Seven centuries of reconstructed Brahmaputra River discharge demonstrate underestimated high discharge and flood hazard frequency

The Brahmaputra River, a major component of the Ganga–Brahmaputra–Meghna system, sustains over 60 million people but frequently causes destructive monsoon-season flooding. High July–September (JAS) discharge is driven primarily by upper-basin precipitation with contributions from snow and glacier melt. Despite expectations from theory and models that warming will intensify the South Asian Summer Monsoon and increase river discharge and flood risk, assessments are hampered by a short, fragmented instrumental discharge record (Bahadurabad gauge, 1956–2011). The study aims to: (i) reconstruct a long JAS discharge history using tree rings to provide a baseline of natural variability over seven centuries; (ii) quantify how the frequency of high-discharge (proxy for flood hazard) during the instrumental era compares to the long-term record; and (iii) compare reconstructed variability with CMIP5 model projections to assess changes in high-discharge recurrence under climate change.

Prior work shows aerosol forcing weakened the South Asian Summer Monsoon in the late 20th century, whereas rising greenhouse gases are projected to strengthen it, increasing runoff and flood risk in Himalayan basins. However, Brahmaputra flood hazard assessments are limited by short instrumental records. Tree-ring based hydroclimate and streamflow reconstructions have successfully extended records in many regions, contextualizing recent extremes within long-term variability and informing risk under climate change. Previous studies in the upper Brahmaputra/Tibetan Plateau provide related hydroclimate reconstructions, but a basin-wide, discharge-focused reconstruction for the lower Brahmaputra linked to flood hazard had been lacking.

Tree-ring predictor network: Tree-ring data between 20°N–35°N and 86°E–101°E (within ~670 km of the basin) were obtained from the ITRDB and two new Myanmar series. Series were standardized using the signal-free method to minimize non-climatic growth effects and retain low-frequency variability; chronologies were truncated where EPS > 0.85. Predictors were retained if they correlated with mean JAS Bahadurabad discharge at p < 0.10 (two-tailed), using both raw and pre-whitened series; lagged (t+1) predictors were allowed to account for prior-year growth response. Reconstruction model: A Bayesian linear regression was used to reconstruct annual mean JAS discharge (yt) from principal components (PCs) of the predictor set with eigenvalues >1 (Kaiser–Guttman). Non-informative priors: a ~ N(0,10^1), β ~ N(0,10^1). Predictors were weighted by powers of their calibration-period correlations with discharge across powers p ∈ {0, 0.1, 0.25, 0.5, 0.67, 1.0, 1.5, 2.0}. A nested approach sequentially removed shorter chronologies to maximize reconstruction length. For each nest and each weight, 50 leave-10-out (of 43 years, 1956–1998; 32 calibration, 10 validation) reconstructions were generated, yielding 400 ensemble members; the final reconstruction is their median. Skill metrics included CRSQ, VRSQ, VRE, and VCE. Linking discharge and floods: Observational analysis established a strong relationship between mean JAS discharge and maximum 10-day mean discharge (Pearson r = 0.82; Spearman r = 0.79), validating JAS discharge as a flood-hazard proxy. Historical flood years (1787, 1842, 1858, 1871, 1885, 1892, 1900, 1902, 1906, 1910, 1918, 1922) and instrumental flood years (1966, 1987, 1988, 1998, 2007, 2010) were compiled. Superposed epoch analysis (SEA) with double bootstrapping tested whether flood years coincided with higher reconstructed discharge. Recurrence intervals: High-discharge threshold set at 48,800 m³/s (mean JAS discharge in 2007, the lowest among instrumental flood years). For each dataset—instrumental (1956–2011), reconstruction over instrumental period (1956–2004), full reconstruction (1309–2004), and CMIP5 RCP8.5 periods (2050–2074, 2075–2099)—1,000 bootstrap samples of 30 years with replacement were drawn. In each, the percentile P of the threshold was computed and return interval estimated as 100/(100−P). Climate model projections: Surface runoff upstream of Bahadurabad was extracted from 20 CMIP5 models (42 members) for historical (1850–2005) and RCP8.5 (2006–2099). Runoff fields were bilinearly interpolated to 0.25° and upper-basin regions with isochrones >16 days were lagged by one month. Model runoff z-scores (relative to 1956–1998 mean and SD) were scaled to discharge units using instrumental mean and SD for direct comparison. Multi-model interquartile ranges and kernel densities were analyzed. Teleconnections with ENSO/IOD were examined but showed no robust relationships.

• Mean JAS discharge reconstruction spans 1309–2004 C.E. and exhibits significant skill (median CRSQ: 65.58%; VRSQ: 45.61%; VRE: 0.41; VCE: 0.31).
• The reconstructed long-term mean discharge (46,993 ± 812 m³/s) is significantly higher than the instrumental mean (1956–2011: 43,350 m³/s); difference = 3,644 m³/s (t = 5.11, p < 0.01). Results hold using 1956–2004 or 1956–1998 instrumental means.
• The early instrumental period (1956–1986) was unusually dry (mean 41,206 m³/s), ranking in the 13th percentile of the 696-year reconstruction; 1956–2004 and 1956–2011 means rank in the 22nd percentile.
• Observations show strong coupling between mean JAS discharge and maximum 10-day mean discharge (Pearson r = 0.82 [0.72, 0.89]; Spearman r = 0.79 [0.69, 0.86]; n = 55, p < 0.001). Instrumental flood years had prolonged periods above median and 95th-percentile daily flows (median 76 and 17 days of exceedance, respectively, within JAS).
• SEA indicates reconstructed discharge during 16 historical/instrumental flood years (pre-2005) is significantly higher than expected by chance (~1 SD; p < 0.001). Pre-1956 association is weaker and not uniformly significant.
• Notable multidecadal wet periods (~1560–1600, 1750–1800, ~1830–1860) have no analogue in the instrumental era; dry periods occur in early 1400s, late 1600s, early and late 1800s.
• CMIP5 multi-model ensembles show increasing discharge under RCP8.5, with larger median and upper-quartile values by late century; kernel densities suggest future and reconstructed distributions are wetter than the instrumental period.
• Recurrence intervals (years) for exceeding 48,800 m³/s: 4.35 (instrumental 1956–2011), 3.57 (reconstruction over 1956–2004), 2.7 (full reconstruction 1309–2004), 2.5 (RCP8.5 2050–2074), 2.17 (RCP8.5 2075–2099).
• Underestimation of high-discharge/flood-hazard frequency by instrumental-era data relative to long-term reconstruction: 24.37%–37.93%. Relative to future projections (RCP8.5): 42.53%–50.11% more frequent than instrumental-era estimates.
• 1998 reconstructed discharge (~60,312 m³/s) ranks among the highest in seven centuries (exceeded five times), consistent with observed extreme flooding.

The study demonstrates that the instrumental record used to assess Brahmaputra flood hazard coincides with an anomalously dry period, leading to an underestimation of high-discharge frequency when used as a baseline. The tight coupling between seasonal mean JAS discharge and peak 10-day discharge justifies using seasonal discharge as a proxy for flood hazard. Paleo-reconstruction reveals higher frequencies of high-discharge years historically than observed in recent decades, and CMIP5 projections indicate further increases in median-to-high flow regimes under continued warming. Together, these findings indicate that both natural variability and anthropogenic climate change contribute to a higher likelihood of high-discharge conditions than suggested by the short instrumental record. The absence of robust ENSO/IOD teleconnections underscores the dominant role of regional monsoon precipitation variability. These insights underscore the need to incorporate long-term paleoclimate context and forward-looking projections into flood hazard assessment and risk management for the Brahmaputra basin.

This work provides a seven-century tree-ring reconstruction of Brahmaputra JAS discharge, quantifies that the instrumental era (especially 1956–1986) is among the driest of the last 700 years, and shows that reliance on recent observations alone underestimates the frequency of high-discharge conditions by about 24–38% relative to natural variability, with climate change likely increasing this underestimation to ~43–50% by late century. The reconstruction, combined with CMIP5 projections, offers a more robust baseline for flood hazard estimation. Future research should: (i) integrate multiple paleohydrological proxies (e.g., flood-scar dendrochronology, geomorphic stratigraphy, speleothems) to reconstruct flood event magnitudes and frequencies; (ii) expand and update regional tree-ring networks, including non-traditional species; (iii) link hydrologic hazard to exposure and vulnerability for comprehensive risk assessment; and (iv) improve transboundary data sharing and real-time discharge monitoring to enhance early warning and adaptive management.

• The reconstruction targets mean JAS discharge, not explicit flood events; translation from high discharge to flood occurrence depends on additional factors (e.g., rainfall intensity/distribution, antecedent soil moisture, land use, infrastructure, policy), which were not modeled.
• The association between reconstructed high discharge and historical flood reports is weaker prior to the instrumental period, possibly due to uncertainties in historical documentation (e.g., events pertaining to broader Bengal floods) and lack of flood magnitude details.
• Instrumental discharge data end in 2011; limited recent observations constrain contextualization of the most recent decades.
• CMIP5-based projections rely on runoff diagnostics and scaling assumptions (e.g., isochrone lagging, z-score scaling to instrumental stats) and on a high-emissions scenario (RCP8.5); model and scenario uncertainties remain.
• Potential future changes in flood management, land use, and infrastructure, which could alter flood risk outcomes, are not accounted for.
• No strong, consistent teleconnections with ENSO/IOD were found, limiting external predictors for interannual variability in forecasting contexts.