Seasonal catchment memory of high mountain rivers in the Tibetan Plateau

The study addresses how high mountain catchments in the Tibetan Plateau memorize precipitation at seasonal scales and how this memory affects streamflow predictability. The context is increasing hydrological extremes and the need for skillful seasonal forecasts to support water resources and food security across Asia. Catchment memory—manifested through storage in soil, snow/ice, lakes, and groundwater—extends streamflow predictability beyond rainfall forecasts. Prior approaches have relied on lagged correlations among hydrologic variables or single-value indices (e.g., baseflow index, groundwater recession), which do not fully capture memory processes or duration. GRACE/GRACE-FO satellite data provide robust observations of terrestrial water storage (TWS), enabling improved analysis of memory. The study aims to disentangle the seasonal linkage among precipitation, streamflow, and TWS, and to quantify both the process and duration of seasonal catchment memory via a new precipitation-to-TWSC model and two metrics (influence time and domination time).

Previous research has characterized catchment memory using lagged correlations between flow signatures and antecedent flows or between groundwater, precipitation, and runoff, providing simple associations but lacking functional relationships. Hysteresis loops among precipitation, streamflow, and storage visualize memory qualitatively. A Gamma-based forgetting curve was proposed to quantify multiyear memory processes but not duration. Hydrological models have been used to simulate storage, while GRACE TWS anomalies offer observation-based storage changes applicable over the Tibetan Plateau. Studies have related GRACE-derived groundwater/storage variations to precipitation to infer short- and long-term memory, and used GRACE to develop indices like flood potential. There remains a need to: (1) clarify precipitation–streamflow–TWS linkages at seasonal scales; and (2) quantify seasonal memory with both process and duration metrics.

- Study area and classification: Eight upstream river basins in the Tibetan Plateau were analyzed: upper Brahmaputra (UBR), Salween (SWR), Lancang (LCR), upper Yangtze (YTR), upper Yellow (YLR), upper Tarim (TRM), upper Indus (UIR), and upper Amu Darya (AMU). Basins were grouped into precipitation-dominated (UBR, SWR, LCR, YTR, YLR) and non-precipitation-dominated/westerlies-dominated (TRM, UIR, AMU). Seasons were defined as pre-monsoon (Apr–May), monsoon (Jul–Sep), post-monsoon (Oct–Nov), and winter (Dec–Mar). - Hydrological modeling (VIC): A modified VIC v4.2.d model (0.25° resolution, 6-hourly) with snow and frozen soil modules was used to simulate runoff and evapotranspiration (ET). A glacier melt module (temperature-index) was added. Model calibration used the E-NSGA-II multi-objective algorithm with NSE and relative bias (BIAS) criteria. Key calibrated parameters included soil layer thicknesses (d2, d3), baseflow parameters (Ws, Dsmax, Ds), infiltration curve bin, and glacier degree-day factor D. Forcing data primarily CMFD; for UIR/AMU, precipitation from IMERG and other meteorological variables from GLDAS. Observed streamflow from national agencies; ET from regional/global products. - TWSC estimation: GRACE/GRACE-FO mascon solutions (CSR RL06, JPL RL06) provided TWS anomalies (TWSA). TWSC was derived via central difference: TWSC(t) = (TWSA(t+1) − TWSA(t−1))/2, with uncertainties propagated accordingly. VIC-derived TWSC used the water budget TWSC = P − ET − R. - Precipitation-to-TWSC conceptual model: Monthly precipitation P(t) is partitioned into immediate runoff and a stored fraction released over subsequent months. The released water from P(t) at lag Δt is P(Δt) = Cr,Δt P(t), with ΣΔt=0..n Cr,Δt = 1 and n = 11 months. The forgetting ratios Cr,Δt decay with lag per a recession-like memory function controlled by parameter bt. Streamflow is modeled as R(t) = ΣΔt P(Δt) − ET(t) + εt, where εt captures additional water fluxes beyond precipitation and ET (e.g., long-term glacier/groundwater changes, human influences). Substituting R(t) into the water balance yields TWSC(t) = P(t) − ΣΔt P(Δt) − εt. In non-precipitation-dominated basins, ε is updated by a temperature-index function ε = αi Ti + εi to capture temperature control on snow/ice storage and melt. - Model calibration/validation: The precipitation-to-TWSC model parameters were calibrated with NSGA-II using 2003–2011 (calibration) and 2012–2018 (validation). Performance was evaluated against GRACE-derived TWSC using multiple precipitation datasets (CMFD, CGDPA, TRMM 3B42v7, IMERG, ERA5, CMORPH, PERSIANN) to quantify input uncertainty; temperature from GLDAS for UIR/AMU. Parameter uncertainty and sensitivity to precipitation inputs were assessed; uncertainties in performance were generally <0.1 among datasets (except some sensitivity in TRM). - Hysteresis analysis: Monthly mean hysteresis loops (2003–2018) among TWSA, streamflow, and precipitation (S-Q, P-S, P-Q) were constructed to visualize seasonal memory and transformation dynamics; anomalies in loop directions were related to precipitation anomalies. - Memory duration metrics: Two metrics quantify seasonal memory duration using contribution rates of precipitation to future releases: (1) Influence time: the maximum lag Δt for which the contribution rate CRt,Δt ≥ 1%; (2) Domination time: maximum lag Δt for which CRt,Δt ≥ 10%. Metrics were computed for sub-basins and seasons to capture intra-annual variability. - Data sources: CMFD (primary forcing), IMERG/GLDAS (UIR/AMU), multiple precipitation products for model uncertainty, SRTM DEM, HWSD soils, land cover datasets (WestDC v2.0, GLC2000), glacier inventories, ET products (TED-TP, REA-ET, ETMonitor), and reconstructed TWS datasets to fill GRACE/GRACE-FO gaps in 2017.

- Basin classification and controls: Precipitation-dominated basins (UBR, SWR, LCR, YTR, YLR) display seasonal dynamics primarily driven by precipitation; non-precipitation-dominated basins (TRM, UIR, AMU) are strongly controlled by temperature via snow/ice storage and melt. The UIR has the highest glacier area fraction (~14.91%). - Hysteresis relationships: In precipitation-dominated basins, mean annual loops typically show S-Q clockwise, P-S anticlockwise, and P-Q anticlockwise, indicating precipitation is partly stored then later released to streamflow. Loop direction anomalies occur in drought/deficit years (e.g., UBR in 2009, 2015) linked to reduced monsoon precipitation (e.g., July deficits up to −58%). In non-precipitation-dominated basins, S-Q anticlockwise loops and winter-spring increases in TWSA without concurrent streamflow increases reflect snow/ice storage; temperature correlates strongly with TWSA and streamflow, with minimal hysteresis. - Model performance (VIC vs GRACE): VIC-derived TWSC generally matches GRACE-derived TWSC; e.g., AMU r = 0.93 (highest) and SWR r = 0.44 (lowest), with GRACE uncertainty envelopes encompassing VIC TWSC. Streamflow and ET simulations achieved NSE mostly >0.8 and |BIAS| < 20% in calibration/validation across gauges (Table 1). - Precipitation-to-TWSC model performance: The precipitation-based model improves TWSC estimation compared to VIC in most basins, with correlation coefficients typically increasing from >0.5 to >0.8 against GRACE. Reported examples include highest r = 0.88 in LCR and lowest r = 0.57 in YLR among precipitation-dominated basins; revised temperature-index model achieved r = 0.93 in AMU. Performance was robust across different precipitation datasets (uncertainty <0.1 in r across inputs). - Memory curve characteristics: Precipitation-dominated basins separate into two groups by curve shape/duration. (1) UBR, SWR, LCR: lower immediate forgetting at 0-month lag and longer memory; (2) YTR, YLR: short memory (<4 months). Across all five, <10% of a given month’s precipitation remains memorized after 4 months. Basins with positive ε (UBR, SWR, LCR) align with decreasing TWSA trends; those with negative ε (YTR, YLR) align with increasing TWSA, consistent with reported regional mass trends. - Non-precipitation-dominated basins: TRM exhibits a steep memory curve with short precipitation memory (~3 months) during warm season (June–September), while UIR and AMU show flat curves, indicating weak seasonal precipitation memory but strong long-term memory driven by melt processes. - Seasonal memory duration metrics: In precipitation-dominated basins, influence time grows from May and declines after September/October; domination time lags slightly, increasing from June. During monsoon, influence time ≈ 6 months and domination time ≈ 4 months. In winter, influence ≈ 3 months and domination ≈ 1 month. TRM shows similar durations to winter conditions (influence ~3 months, domination ~1 month) during warm-season precipitation memory. Influence/domination times were not assessed for UIR and AMU due to dominance of long-term melt-driven memory rather than seasonal precipitation memory. - Implications: Seasonal catchment memory can inform efficient lead times for seasonal streamflow forecasting—winter precipitation minimally affects monsoon flows, while monsoon precipitation can influence winter flows.

The findings clarify seasonal storage–runoff transformation in Tibetan Plateau catchments. In precipitation-dominated basins, hysteresis loop geometry reflects the runoff generation mechanism where precipitation governs both TWSA and streamflow. Loop direction anomalies are primarily tied to monsoon precipitation deficits but the long-term loop direction remains stable, reflecting persistent basin storage characteristics. In non-precipitation-dominated basins, strong temperature control leads to winter accumulation of snow/ice (increasing TWSA without streamflow increases) and summer release, producing anticlockwise S-Q loops and weak seasonal precipitation memory; TRM uniquely exhibits mixed precipitation–melt control with a short warm-season precipitation memory. The precipitation-to-TWSC model, including a temperature-index update for melt-dominated basins, captured TWSC dynamics more accurately than VIC-based water balance in most basins, validating the precipitation memory curve as a quantitative descriptor of seasonal memory processes. The new influence and domination time metrics reveal that monsoon precipitation exerts effects over approximately half a year, with a core dominance of around four months, while winter precipitation has much shorter influence/dominance. These insights support identifying practical lead times for seasonal streamflow forecasts and optimizing water resources management under monsoon-driven hydrology and melt influences.

This work proposes a precipitation-to-TWSC framework that quantifies seasonal catchment memory via a precipitation memory curve and two duration metrics (influence and domination times). Across eight Tibetan Plateau basins, precipitation-dominated systems exhibit anticlockwise P-Q and P-S and clockwise S-Q hysteresis loops, with memory durations up to several months; non-precipitation-dominated systems are temperature-controlled with weak seasonal precipitation memory but notable long-term memory. The model reproduces GRACE-observed TWSC with improved accuracy over VIC in most basins and robustly across multiple precipitation datasets. Influence time is about 6 months and domination time about 4 months during monsoon, shrinking to ~3 and ~1 month in winter; TRM shows short warm-season memory. These results can guide effective lead times for seasonal streamflow forecasting and water resources operations. Future research could extend the framework to explicitly separate and quantify contributions of snow, glacier, groundwater, and human activities to ε, address spatial heterogeneity within basins, incorporate time-varying parameters where data permit, and examine long-term (multiyear) memory alongside seasonal processes.

- Scope limitation: For UIR and AMU, seasonal precipitation memory metrics (influence/domination times) were not analyzed due to dominance of melt-driven long-term memory, focusing the study on seasonal processes in precipitation-dominated basins and TRM warm season. - Parameterization constraints: To limit uncertainty with available data, the precipitation memory shape parameter bt and ε were treated as time-invariant in non-precipitation-dominated basins; the longest memory horizon was set to n = 11 months, potentially truncating longer memory components. - Data and model limitations: VIC lacks native glacier dynamics (addressed with a temperature-index module). GRACE–GRACE-FO data gaps (2017) were filled with reconstructed TWS products; uncertainties in GRACE mascons and reconstructions persist. While precipitation dataset variations had small impact on performance overall, TRM showed greater sensitivity. Hysteresis loop anomalies can arise from climate anomalies or human impacts, which are not fully isolated here.