021 and by 156.1 m in 2021 compared to 2015 (Fig. 5a). The average widths of representative river cross-sections for the three years were 25 m, 15.18 m, and 24.25 m, respectively (Fig. 5c), highlighting the spatial and temporal dynamic changes in river morphology.

Fig. 5: Changes in the outlet region of Zonag Lake.
figure 5

a Gaofen-2 images (35.529°N–35.531°N, 92.032°E–92.035°E) of the Zonag Lake outlet region on 20 June 2015, 21 June 2021, and 6 September 2024. I–III represent the water surface extraction results, with red arrows indicating the measured distances from the current lakeshore to the original reference points. The three same black dashed boxes in 1–3 denote the selected river cross-sections used for calculating river width. b Photographs of the Zonag Lake outlet (red dashed boxes) taken on 24 May 2024 and 17 July 2024. c The three river cross-sections selected in Figures I–III, along with the average river width for each cross-section (Photograph in (b) taken by Zhaxi Nima, Public Security Bureau of Hoh Xil Nature Reserve, Qinghai Province, reproduced with permission. © 2024 Zhaxi Nima/Public Security Bureau of Hoh Xil Nature Reserve, Qinghai Province).

Full size image

The rise in water level and enlargement of the outlet of Zonag Lake, driven by extreme precipitation events, led to a rapid discharge of a large volume of lake water, directly affecting the hydrological regime of the downstream Yanhu Lake and highlighting a regional-scale linkage effect. The water area of the Yanhu lake has shown the first recorded expansion since the construction of the drainage diversion project in 2019, increasing from 207.43 km² to 209.35 km², a net gain of 1.92 km² (Supplementary Table 2). This process not only reflects the strong coupling between upstream and downstream hydrological processes in the Zonag and Yanhu Lake drainage basins, but also reveals the unstable response of the upstream Zonag Lake hydrological state to short-term heavy precipitation, and consequently increases the potential flood risk downstream of Yanhu Lake.

Sandstrom changes caused by wind erosion on the lake shores of the water recession area

Remote sensing analyses from 2011 to 2024 reveal a continuous shrinkage in Zonag Lake’s water area and a decline in water level, accompanied by annual retreat of the lake shoreline. The recession has exposed lakebed sediments, forming a sandstorm source area of approximately 103 km² on the western and southern shores10,23,24. The sediments in the exposed lakebed are mainly composed of silty clay, silty fine sand and organic material25,26, and are prone to wind erosion under strong winter winds, resulting in frequent sandstorm events. Such aeolian processes have caused significant ecological damage to the calving habitat of Tibetan antelopes, threatening the survival of population10,23.

This study used MODIS images to monitor sandstorm events in the Zonag Lake region from late 2011 to early 2024 (Fig. 6b). Time series analysis indicates that the annual number of sandstorm days was generally fewer than 30 from 2011 to 2016, and it increased significantly after 2017. From 2017 to 2024, the frequency of sandstorm events nearly exceeded 20% of the observation days each year except 2023 (Supplementary Table 4). In total, 434 sandstorm days were recorded during the observation period, accounting for approximately 22.7% of all observation days, indicating persistently high sandstorm activity in the region (Fig. 6a).

Fig. 6: Statistics of the sandstorm in the Zonag Lake region, 2011–2024.
figure 6

a Scatter plot showing the trend (red trendline) of the frequency (gray dot) of sandstorm in Zonag Lake between 2011 and 2024. b The red lines in the plot indicate the dates when the sandstorm was observed from November to March for each year between 2011 and 2024, where the X-axis is the maximum number of days in a month (31 days), and the Y-axis on the left represents the observed months in each year from 2011–2024; the Y-axis on the right represents the frequency of sandstorm in different months, and the blue color represents the frequency, with bluer colors indicating that more sandstorm have been observed in the corresponding months.

Full size image

Multitemporal Landsat imagery reveals a marked expansion of sand cover on the winter ice surface of Zonag Lake (Supplementary Table 5), concentrated mainly in the western and central sectors of the lake (Fig. 7). Coverage increased from 38.71 km² and 27.19 km² in 2014 and 2016, respectively, to 108.78 km² and 111.1 km² in 2023 and 2024 (Supplementary Table 5). Since 2017, sand deposition on the ice has consistently exceeded half of the total ice-covered area, indicating a sustained increase in aeolian transport from exposed shoreline sources. This process may reduce the ice surface albedo and enhance heat absorption, accelerating spring ice melt.

Fig. 7: Map of sand deposition on the frozen ice surface of Zonag Lake during the winter of 2011–2024.
figure 7

The red line in the satellite images at different times represents the lake boundary of Zonag Lake in that year; the frozen ice surface (blue) and the sand deposited on the ice surface (yellow) are shown in the corresponding vector map on the right.

Full size image

Future trends in precipitation and temperature and related potential disaster risks

The analysis of the factors influencing changes in the area of lakes in the basin revealed that average annual precipitation, average annual evapotranspiration, and glacial meltwater were the primary contributors to these changes (Supplementary Table 6). Given that the annual average temperature plays a key role in both evapotranspiration and glacial meltwater, the study used a 10-year observation window to conduct medium- to short-term climate predictions of future precipitation and temperature trends27. The findings indicate that both precipitation and temperature are expected to exhibit a continuous, moderate upward trend, with stable seasonal characteristics, between 2025 and 2035 (Fig. 8; Supplementary Figs. 1, 2). This trend suggests that the lake area changes in the basin will likely continue.

Fig. 8: Forecasting of average monthly precipitation and temperature for 2025–2035.
figure 8

a The optimal precipitation SARIMA (3, 1, 1) × (3, 1, 1, 12) is obtained from the minimum AIC = 2508.22; gray represents the forecast values, red denotes the trendline. b The optimal temperature SARIMA (3, 1, 1) × (3, 1, 1, 12) is obtained from the minimum AIC = 1898.97; gray represents the forecast values, red denotes the trendline.

Full size image

Under future scenarios of increased precipitation and temperature, the hydrological processes and environments of the Zonag and Yanhu Lake drainage basins will face new challenges and risks. In the winter of 2019, the thickness of sand accumulation on the ice surface of Zonag Lake ranged from 10 to 50 cm, with the estimated volume of water displaced by sand deposition being 0.006–0.03 billion m³ 24. This indicates that sand is not only an important indicator of wind erosion risk, but may also alter the local lakebed topography due to substantial accumulation in the western and central parts of the lake, thereby affecting hydrodynamic processes and outflow. In addition, the heavy precipitation events of 2024 triggered another significant linked hydrological response between the upstream and downstream systems. If such heavy precipitation occurs again, the lake outlet may expand and deepen once more, or even lead to a secondary shoreline failure of Zonag Lake, and the destruction of the downstream Yanhu Lake drainage engineering. Such floods could destroy or inundate the Qinghai-Tibet Railway and downstream highways14,28, disrupting the regional transportation network. The continuous inflow of Yanhu lake water may also compromise the thermal stability of the permafrost beneath the Qinghai-Tibet Railway13 and affect the hydraulic capacity of the original culvert at the Qingshui River Bridge on the Qinghai-Tibet Railway28. Furthermore, Yanhu Lake is one of the most saline lakes on the Tibetan Plateau28,29. This highly mineralized and saline water is discharged into the Qingshui River, increasing the TDS of the Qingshui River from 2689 mg/L to 5384 mg/L29. Continued saline water input could alter the water quality and ecosystem structure of the northern source of the Yangtze River.

Scientific cognition and management suggestions

This study advances the scientific understanding of hydrological reorganization under the dual influences of climate change and human intervention in high-altitude endorheic basins on the Tibetan Plateau. The evidence supports a “climate stress–system response” model, whereby natural events—such as the 2011 shoreline breach—coupled with engineering interventions like the 2019 Yanhu Lake drainage diversion project, create complex feedback loops within the regional hydrological network. Similar reorganization processes have been observed in other mountain regions8,30, highlighting the global relevance of our findings.

Our analysis shows that channelisation has, in this case, successfully mitigated immediate flood risks and restored vital hydrological connectivity between endorheic and exorheic systems. However, the intervention is not without its risks. The rapid changes observed at Zonag Lake’s outlet—marked by increased erosion, widening, and deepening—pose potential risks for secondary outbursts and further instability. Moreover, alterations in sediment flux and downstream water chemistry could have cascading ecological consequences, impacting aquatic habitats and biogeochemical cycles.

To address these challenges, scientifically valid engineering measures are recommended. Structural interventions to reinforce the outlet of Zonag Lake can enhance stability and reduce the risk of uncontrolled failure. Furthermore, because gravel cover has the potential to control wind erosion and maintain the stability of the soil-vegetation system, its use to protect the shores of Zonag Lake constitutes a shore-protection measure that combines physical effectiveness with ecological compatibility31. This strategy not only effectively suppresses wind erosion and reduces the risk of lakeshore desertification32 but also helps to protect the ecological integrity and environmental stability of the Lake’s shores. These measures should be integrated with long-term hydrological and ecological monitoring to ensure that adaptive management strategies remain responsive to emerging environmental dynamics.

Conclusions

This study has provided a detailed account of the hydrological reorganization in the Zonag and Yanhu Lake drainage basins and their subsequent connection with the Yangtze River. By integrating satellite monitoring with modeling techniques, we have documented how both natural forces—such as increased glacier melt and heavy precipitation—and human interventions, notably the 2019 drainage diversion project, have transformed the regional hydrological landscape and how climate change is accelerating this transformation. This transformation underscores a shift from isolated endorheic systems to an interconnected network, which has broad implications for downstream water security and ecological stability.

Our analysis confirms the “climate stress–system response” model on the Tibetan Plateau, where the interplay between climate variability and engineered solutions drives dynamic feedbacks within the hydrological network. The observed changes not only enhance our scientific understanding of high-altitude water systems but also provide essential insights for developing adaptive management strategies in regions vulnerable to climate change.

Ultimately, this research serves as a crucial reference for both the scientific community and water resource managers. It emphasizes the need for continued monitoring, interdisciplinary research, and proactive policy measures to safeguard the delicate balance of the “Asian Water Tower”. By highlighting the complex interactions between natural processes and human activity, our work lays the groundwork for future efforts aimed at mitigating risk and promoting sustainable water management in climate-sensitive regions worldwide.

Methods

Calculation of lake area and glacier area

Remote sensing data from Landsat 8 and Landsat 9, acquired through the United States Geological Survey (USGS) (https://earthexplorer.usgs.gov/), were used to extract the lake boundaries of Zonag Lake, Kusai Lake, Haidingno’er Lake, and Yanhu Lake from 2020 to 2024 (Supplementary Tables 8, 11), the glacier boundaries of the Zonag and Yanhu Lake drainage basins(Supplementary Table 9).

The threshold segmentation method, combined with the Modified Normalized Difference Water Index (MNDWI)33, based on the Normalized Difference Water Index (NDWI), was used to identify the water bodies and extract lake boundaries23. Due to variations in image quality across years, different thresholds were applied for each year according to the actual conditions, rather than using a single threshold, to more accurately extract the lake and glacier boundaries. The range of thresholds used is 0.2–0.3 (Supplementary Tables 8, 11). The formula for MNDWI is:

$${MNDWI}=\frac{({Green}-{SWIR})}{({Green}+{SWIR})}$$
(1)

where \({Green}\) refers to the green band and \({SWIR}\) refers to the short-wave infrared band.

The boundaries of the glacier were identified and extracted using the Normalized-Difference Snow Index (NDSI) and threshold segmentation methods34. Thresholds ranging from 0.4 to 0.45 were used for extracting glacier boundaries (Supplementary Table 9). The formula for NDSI is:

$${NDSI}=\frac{{Green}-{NIR}}{{Green}+{NIR}}$$
(2)

where \({Green}\) refers to the green band and \({NIR}\) refers to the near infrared band.

Calculation of lake water volume and glacial meltwater

Further calculations of lake water volume and glacier meltwater were conducted based on the lake and glacier areas. The lake water volume was calculated following the methods outlined by Ma et al.35 and Lu et al.10. This approach uses a 30 m DEM to calculate the area Si and volume Vi between different water surface elevation planes, and constructs the area-volume relationship equations Si = f(Hi) and Si = f(ΔVi). The surface area Si’ and underwater volume ΔVi’ corresponding to different underwater elevations Hi’ were then calculated using these two equations, leading to the final functional relationship A = f(V) for water area and water volume.

The calculation of glacier meltwater was based on methods provided by Hock36 and Lu et al.10. This method uses the degree-day factor (DDF) and the sum of positive accumulated temperatures at each time interval (PDD) to establish the equation ΔH = DDF × PDD × 10−3 for the amount of ice melt. Finally, the glacier meltwater G was calculated using the glacier area S and the amount of ice melt with the equation G = S × ΔH.

Calculation of lake shoreline changes and river width

This study employed Gaofen-2 satellite imagery (0.8 m spatial resolution) provided by the China Centre for Resources Satellite Data and Application (https://data.cresda.cn/#/home; Supplementary Table 10) to monitor changes in the water boundaries of the Zonag Lake failure area in 2015, 2022, and 2024. Image processing uses the Segment Anything Model (SAM). This model is universal and capable of image segmentation across remote sensing platforms and different spatial resolutions37,38. In this study, the images from three time scales were first aligned and standardized. Subsequently, point prompts were used as input to identify and extract the lake shoreline and river boundaries. In addition, water level data of Zonag Lake from the Database for Hydrological Time Series of Inland Waters (DAHITI) provided by the German Geodetic Research Institute and the Technical University of Munich in Germany (https://dahiti.dgfi.tum.de/en/41199/water-level-altimetry/) were used to analyze the water level changes of Zonag Lake from 2016 to 2024.

To quantitatively evaluate the shoreline retreat at the failure site of Zonag Lake in 2015, 2021, and 2024, a reference point (35.530° N, 92.0342° E) was selected at the site in the 2015 Gaofen-2 image. Measuring points were then set up in the 2021 and 2024 images, respectively, parallel to the direction of the river runoff from the reference point, with coordinates of 35.530° N, 92.033° E and 35.530° N, 92.032° E. By connecting reference points with measurement points from different years, the shoreline retreat distance of the lake shoreline was measured. To further evaluate river channel changes at the failure section, a representative cross-section was selected for analysis of river channel width changes. Based on the extracted river boundaries, the latitude and longitude coordinates of the left and right banks were obtained at equal intervals. The river centreline was constructed using Python’s Centerline-Width package, and measuring lines were drawn perpendicular to the centreline to measure multiple widths and calculate the mean, thereby quantifying changes in channel width.

Monitoring of sandstorm and calculation of sand area

The 250 m Moderate Resolution Imaging Spectroradiometer (MODIS) imagery provided by the Terra and Aqua satellites (https://worldview.earthdata.nasa.gov/) has a high temporal resolution with a revisit frequency of up to 1–2 days24,39,40. Based on this characteristic, these daily remote sensing images can help identify sandstorm features and accurately observe and track sandstorm occurrence23,41. This study focuses on the exposed lakebed formed by the dam failure of Zonag Lake, observing sandstorm events in the region from November to the following March each winter between 2011 and 2024. Through visual interpretation, the gray and yellow plumes appearing on the lake bed and surrounding areas within each scene were captured as sandstorms24,39,42,43. The number of sandstorm days was determined by counting the number of times each pixel was identified as a sandstorm.

The sand area on the lake surface of Zonag Lake during the winter from 2011 to 2023 was detected by subtracting the ice surface area (Supplementary Table 5) from the total lake area. Specifically, the Normalized-Difference Snow Index (NDSI) and threshold segmentation methods34 were applied to Landsat-8 and -9 satellite data (https://earthexplorer.usgs.gov/) to extract the ice surface area. The threshold extraction range for lake ice boundaries was −0.01 to −0.10 (Supplementary Table 12).

Analysis of factors associated with changes in lake area

The XGBoost-SHAP model was used for explanatory analysis of nonlinear relationships44. XGBoost is a machine learning algorithm that implements gradient boosting decision trees, which are scalable across various scenarios45. SHapley Additive exPlanations (SHAP) is an interpretation framework based on cooperative game theory46, which allows for the interpretation of the XGBoost predictive model by quantifying the contribution of each hydrological factor to the water area and assigning different weights to each influencing feature (Supplementary Table 6). The Shapley equation is:

$$g\left({z}^{\prime} \right)={{{\varnothing }}}_{0}+{\sum }_{m=1}^{M}{{{\varnothing }}}_{m}{z}_{m}^{{\prime} }={bias}+\sum {C}_{i}$$
(3)

where \(g\left({z}^{{\prime} }\right)\) is the model output (XGBoost), and \({\varnothing }_{m}{z}_{m}^{{\prime} }\) represents the contribution of each feature to the model output.

SARIMA forecasting model for precipitation and temperature

The Seasonal Autoregressive Integrated Moving Average (SARIMA) model is a stochastic model that excels at capturing cyclical features, trends, and serial correlations in time series data47,48,49. SARIMA extends the Autoregressive Integrated Moving Average (ARIMA) model by adding seasonal parameters, making it suitable for forecasting seasonal data. The SARIMA model construction for precipitation and temperature follows five key steps.

Step 1: Time series decomposition

Time series decomposition helps identify patterns in the data and determine trends and seasonality50. Seasonality refers to recurring patterns at fixed intervals influenced by seasonal factors51. If seasonal behavior is detected, the observed seasonal period is added to the ARIMA model. Classical multiplicative and additive decompositions were used for the time series decomposition, where a multiplicative model was applied to precipitation data52,53 (Supplementary Fig. 1a–d):

$${X}_{t}={T}_{t}\times {S}_{t}\times {C}_{t}\times {e}_{t}$$
(4)

Temperature data were decomposed using an additive model (Supplementary Fig. 2a–d):

$${X}_{t}={T}_{t}+{S}_{t}+{C}_{t}+{e}_{t}$$
(5)

where t is the observed time series, \({T}_{t}\) is the trend, \({S}_{t}\) is the seasonal effect, \({C}_{t}\) is the cyclical component, and \({e}_{t}\) is the irregular component.

Step 2: Description and identification of time series correlation, lag, and trend

The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to describe the time series correlation, lag, and trend. The seasonal lags and the effect of lagged values demonstrated by the ACF (Supplementary Figs. 1e, 2e) and PACF (Supplementary Figs. 1f, 2f) help determine the number of moving average (MA) and autoregressive (AR) terms in the model, which correspond to q and p in the SARIMA model51,54,55.

The equation ACF is:

$${c}_{k}={\hat{\gamma }}_{k}=\frac{1}{N}{\sum }_{t=1}^{N-k}\left({z}_{t}-\bar{z}\right)\left({z}_{t+k}-\bar{z}\right)\,k=0,1,2,\ldots ,K$$
(6)

where \({\gamma }_{k}\) is the estimate of the autocovariance, \({z}_{t}\) is the time series value at time t, \(\bar{z}\) is the sample mean of the time series, N is the number of sample data, and k is the lag.

The PACF equation is:

$${\phi }_{{kk}}={corr}[{z}_{t}-{\hat{z}}_{t},\,{z}_{t-k}-{\hat{z}}_{t-k}]$$
(7)

where \({\phi }_{{kk}}\) is the last coefficient, and the other values are defined as in the ACF equation.

Step 3: Smoothness analysis of time series

The Augmented Dickey–Fuller test (ADF) is used to check the smoothness of the time series. If p < 0.5, the time series is stationary50. If the series is non-stationary, differencing the data is required; otherwise, SARIMA modeling can proceed directly55,56. The equation for ADF test is:

$${\Delta }^{2}{Y}_{t}=\alpha +\gamma {y}_{t-1}+\beta t+{\sum }_{i=1}^{p}{\beta }_{1}\Delta {y}_{t-1}+{\varepsilon }_{t}$$
(8)

Step 4: SARIMA model and optimal parameter combination

Average monthly precipitation and temperature data from 1961 to 2024 were used for this study; the data were provided by Wudaoliang National Meteorological Station, Qinghai Province, China. The data from 1961 to 2014 served as the training set, and the data from 2015 to 2024 served as the test set. In the SARIMA model, \(\left(p,{d},{q}\right)\) represents the non-seasonal part, and \({\left(P,{D},{Q}\right)}_{m}\) represents the seasonal part of the model51,55,57. The SARIMA equation is:

$${\phi }_{p}(B){\Phi }_{P}({B}^{s}){(1-B)}^{d}{(1-{B}^{s})}^{D}{Y}_{t}={\theta }_{q}(B){\Theta }_{Q}({B}^{s}){\varepsilon }_{t}$$
(9)

where p is the autoregressive component of order p, q is the moving average component of order q, d is the integrated component of order d, P is the number of seasonal AR terms, D is the number of seasonal differences, Q is the number of seasonal MA terms, and s is the seasonal period.

The accuracy of SARIMA predictions is assessed using the Mean Absolute Deviation (MAD), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2)58,59,60 (Supplementary Table 7).

The optimal parameter combination is found using Akaike’s Information Criterion (AIC). The minimum AIC value is considered the best model (Hyndman, 2018), with the equation:

$${AIC}=T\log \left(\frac{{SSE}}{T}\right)+2(k+2)$$
(10)

where T is the number of observations, and k is the number of predictors.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.