Recent drying of the Ganga River is unprecedented in the last 1,300 years

Significance

The Ganga River has profound religious, cultural, and economic significance for millions living in India, Nepal, and Bangladesh. The Ganga River has been facing drying trends; however, it remains unclear whether the recent drying is unprecedented. We examined the recent drying of the Ganga River using reconstructed streamflow for the past 1,300 y and found that the river basin has faced its worst droughts in the last few decades. The recent drying is well beyond the realm of last millennium climate variability, and most global climate models fail to capture it. Our findings highlight the need to represent the complex interactions among natural and anthropogenic factors in climate models, which can assist water management strategies under climate warming.

Abstract

The Ganga River basin, critical to over 600 million people, is experiencing a severe and unprecedented drying trend, threatening water and food security. Using streamflow reconstructions spanning 1,300 y (700–2012 C.E.) from instrumental data, paleohydrological records, and hydrological modeling, we show that drying from 1991 to 2020 is unmatched in the past millennium. Streamflow decline since the 1990s, driven by frequent and prolonged droughts, is 76% more intense than the 16th-century drought—the closest historical analogue. This drying exceeds natural variability, highlighting the dominant role of anthropogenic factors. Despite CMIP6 models projecting increased streamflow under warming scenarios, the recent decline indicates complexities associated with future water availability projections. Our findings underscore the urgent need to examine the interactions among the factors that control summer monsoon precipitation, including large−scale climate variability and anthropogenic forcings. Better constraints on these processes in climate models will be essential for improving future monsoon projections and implementing adaptive water management strategies to secure the Ganga basin’s freshwater availability under a changing climate.

Streamflow observations show widespread drying in several regions across the globe (1–5). Droughts impact societies and national economies through agricultural losses, reduced hydropower, and threats to water and food security (6–10). A large population of India relies on agriculture for their livelihoods and, therefore, is vulnerable to droughts (11). In recent decades, weakening of the summer monsoon (12, 13) has caused frequent droughts and drying of the Ganga River (14). The Ganga River basin (GRB) is a lifeline for more than 600 million people in the Indian subcontinent, accounting for approximately 40% of India’s gross domestic product (15). The GRB supports extensive agro-industrial activity; therefore, recent drying experienced by the region has had pervasive and downstream impacts on several sectors. For instance, during 2015–2017, historically low water levels across the middle and lower reaches of the Ganga River severely disrupted drinking water supply, power generation, irrigation, and navigation, affecting over 120 million people (16). The observed drying has further exacerbated the already stressed groundwater storage of the region (17, 18). Furthermore, the decrease in river flow has reduced the freshwater contribution to the Bay of Bengal, potentially threatening one of the world’s most productive marine ecosystems (3).

Observed changes in flow in the GRB are associated with climatic changes and human interventions. For instance, weakening of the summer monsoon, which is linked to Indian Ocean warming and anthropogenic aerosol emissions (12, 19, 20), has partly led to the drying of the Ganga River (21, 22). Increased warming combined with the summer monsoon drying has resulted in massive groundwater pumping that has led to rapid groundwater depletion in the basin (18, 23–25). Reduced baseflow due to groundwater pumping might have also contributed to the reduction in flow (17). Therefore, the main drivers of the recent drying of the Ganga River remain poorly understood. The severity of the recent decline in streamflow in the GRB remains unknown, primarily due to short precipitation and streamflow observations. Understanding the causes and severity of drying of the Ganga River is vital for developing effective adaptation strategies.

Paleoclimate reconstructions of streamflow are vital for understanding the full nature of hydroclimatic variability over a basin (26–33). Despite several streamflow reconstructions for major rivers in the Indian subcontinent (26, 34–38), long-term paleoclimatic reconstructions of river flow, including the GRB, have been lacking. Traditional paleohydrology reconstructions in the Indian subcontinent have primarily relied on tree-ring chronologies that exhibit spatial irregularity and temporal inconsistency (38–41). The use of systematically derived datasets (42–44) from paleoclimate proxies provides a more robust alternative for streamflow reconstruction (28, 29, 37, 45–47). The central objective of our study is to understand the causes, severity, and future risk of streamflow drying in the GRB based on long-term and recent drought datasets. Here, we reconstruct the annual streamflow for the GRB for the last 1,300 y using the Monsoon Asia Drought Atlas (44) to examine the causes and severity of the recent drying in the Ganga River (SI Appendix, Fig. S1). We utilize this dataset to investigate the large-scale ocean-atmospheric drivers of drought in this region, thereby contextualizing the recent drying trend. We then used hydrological modeling and climate projections to assess how well current models capture the observed drying trend and evaluate potential future changes in water availability and streamflow in the GRB.

Reconstruction of Ganga River Streamflow

We reconstructed streamflow in the GRB at Farakka using a log–linear model between log-transformed streamflow and a parsimonious subset of Palmer Drought Severity Index (PDSI) predictors (see Materials and Methods for more details). The observed streamflow shows a strong relationship (r = 0.89) with observed precipitation in the GRB (SI Appendix, Fig. S2B). The long-term reconstruction shows that the mean streamflow for the Ganga River between 700 and 2012 C.E. is significantly higher (P-value < 0.05) than for 1951–2020 (Fig. 1C). The observed decline in mean streamflow for the recent period may potentially be linked to the frequent occurrence of recent meteorological droughts (48), which we further explore in this work.

Fig. 1.

Streamflow reconstruction and cross-validation at Farakka. (A) Comparison of VIC-simulated streamflow (black) and reconstructed streamflow (red) at Farakka for the observed period 1951−2012. The gray-shaded region represents the 95% CI of the reconstructed streamflow. The solid blue line represents the observed streamflow at Faraka for the available period. (B) Box plot of the reduction of error (RE) during validation, Nash-Sutcliffe efficiency (NSE) during validation, Kling-Gupta efficiency (KGE) during validation, validation period square of Pearson correlation (VRSQ) for 40 ensemble reconstructions. (C) Reconstructed streamflow (blue) for each year between 700 and 2012 C.E., along with all 40 ensemble reconstructions (gray). The black and blue dashed lines represent the mean of the reconstruction (700−2012) and the VIC-simulated mean (1951−2020), respectively. Red stars denote the significant change points in the streamflow time series. (D) Standardized streamflow anomaly (SSA) between 700 and 2020 using the full reconstruction (700−2012) extended with the VIC simulations (2013−2020), with a 30-y backward moving mean (black). SSA is estimated as the Z-score of the streamflow series, which follows a normal distribution. The blue horizontal line indicates the 30-y moving mean SSA for 2020. Orange circles denote documented drought years from 700 to 2020 in the Indo-Gangetic Plain.

Over the past 1,300 y, extreme wet years (SSA or Z-score > 1) in the GRB have generally exceeded extreme dry years (SSA < −1) (SI Appendix, Fig. S4C). However, a considerable departure from this long-term trend has occurred during the last three decades (1991–2020), which experienced a significant increase in extreme dry years (SI Appendix, Fig. S4C). Moreover, the GRB did not witness a single extreme wet year during 1991–2010, whereas only two extreme wet years occurred during 2011–2020. We examined changepoints in the reconstructed streamflow using Bayesian changepoint detection, considering the reconstructed flow of the last 1,300 y (49). The most notable changepoint occurred in 1991, characterized by a sudden decline of ~620 m³/s in mean annual streamflow, with an occurrence probability (P) of 0.74 (Fig. 1C). Collectively, these findings emphasize the major hydrological shifts that have occurred in the GRB, particularly the emerging dry conditions over recent decades. Overall, the GRB has witnessed an unusual drying that is likely caused by the weakening of summer monsoon precipitation in the region (12, 20).

The reconstructed streamflow data for the GRB reflect significant droughts that occurred during the observational period with high fidelity (50) (Fig. 1D). It also identifies major famines (51, 52) caused by repeated monsoon failures, such as those between 1344 and 1355, which had a cumulative SSA of −1.85, and the Bengal famine from 1769 to 1771 (cumulative SSA of −2.91). These findings highlight the reliability of the reconstructed streamflow record. The famines in India resulted in lasting consequences, including high human mortality due to crop failures attributable to these droughts (53). Other significant droughts that had enduring effects include the famine of 1802–1804 (SSA of −2.87) and the Indian famine from 1899 to 1900 (54), notably impacting the central and western regions (SSA of −1.2). In summary, the historically documented droughts and famines are well represented in the reconstructed streamflow data.

Severity of the Recent Drying of the Ganga River

Next, we examined the rarity of the recent drying of flow in the Ganga River using the 30-y backward-moving means of SSA to examine episodes of river drying similar to the 1991−2020 period (Fig. 1D). We used a 30-y backward-moving mean to identify analogues of the recent drying and found that it is unmatched across the reconstruction over 700−1990 C.E., highlighting the severity and rareness of the event (the 30-y moving mean SSA for the period 1991–2020 was −1.23σ; blue line, Fig. 1D). The second most severe drying of the Ganga River based on the 30-y flow anomaly occurred during 1501–1530, aligning with findings of increased aridity in northern India during the early 16th century (55). The third driest 30-y flow in the river occurred during the mid-14th century (1344–1373). The drying of the Ganga River during the mid-14th century is consistent with the δ¹⁸O records of Dandak cave from the core monsoon region of India (56). Overall, the river drying during 1991–2020 is unprecedented and more severe than its nearest two historical analogues during the entire record of 1,300 y.

After identifying the enormity of the recent drying in the GRB, we applied a “runs theory” approach (27, 57) considering consecutive years with river flow less than the mean flow of the 700−2020 period to analyze the severity of recent droughts in the GRB (Fig. 2A). The early 16th (1520–1530 C.E., 11 y) and the late 15th (1497–1511 C.E., 15 y) centuries witnessed the two longest droughts spanning over a decade (Fig. 2D). Short-term droughts of 1-y, 2-y, and 3-y durations typically occur once in every 6, 10, and 28 y, respectively (Fig. 2A). Long-term droughts lasting 4 y or more are unusual in the GRB, with a median recurrence interval of 70 to 200 y over the last millennium (Fig. 2 A and B). In striking contrast, the GRB faced four droughts of duration of more than 3 y between 1991 and 2020 (Fig. 2A), signifying the anomalous nature of drying in recent decades.

Fig. 2.

Severity–duration analysis for reconstructed streamflow (700−2020 C.E.) using SSA based on runs theory. (A) Cumulative severity (filled light-brown circles) of individual drought events (defined as consecutive years below the long-term 700–2020 mean) of the duration shown on the x-axis, corresponding to the Left-hand y-axis. Stars indicate drought events that occurred between 1991 and 2020. The box and whisker plots correspond to the Right-hand y-axis represent the interval between droughts of duration shown on the x-axis. (B) Distribution of drought events of different durations, and (C) the percentage by which the selected drought events between 1991 and 2020 have exceeded the severity of corresponding historical drought events. (D) Trajectories of the top 10 hydrological drought events and their cumulative severity between 700 and 2020 C.E. The trajectories highlighted with stars indicate two recent drought events between 1991 and 2020.

The GRB experienced several notable droughts spanning different durations during the recent period. Notably, the 1999–2002 drought ranks as the second most severe 4-y drought of the past 1,300 y, surpassed only by the 1720–1723 event. The latter included the 1721–1723 monsoon-season drought, which affected 76% of India and remains the most severe monsoon failure on record (58). The 2014–2018 drought stands out as the most intense 5-y drought in the GRB over the last millennium. Even more striking are two unprecedented 7-y drought events that occurred between 1991 and 2020: 1991–1997 and 2004–2010. Among these, the 2004–2010 drought ranks as the most severe drought of the past 1,300 y, whereas 1991–1997 was the second-worst 7-y drought period. Both rank among the ten longest droughts in the Ganga basin’s 1,300-y history (Fig. 2D). Our findings reveal that severe and prolonged droughts have become more frequent in recent decades.

Large-Scale Drivers of Drying of the Ganga River

Observational records of precipitation and streamflow support the recent drying of the GRB (SI Appendix, Supplementary Text 1). The GRB has experienced a significant decrease of 9.5% (±9.6%) (P-value <= 0.05) in total annual precipitation during 1951−2020 (Inset panel in Fig. 3 A), with a more prominent decline of over 30% in the western region (Fig. 3A). Conversely, precipitation has increased by 10% in parts of the GRB located in Nepal and east-central India. Recent data from the India Meteorological Department (IMD) (SI Appendix, Supplementary Text 2) show a decline in mean annual precipitation in the GRB, aligning with trends observed in other precipitation products (SI Appendix, Fig. S2A). Whereas anthropogenic warming has raised temperatures across India, the increase in the Gangetic plains is less pronounced compared to other regions (Fig. 3B). Notably, several areas in the basin have cooled between 1951 and 2020, attributed to increased irrigation practices (59–62). Enhanced evaporation due to higher moisture availability from irrigation has contributed to this surface cooling (61). The Himalayan regions of the basin have warmed by over 1 °C since 1951, contributing to an overall warming trend of 0.62 °C (P-value < 0.05) in the basin (Fig. 3B). The interannual variability of precipitation and temperature considerably affect flow in the GRB (Fig. 3C). Precipitation anomalies are strongly associated (R² = 0.94) with river flow anomalies in the GRB. Moreover, dry years are associated with accelerated warming (red dashed line) in the basin. The decline in monsoon precipitation has also reduced groundwater storage in the GRB, affecting both recharge and abstraction (18, 23, 24) and contributing to reductions in total streamflow (63). In summary, we identified 15 hydrological drought years (SSA < −0.5) and 20 y with below-average streamflow during the 30-y period of 1991–2020, which were mainly driven by precipitation deficits (Fig. 3C).

Fig. 3.

Observed changes in precipitation and temperature between 1951 and 2020. Spatial distribution of change in (A) annual precipitation (%) and (B) annual mean temperature (°C) between 1951 and 2020 based on the Sen’s slope. Grids with statistically significant trends (P <= 0.05), based on the Mann–Kendall test, are highlighted with stippling. The Inset panels in (A) and (B) represent the interannual variability in precipitation anomaly (%) and temperature (°C) averaged for the GRB (blue boundary). The total change in average precipitation and temperature over the GRB during 1951−2020, estimated using the Sen’s slope, is statistically significant (P-value <= 0.05) based on the Mann–Kendall test. (C) Standardized anomaly of precipitation (P) (dashed brown), temperature (T) (dashed red), and streamflow (S) (bars) for each year between 1991 and 2020. Standardized anomaly was estimated using 1951−2020 mean streamflow and SD, and years with SSA less than −0.5 were considered hydrological drought years based on VIC-simulated streamflow.

Previous major monsoon droughts that led to GRB drying were associated with warm sea surface temperatures (SSTs) in the Indo-Pacific Oceans (resembling an El-Niño-like SST pattern) (64), which induced high sea-level pressure and weakened moisture transport into the region (SI Appendix, Fig. S5 A−D). Conversely, extreme wet years exhibited cooler SSTs (resembling a La-Niña-like SST pattern) in the Indo-Pacific (65), facilitating moisture-laden south easterlies that lead to above-normal precipitation across the Indian subcontinent (SI Appendix, Fig. S5 E−H). We examined oceanic and atmospheric conditions during the 15 hydrological drought years (SSA < −0.5) (66), categorizing them into El Niño, La Niña, and neutral years (SI Appendix, Table S1 and Fig. S6). As expected, El Niño events in 1997, 2002, 2009, and 2015 caused widespread rainfall deficits across India (SI Appendix, Figs. S6 A and B and S7 A and B). Neutral years showed weak El Niño patterns, coupled with a negative Indian Ocean Dipole (IOD), which further weakened moisture transport from the Indian Ocean, leading to suppressed rainfall (SI Appendix, Table S1 and Figs. S6 E and F and S7 E and F). Interestingly, during the strong La Niña years of 2007 and 2010, whereas most of India experienced above-average rainfall, the GRB received below-average precipitation, contrasting with expectations of monsoon variability arising from internal climate dynamics (SI Appendix, Fig. S6 C and D). Elevated sea level pressure (SLP) and corresponding surface air temperature (SAT) patterns over northern India (north of 15°N) induced subsidence and an anticyclonic pattern, resulting in northeasterly winds that weakened monsoon circulation and moisture transport into central India and the GRB (SI Appendix, Fig. S7 C and D), causing reduced streamflow. We identified hydrological wet events, and in contrast to the rainfall deficit years, none of the six wet years (SSA > 0.5; 1998, 1999, 2003, 2011, 2013, and 2020) occurred during the El Niño phases. Wet years were primarily associated with La Niña events (1998, 1999, 2011, and 2020). Neutral years such as 2003 and 2013 also exhibited weak La Niña-like conditions, which enhanced rainfall across the subcontinent (SI Appendix, Fig. S8).

Ganga River Flow Under a Warming Climate

The GRB is projected to experience substantial hydrological shifts with future warming (Fig. 4). Streamflow sensitivity analysis (using the VIC model) in the GRB under various warming and precipitation scenarios indicates that summer monsoon deficits contribute significantly to declining streamflow, and warming plays a relatively lesser role (SI Appendix, Fig. S9). Nevertheless, severe reductions in streamflow (5 to 35%) are projected if warming coincides with ongoing precipitation declines under extreme climate scenarios (SI Appendix, Fig. S9). The multimodel mean from 25 CMIP6 models projects that basin temperatures could increase by over 4.5 °C (±1.19 °C) by the end of the 21st century, with a 30% rise in precipitation (±21%) under the SSP5-8.5 scenario (Fig. 4 D and E). While these models accurately capture the observed warming trend (Fig. 4E), a majority of climate models do not capture the observed drying over the GRB (Fig. 4 A, B, and E). For instance, only five out of twenty-five models capture the ongoing drying trend in precipitation and streamflow (Fig. 4 C, D, and F and SI Appendix, Table S2). Projections from these selected models suggest a substantial increase in streamflow relative to the 1951−2020 mean, which is expected to continue and intensify toward the end of the century (26.16%, ±17%). Increasing precipitation under warming conditions may reverse the ongoing drying trend by 2040 (Fig. 4F). Streamflow reconstructions and projections suggest that the recurrence interval of low-flow years (defined as years with flows below the 1951−2020 mean, SI Appendix, Supplementary Text 3) could extend from the current 2 y interval to approximately 6 y in the future (Fig. 4G). Under regional warming of 1 to 4 °C relative to 1951–2020, precipitation is expected to increase by 3 to 25%, potentially driving a 4 to 30% rise in streamflow in the GRB (SI Appendix, Fig. S10).

Fig. 4.

Changes in water availability under a warming climate. Spatial distribution of 30-y mean precipitation anomalies over the GRB based on (A) observations from IMD, (B) the CMIP6 multimodel mean, and (C) selected CMIP6 GCMs for the period 1991–2020. (D) 30-y moving mean precipitation anomalies from IMD (red) for the period 1980–2020, with solid and dashed blue lines representing the 30-y moving-mean CMIP6 multimodel mean and the uncertainty based on one SD among the GCMs between 1980 and 2100, respectively. The green line shows the 30-y moving mean of selected CMIP6 GCMs capturing the drying trend. (E) Same as (D), but for mean temperature. Gray lines in (D and E) indicate trends from twenty-five individual CMIP6 models. (F) 30-y mean streamflow anomalies from IMD (red) for the period 1980–2020 and from the selected CMIP6 GCM multimodel mean (dark green) for 1980–2100. Anomalies in (A–F) were estimated with respect to the reference period of 1951–2020. (G) Recurrence intervals (in years) of dry periods with streamflow below the mean streamflow of the 1951–2020 period, calculated from 1,000 draws of 30 y with replacement from VIC-simulated streamflow based on observed IMD forcing (1951–2012), the reconstruction over the observed period (1951–2012 C.E.), the multimodel mean of selected CMIP6 streamflow simulations for 1951–2012, the full reconstruction period (700–2012), and the multimodel mean of selected CMIP6 streamflow simulations for 2021–2100.

Despite these projections of wetter conditions, considerable uncertainties in the CMIP6 models limit our understanding of future water availability in the GRB. The models struggle to capture the observed drying trend from recent decades, likely due to limitations in spatial resolution and the representation of regional climate processes (e.g., anthropogenic aerosol emissions) and human intervention (67–70). Only a subset of models captures the ongoing decline in precipitation, and even these models tend to underestimate the rate of decline (Fig. 4D). This misalignment between observed and simulated trends raises concerns about the uncertainty in future precipitation projections. To examine whether the CMIP6 models’ failure to capture the recent drying could be due to internal climate variability or not, we analyzed streamflow simulations from 100 ensemble members of the Community Earth System Model version 2 Large Ensemble (CESM2-LENS2) (SI Appendix, Supplementary Texts 2 and 4). None of the ensemble members reproduces the observed drying trend over the GRB (Fig. 1D and SI Appendix, Fig. S11A). We also examined preindustrial (850–1849) control runs from four Paleoclimate Modeling Intercomparison Project Phase 4 (PMIP4) CMIP6 models (SI Appendix, Supplementary Texts 2 and 4) to assess whether these models can simulate drying events of comparable magnitude over the past millennium. Across these simulations, no 30-y period exhibited streamflow anomalies as severe as the recent drying (Fig. 1D and SI Appendix, Fig. S11B), suggesting that the observed trend likely exceeds the range of internal variability captured by current climate models.

Consequently, despite a projected increase in water availability, the persistence of uncertain model outputs and the observed sensitivity of streamflow to warming suggest that river drying could become more intense under warming, especially if precipitation declines continue. Given these uncertainties, caution is warranted in interpreting projections, emphasizing the need for adaptive water management strategies to mitigate possible water scarcity challenges in the GRB under a warming climate.

Discussion

Human activities have significantly altered the hydrology of the GRB (14, 15, 71–73). Paleoclimate reconstructions, which provide insights into preinstrumental climate variability, are essential for understanding these changes and establishing baselines for future projections. We reconstructed 1,300 y (700–2012 C.E.) of annual streamflow at the Farakka station downstream in the GRB (West Bengal) using a log–linear model between streamflow and MADA-PDSI, to contextualize the 1991–2020 recent drying in historical drought patterns. The streamflow in the GRB during 1951–2020 was significantly drier than the mean reconstructed streamflow for 700–2012. The 30-y moving flow anomaly highlights pronounced nonstationarity, with recent decades (1991–2020) exhibiting anomalous and unprecedented dry conditions. This dry period was 76% and 151% more severe than its two nearest historical analogues (1501–1530 C.E. and 1344–1373 C.E.). Five droughts lasting 1 to 7 y were recorded during this period, which were unprecedented in the last millennium, contributing to the streamflow decline.

We find that the recent drying of the GRB was primarily driven by reduced southwest monsoon precipitation (12, 14, 73, 74). Long-term reconstruction of basin-averaged precipitation over the GRB also supports the declining precipitation over the basin in the recent period (SI Appendix, Fig. S12). Monsoon strength is influenced by Indo-Pacific SST anomalies, where warmer SSTs strengthen the monsoon, and cooler SSTs weaken it (65, 75, 76). However, two of the 15 droughts from 1991 to 2020 coincided with cooler Indo-Pacific SSTs, indicating the role of anthropogenic factors in driving GRB precipitation changes beyond natural interannual variability. Whereas anthropogenic warming is expected to intensify tropical rainfall (77, 78), rapid warming in the Indian Ocean and subdued warming across the subcontinent have weakened the land-sea thermal gradient, resulting in reduced monsoon rainfall over north India (12). Regional monsoon variations are also linked to anthropogenic aerosols that suppress local precipitation (19, 79). Dust and aerosols over central India and the Indo-Gangetic Plain additionally contribute to declining rainfall (80). Finally, reduced groundwater recharge due to declining precipitation, coupled with excessive groundwater extraction for irrigation, also exacerbates streamflow reduction (24, 25, 63).

Our 1,300-y streamflow reconstruction provides a critical baseline for assessing recent hydrological shifts in the basin. Whereas the reconstruction model demonstrates strong cross-validation skills, potential limitations remain. The MADA-PDSI records rely on heterogeneous tree-ring chronologies with uneven temporal coverage, which may affect accuracy, especially in earlier periods with fewer available data (44, 81). Notably, the number of tree-ring records drops sharply before the 14th century, likely reducing reconstruction fidelity in earlier centuries (44). This limitation underscores the need for caution when interpreting long-term variability and extremes in the early part of the record. Incorporating more tree-ring chronologies and other proxies (82) could improve reconstruction accuracy. In addition, refining climate models to better account for regional dynamics and human interventions (aerosols and irrigation) is essential for reliable future projections that can guide climate adaptation (67, 69, 83, 84). Advancements in both paleoreconstructions and climate models are vital for informing climate policies and developing effective water management strategies in response to ongoing anthropogenic climate change and the increasing drying trend in the GRB.

Materials and Methods

Variable Infiltration Capacity (VIC) Model.

The long-term reconstruction of streamflow using paleoclimate proxies requires a temporally consistent record of observed streamflow spanning several decades (more than 40 y) for skillful reconstruction. However, as Ganga is a transboundary river, the streamflow observation downstream in the GRB is only available between 1951 and 1973 (missing data between 1961 and 1964) at Farakka station (85, 86). A downstream streamflow station captures basin-wide climate characteristics by integrating cumulative upstream hydrological responses. We supplement the streamflow records at Farakka using streamflow simulations from a well-calibrated and evaluated hydrological model. We used the VIC land surface model (87) combined with the standalone routing model (88) to simulate streamflow at Farakka for the period 1951−2020. The VIC model, set up at a 0.25° spatial resolution, was driven by meteorological inputs, including P, Tmax, Tmin, and wind speed (SI Appendix, Supplementary Text 2). The model demonstrated robust performance in simulating the monthly streamflow (NSE > 0.95) and other water balance fluxes, including soil moisture and evapotranspiration (ET). The detailed calibration and validation statistics of the VIC model are available in SI Appendix, Supplementary Text 4. We also used the VIC model to simulate streamflow at Farakka based on CMIP6, PMIP4-CMIP6, and CESM2-LENS2 precipitation and temperature data of historical and future periods (SI Appendix, Supplementary Texts 2 and 4).

Paleoclimate Proxy Data.

We used the Monsoon Asia Drought Atlas version 2 (MADA v2) (43) as a paleoclimate proxy to reconstruct streamflow at Farakka. MADA provides an annual time series (till 2012) of gridded self-calibrated PDSI (89) for the mean June–July–August (JJA) period (44). The MADA dataset was developed using tree ring observations and is available at a 1° resolution for the Asian monsoon region, covering the past millennium or longer and up to 2012. The number of tree-ring observations used in MADA reconstruction decreases back in time, which can potentially reduce the skill of MADA-PDSI in the earlier centuries (SI Appendix, Fig. S15A). We used gridded MADA-PDSI instead of tree ring observations as it provides spatially consistent grid cells (44) and has been systematically calibrated with instrumental PDSI (90). MADA-PDSI is available from 700 C.E. onward for northern India, which includes the GRB region (SI Appendix, Fig. S15B).

We evaluated the performance of MADA-PDSI by comparing it with observed PDSI from the IMD-ERA5 datasets and found a strong agreement in the two datasets (SI Appendix, Fig. S15 C and D. We also compared MADA-PDSI with IMD-ERA5 precipitation and temperature and found good correlations (r = 0.68 and r = −0.52, respectively), confirming that tree-ring-based MADA-PDSI captures the precipitation and temperature variability (SI Appendix, Fig. S16). Additionally, the comparison of MADA-PDSI predictors and tree-ring chronologies with observed streamflow suggests that PDSI predictors exhibit a higher correlation with streamflow than the tree-ring chronologies (45). We compared MADA-PDSI with the PDSI from the Great Eurasian Drought Atlas (GEDA-PDSI) (91) to further assess its consistency with other tree-ring-based PDSI datasets. GEDA provides the PDSI dataset covering the entire Eurasian region and parts of sub-Saharan Africa for the last millennium (1000–2020 C.E.) at 0.5° resolution. Both MADA and GEDA show a consistent long-term decline in PDSI over the GRB. However, MADA tends to slightly overestimate, and GEDA slightly underestimates, the observed drying trend (SI Appendix, Fig. S17). The relationship between streamflow and PDSI is underpinned by the fact that both respond similarly to climatic anomalies (28, 37). Consequently, the PDSI reconstructed from tree rings, which are excellent proxies for past climate conditions, can be effectively used to reconstruct streamflow.

Paleo Proxy Predictor Selection.

Previous studies on streamflow reconstruction in Europe and the conterminous United States using gridded PDSI have typically employed PDSI predictors within a search radius of 450 km (28, 29, 46, 92). The gridded PDSI datasets used in these studies, such as the Living Blended Drought Atlas (LBDA) (42), rely on tree-ring chronologies within the 450 km radius from each grid point. Wu et al. (45) used the MADA-PDSI dataset to reconstruct streamflow for five rivers in the Tibetan Plateau with a search radius of 700 km. Conversely, Nguyen et al. (37) considered MADA-PDSI predictors based on the hydroclimatic similarity between streamflow stations in monsoon Asia regions and MADA grid points within a fixed search radius of 2500 km. Nguyen et al. (37) defined hydroclimatic similarity as the Euclidean distance (KWF) in hydroclimate space between the station and grid point, considering six KWF values (0.1, 0.15, 0.20, 0.25, 0.30) for ensemble reconstruction. We used an ensemble reconstruction approach and considered MADA-PDSI predictors for five different search radii, i.e., 1,000 km, 1,500 km, 2,000 km, 2,500 km, and 3,000 km. The MADA-PDSI dataset itself is developed using an ensemble reconstruction approach based on different search radii (44). Here, we assume that the time series of reconstructed streamflow based on the median of ensemble reconstructions will be superior to any individual ensemble member (44).

For each search radius, we retained the potential PDSI predictor if the Pearson correlation between the annual flow (1951−2012) at Farakka and the PDSI grid (1951−2012) was statistically (at 5% level) significant (SI Appendix, Fig. S18 A−E). Most of the PDSI predictors are clustered within the basin boundary of the Ganga River for all search radii, thus preserving the basin climate characteristic (SI Appendix, Fig. S18 A−E). We selected the streamflow reconstruction period from 700 to 2012 C.E. (thirteen centuries) due to the consistent PDSI data available for all potential predictors during this period. PDSI exhibits a good correlation with the annual, JJA, and other 9-mo streamflow (SI Appendix, Fig. S18 F and G), ensuring that it can effectively capture the interannual variability in streamflow.

Streamflow Reconstruction.

We used a linear regression model described below to reconstruct the annual streamflow at Farakka between 700 and 2012. In this model, Y_t represents the log-transformed streamflow for year t, α is the intercept, β is the slope, and X is the matrix of retained principal vectors of the MADA-PDSI predictors. The regression equation is as follows:

$$Y_t=\alpha +\beta \mathrm{X}.$$

We applied weighted principal component analysis (PCA) to the selected PDSI predictors to reduce the dimensionality of the predictors and explicitly incorporate covariance between annual streamflow and PDSI predictors (38). The weights were determined based on the strength of the correlations between the PDSI values and the streamflow data. The expression for applying weightage is given as follows:

$$X_i=x_i{r_i}^w.$$

Here, x_i and X_i are the original and weighted PDSI for the i-th PDSI predictor time series, r_i is the Pearson correlation coefficient between the i-th PDSI predictor and the streamflow, and w is the current weight being applied. We considered eight weighting factors (0, 0.1, 0.25, 0.5, 0.67, 1.0, 1.5, 2.0), the same as those used by Rao et al. (38) in reconstructing streamflow for Brahmaputra. For each searching radius (1,000 to 3,000 km) and weighting factor (0 to 2.0), we performed PCA on the weighted PDSI matrix and retained those principal vectors with eigenvalues greater than 1. Given the large number of retained principal components (PCs), we used the variable selection algorithm from Random Forest (RF) to select the parsimonious subset of the key PCs for predicting streamflow based on their importance scores and cross-validation skills (93, 94). The RF algorithm assigns an importance score to each PC, and we excluded PCs with importance scores less than 0 in our analysis, ensuring that only those positively contributing to streamflow prediction were included.

We used the leave-ten-out cross-validation approach to validate the model rather than the split-sample approach due to fewer data points. The leave-n-out approach has been efficient in the rigorous validation of the model in streamflow reconstruction studies (29, 45, 46). We calculated four model calibration-validation metrics: i) reduction of error (RE) during validation, ii) Nash-Sutcliffe efficiency (NSE) during validation, iii) Kling-Gupta efficiency (KGE) during validation, and iv) square of Pearson correlation during validation (VRSQ). We developed 50 such bootstraps using the leave-ten-out approach and estimated the median RE, NSE, KGE, and VRSQ of these 50 bootstraps. Finally, for each radius and for each weighting factor, we used the trained model to reconstruct streamflow for the entire period of available PDSI data (700−2012). Thus, five different search radii and eight different weighting factors give us a total of 40 ensemble reconstructions. The number of retained principal vectors (eigenvalues greater than 1) and a parsimonious subset of those retained principal vectors selected using the Random Forest algorithm for each searching radius and for each weighting factor are provided in SI Appendix, Fig. S19. We considered the median value of these 40 ensemble reconstructions as the final long-term series of reconstructed streamflow for the period 700−2012. We extended the reconstructed streamflow series up to 2020 using VIC-simulated streamflow between 2013 and 2020. The flow chart of the above methodological framework to reconstruct past streamflow is shown in SI Appendix, Fig. S1.

The reconstructed annual streamflow explains 69% of the variability in the well-calibrated VIC-simulated streamflow (observed period streamflow) during 1951–2012, while the 95% CI band from the 40-member ensemble accounts for 91% of the variance (Fig. 1A and SI Appendix, Supplementary Text 5). The reconstructed streamflow effectively captures the interannual variability and temporal trend in the VIC-simulated streamflow during the observed period (Fig. 1A). The reconstruction overestimates dry and wet extremes in raw streamflow (SI Appendix, Fig. S20B), which is partly due to inherent biases in the PDSI data. However, this bias is not evident in the standardized streamflow, which is the basis for all subsequent analyses (SI Appendix, Fig. S20 C and D). The box plot of model calibration-validation skills for 40 ensemble reconstructions demonstrates the satisfactory performance of the reconstruction model (Fig. 1B). The skill of each ensemble member is based on the median of the leave-ten-out cross-validation approach with 50 bootstraps. Four reconstruction metrics, including RE, NSE, KGE, and VRSQ, range between 0.42 and 0.72. The median RE based on 40 ensemble reconstructions was 0.56, and the median NSE was 0.50. The comparable values of RE and NSE indicate that the model does not overfit the data during calibration. The median KGE and VRSQ were also 0.59 and 0.60, indicating good reconstruction skills. The RE, NSE, KGE, and VRSQ values exceeding 0.40 for all ensemble reconstructions confirm the model’s effective reconstruction performance. These skills are statistically significant at a 0.05 significance level based on two-tailed t test statistics. The KGE metric, which integrates bias, correlation, and variance, shows a higher median value, indicating that the model performs well across these aspects of streamflow. The median reconstructed streamflow (final series) for the full reconstruction period between 700 and 2012 C.E., along with the 95% CI based on 40 ensemble reconstructions using MADA-PDSI is represented in SI Appendix, Fig. S4A. The median reconstructed streamflow follows the normal distribution based on the Anderson–Darling test and the Lilliefors/Kolmogorov–Smirnov test (at a 5% level of significance).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Observed-IMD and ERA5-reanalysis datasets used in the study can be downloaded from the following links: IMD: https://imdpune.gov.in/lrfindex.php (95), and ERA5 reanalysis on single levels: https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=download (96). CMIP6 data are freely available for download from the Earth System Grid Federation database (https://aims2.llnl.gov/search) (97). CESM-LENS2 data can be downloaded from the Research Data Archive portal (https://rda.ucar.edu/datasets/d651056/dataaccess/#) (98). Observed streamflow for the Ganga River at Farakka station was obtained from the Center for Sustainability and the Global Environment (SAGE) (https://sage.nelson.wisc.edu/riverdata/index.php) (99). Source codes for the VIC model simulations can be downloaded from the GitHub repository (https://github.com/UW-Hydro/VIC) (100). The codes used and analyzed data are published on the Zenodo repository (https://zenodo.org/records/15206221) (101). All other data are included in the manuscript and/or SI Appendix.