Shi et al. (2026) Spatio – temporal hydrological cycle characteristics in the upper reaches of the Yangtze River: A multi-source remote sensing and machine learning perspective

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-16
• Authors: Yang Shi, Yousheng Zhang, Minglei Hou, Jiahua Wei
• DOI: 10.1016/j.ejrh.2026.103142

Research Groups

• State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing, China
• State Key Laboratory of Plateau Ecology and Agriculture, School of Civil Engineering and Water Resources, Laboratory of Ecological Protection and High Quality Development in the Upper Yellow River, Key Laboratory of Water Ecology Remediation and Protection at Headwater Regions of Big Rivers, Ministry of Water Resources, Qinghai University, Xining, China

Short Summary

This study investigates spatio-temporal hydrological cycle variations in the Upper Yangtze River (URYR) from 1980 to 2030 using multi-source remote sensing and a novel LSTM-AT machine learning model to predict exploitable water resources (EWR). It finds significant regional hydrological heterogeneity and projects a post-2015 decline in basin-wide EWR, with some sub-basins showing increases potentially linked to long-term storage depletion.

Objective

• To investigate the spatio-temporal variations of hydrological cycle components (precipitation, evapotranspiration, runoff, terrestrial water storage) in the Upper Reaches of the Yangtze River (URYR) from 1980 to 2030.
• To develop and apply an attention-enhanced Long Short-Term Memory (LSTM-AT) network for predicting exploitable water resources (EWR) in the URYR.
• To analyze the inter-element correlations, water budget closure, and the influence of natural and human factors on the hydrological cycle and EWR.

Study Configuration

• Spatial Scale: Upper Reaches of the Yangtze River (URYR), China (90°32′–111°33′E, 24°42′–35°55′N), covering approximately 1 million square kilometers. The region is divided into sub-basins: Jinsha River Basin (JS, including upper (UJS) and lower (LJS) sections), Min-Tuo River Basin (MT), Jialing River Basin (JL), Wu River Basin (WJ), and the Yangtze River main stem (YR).
• Temporal Scale: Analysis period: 1980–2015. Prediction period: 2016–2030. Total study period: 1980–2030.

Methodology and Data

• Models used:
  • Long Short-Term Memory network with attention mechanism (LSTM-AT) for exploitable water resources (EWR) prediction.
  • Variational Mode Decomposition (VMD) for sequence decomposition.
  • Multiple Linear Regression (MR) for precipitation correction.
  • Mann-Kendall (M-K) trend test and Pettitt’s method for trend and change-point detection.
  • Morlet wavelet analysis for periodicity.
  • Water balance equation (P − ET − R − TWSC = ε) with Proportional Redistribution (PR) for error correction.
  • VIC v4.2.d land surface hydrological model (for VIC-CN05.1 runoff data).
  • WaterGAP Global Hydrology Model (WGHM) for human activity impacts.
• Data sources:
  • Precipitation: 1-kilometer monthly precipitation dataset for China (derived from Climatic Research Unit (CRU) and WorldClim, corrected with >80 rain gauge stations), China Meteorological Administration (CMA) station data.
  • Evapotranspiration (ET): GLEAM_v3.2a (selected after validation against GLDAS, REA ET, and 73 evaporation stations), CMA station data.
  • Runoff (R): VIC-CN05.1 daily 0.25° terrestrial hydrological dataset for China (simulated using VIC v4.2.d, validated with in-situ and remote sensing products), CMA station data.
  • Terrestrial Water Storage (TWS): Reconstructed total water storage anomaly (TWSA) fields (0.5° resolution, 1979–2020) (validated with GRACE-FO mascon and Satellite Laser Ranging (SLR)).
  • Meteorology: CN05.1 daily dataset (0.25° resolution) (temperature, precipitation, humidity), 79 meteorological stations from China Meteorological Data Sharing Service (temperature, sunshine duration, wind speed, precipitation).
  • Human Activities: WaterGAP Global Hydrology Model (WGHM) outputs (human consumption, reservoir storage).
  • Topography: Digital Elevation Model (DEM) from Geospatial Data Cloud (90 meter resolution).

Main Results

• The URYR exhibits strong spatial heterogeneity in hydrological elements: precipitation ranges from 232.68 millimeters to 1608.61 millimeters, evapotranspiration from 91.74 millimeters to 820.94 millimeters, and runoff up to 725.23 millimeters. Terrestrial water storage peaks at 11.26 millimeters in the upper Jinsha River basin.
• Hydrological elements show clear seasonality with summer peaks. From 1980 to 2015, precipitation generally increased, driving higher evapotranspiration across all sub-basins (Mann-Kendall Z-values up to 11.22). However, runoff declined in the upper Jinsha River (Z ≈ -3.33), influenced by climate change and human activities.
• Precipitation is a strong driver for runoff (correlation up to 0.93) and terrestrial water storage. Evapotranspiration and terrestrial water storage show complex relationships, including negative correlations (up to -0.65) in highly evaporative regions, indicating water reserve depletion during dry periods.
• Water budget closure analysis revealed significant residuals, particularly in high precipitation/runoff regions (Jinsha and Jialing River basins), with absolute errors up to 90.97 millimeters and relative errors exceeding 18% in some sub-basins.
• LSTM-AT model projections indicate a post-2015 decline in basin-wide exploitable water resources (EWR) at -1.95 ± 1.41 millimeters per year. Regionally, the upper Jinsha River shows an anomalous increase (+2.98 ± 0.92 millimeters per year), potentially due to long-term storage depletion (e.g., glacier mass loss), while Min-Tuo River and the Yangtze River main stem show decreases mainly driven by runoff reduction.
• The LSTM-AT model outperformed plain LSTM and Temporal Convolutional Network (TCN), and showed comparable, more stable performance than Transformer models for EWR prediction. Historical EWR records and human activity factors (reservoir storage, human spending on water) were identified as dominant drivers.

Contributions

• Developed and applied a novel attention-enhanced Long Short-Term Memory (LSTM-AT) network for annual-scale prediction of exploitable water resources (EWR) in a complex, highly regulated basin.
• Integrated multi-source remote sensing data, hydrological models, meteorological observations, and human activity indicators into a unified data-driven framework for comprehensive hydrological cycle analysis and EWR forecasting.
• Provided new insights into the dominant hydro-climatic and human-regulation drivers of annual water availability through interpretable feature-weight analysis using the attention mechanism.
• Quantified spatio-temporal variations, trends, and periodicities of key hydrological cycle elements (P, ET, R, TWSC) and evaluated water budget closure in the Upper Yangtze River.
• Highlighted the critical need to distinguish between sustainably renewable and stock-consuming components of EWR for future water resource planning, particularly in glacier- and reservoir-dominated regions.

Funding

• National Key Research and Development Program (Grant No. 2022YFC3202400, 2023YFC3206700)
• Young Elite Scientist Sponsorship Program by China Association for Science and Technology (Grant No. YESS20200094)

Citation

@article{Shi2026Spatio,
  author = {Shi, Yang and Zhang, Yousheng and Hou, Minglei and Wei, Jiahua},
  title = {Spatio – temporal hydrological cycle characteristics in the upper reaches of the Yangtze River: A multi-source remote sensing and machine learning perspective},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103142},
  url = {https://doi.org/10.1016/j.ejrh.2026.103142}
}