Yu et al. (2026) Flash drought early warning by a two-stage coupled framework integrating Causal-ConvLSTM in the Han River basin

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-11-02
• Authors: Yue Yu, Xini Zha, Lihua Xiong, Xiaokang Xin
• DOI: 10.1016/j.agwat.2026.110915

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, Wuhan, 430072, PR China
• Changjiang Water Resources Protection Institute, Wuhan 430051, PR China

Short Summary

This study develops a two-stage coupled framework for 15-day flash drought early warning in the Han River Basin, integrating a Causal-ConvLSTM model for root-zone soil moisture forecasting and a LightGBM classifier with seasonal optimal thresholds. The framework achieves competitive performance (AUC of 0.788, accuracy of 0.715) and successfully identifies severe flash drought events 15 days in advance, providing a reliable operational tool for drought mitigation.

Objective

• To propose a 15-day short-lead prediction method for flash droughts in the Han River Basin based on the soil moisture cumulative deficit, addressing rapid error propagation in long-series soil moisture forecasting and insufficient adaptation to seasonal heterogeneity.

Study Configuration

• Spatial Scale: Han River Basin, China (106°15′–114°20′E, 30°10′–34°20′N), with a total basin area of 159,000 km².
• Temporal Scale:
  • Data span: January 1, 1981, to December 31, 2022.
  • Training period: 1981–2014 (earliest 80% of pentad time steps).
  • Test period: 2015–2022 (most recent 20% of pentad time steps).
  • Forecast lead time: 15 days (three pentads).
  • Temporal resolution: Daily for Causal-ConvLSTM, pentad (5-day) for flash drought identification and LightGBM.

Methodology and Data

• Models used:
  • Causal-Convolutional Long Short-Term Memory (Causal-ConvLSTM): For daily root-zone soil moisture forecasting, integrating Latent PC-Momentary Conditional Independence (LPCMCI) causal inference unit.
  • Boosted Regression Tree (BRT): For screening season-specific core hydrometeorological driving factors.
  • Light Gradient Boosting Machine (LightGBM): For flash drought occurrence probability prediction.
  • Youden index: For determining seasonal optimal discriminant thresholds.
  • Benchmark models for comparison: Long Short-Term Memory (LSTM), Convolutional Long Short-Term Memory (ConvLSTM), Predictive Recurrent Neural Network Version 2 (PredRNN-V2), Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost).
• Data sources:
  • ERA5-Land reanalysis dataset: Global reanalysis dataset for land applications.
    • Time span: January 1, 1981, to December 31, 2022.
    • Spatial resolution: 0.1° × 0.1°.
    • Temporal resolution: Daily.
    • Variables: Root-zone soil moisture (integrated 0–1000 mm from 0–70 mm, 70–280 mm, 280–1000 mm layers), total precipitation (mm pentad⁻¹), potential evapotranspiration (mm pentad⁻¹), total evaporation (mm pentad⁻¹), 10 m wind speed (m s⁻¹), snowmelt (mm pentad⁻¹), surface net thermal radiation (MJ m⁻² pentad⁻¹), surface air pressure (Pa), 2 m dewpoint temperature (K), snow depth (m), snow water equivalent (m), skin temperature (K), surface net solar radiation (MJ m⁻² pentad⁻¹), surface downward solar radiation (MJ m⁻² pentad⁻¹), surface downward thermal radiation (MJ m⁻² pentad⁻¹), 2 m air temperature (K).

Main Results

• The proposed framework achieved an Area Under the Curve (AUC) of 0.788 and an overall accuracy of 0.715 on the held-out test period (2015–2022).
• The Causal-ConvLSTM model for root-zone soil moisture forecasting attained an R² of 0.89 at 1–5 day lead times, decreasing to 0.71 at 6–10 days and 0.57 at 11–15 days (1–15 day mean: 0.73), demonstrating effective suppression of error propagation.
• The framework successfully identified the severe summer 2022 flash drought event 15 days in advance, accurately capturing its spatiotemporal evolution.
• Seasonal performance was consistently reliable, with spring and summer showing the highest skill (AUC values of 0.806 and 0.764, and accuracy rates of 0.73 and 0.69, respectively).
• The innovative three-pentad cumulative soil moisture deficit feature improved prediction performance, increasing AUC by 12.9%, accuracy by 16.4%, and precision by 25.6% compared to using original soil moisture data.
• LightGBM exhibited the least training time (6.00 seconds) among compared classifiers while maintaining high performance.
• Optimal seasonal discriminant thresholds determined by the Youden index were 0.497 for spring, 0.500 for summer, 0.469 for autumn, and 0.505 for winter, adapting to seasonal differentiation and balancing missed detection and false alarms.

Contributions

• Developed a novel two-stage coupled framework for 15-day flash drought early warning, effectively addressing rapid error propagation in long-series soil moisture forecasting and seasonal heterogeneity.
• Integrated Causal-ConvLSTM with causal priors for accurate daily root-zone soil moisture forecasting, enhancing physical interpretability and suppressing error accumulation over longer lead times.
• Innovatively extracted the three-pentad cumulative soil moisture deficit as a core discriminant feature, improving the model's ability to characterize continuous soil moisture depletion and reducing interference from random fluctuations.
• Implemented seasonal optimal discriminant thresholds based on the Youden index, adapting to the seasonal differentiation characteristics of flash droughts in the Han River Basin and balancing missed detection and false alarm rates.
• Utilized Boosted Regression Tree (BRT) for season-specific core driver screening, reducing input dimensionality while preserving physical interpretability and enhancing model robustness.
• Provided a reliable and operational tool for proactive agricultural drought mitigation, optimal irrigation scheduling, and refined water resources allocation in the Han River Basin and similar semi-humid regions.

Funding

• National Key R&D Program of China (No. 2024YFD1702004)
• Natural Science Foundation of Hubei Province of China (No. 2025AFB319)
• Major Science and Technology Project of the Ministry of Water Resources of China (No. SKS-2025066, SKS-2025050)
• Visiting Researcher Fund Program of State Key Laboratory of Water Resources Engineering and Management (Grant No. 2023SWG04)

Citation

@article{Yu2026Flash,
  author = {Yu, Yue and Zha, Xini and Xiong, Lihua and Xin, Xiaokang},
  title = {Flash drought early warning by a two-stage coupled framework integrating Causal-ConvLSTM in the Han River basin},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110915},
  url = {https://doi.org/10.1016/j.agwat.2026.110915}
}