Zhu et al. (2026) A hybrid VMD–BiLSTM–XGBoost approach for multi-scale drought forecasting in urbanizing monsoon transition zones

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-04-01
• Authors: Zhe Zhu, W. Lin, Fujiang Liu, Yan Guo, Bo Li
• DOI: 10.1007/s11269-026-04621-3

Research Groups

• School of Geography and Information Engineering, China University of Geosciences (Wuhan), China
• Hubei Key Laboratory of Regional Ecology and Environmental Change, School of Geography and Information Engineering, China University of Geosciences, China
• School of Computer Science, China University of Geosciences (Wuhan), China

Short Summary

This study proposes a hybrid VMD–BiLSTM–XGBoost model for multi-scale meteorological drought forecasting in urbanizing monsoon transition zones. The model effectively disentangles non-stationary drought signals and adaptively captures both long-term trends and short-term extremes, demonstrating superior accuracy and reliability compared to benchmark models.

Objective

• To improve the accuracy and reliability of multi-scale meteorological drought forecasting in urbanizing monsoon transition zones by developing a hybrid VMD–BiLSTM–XGBoost model that synergistically integrates signal decomposition, deep learning, and ensemble machine learning.

Study Configuration

• Spatial Scale: Three urban meteorological stations in Henan Province, China (Anyang, Zhengzhou, and Xinyang), representing a climatic and urbanization gradient.
• Temporal Scale: Daily Standardized Precipitation Evapotranspiration Index (SPEI) data from 1961 to 2018, with forecasting at 1-month, 3-month, 6-month, and 12-month timescales. The training period was 1961–2007, and the independent test period was 2008–2018.

Methodology and Data

• Models used:
  • Proposed: Hybrid VMD–BiLSTM–XGBoost (Variational Mode Decomposition, Bidirectional Long Short-Term Memory, Extreme Gradient Boosting).
  • Components: VMD for signal decomposition, BiLSTM for low-frequency components, XGBoost for high-frequency components.
  • Benchmark models: VMD–ARIMA, VMD–LSTM, VMD–BiLSTM.
  • Potential evapotranspiration (PET) estimated using the Hargreaves model.
  • SPEI normalized with the generalized extreme value (GEV) distribution.
• Data sources:
  • Daily multi-scale Standardized Precipitation Evapotranspiration Index (SPEI) dataset from Wang et al. (2021), published in Earth System Science Data.
  • The dataset covers 427 national meteorological stations in mainland China from 1961 to 2018, derived from quality-controlled daily observations of precipitation, temperature, and sunlight duration.

Main Results

• The VMD–BiLSTM–XGBoost model consistently outperformed benchmark alternatives (VMD–ARIMA, VMD–LSTM, and VMD–BiLSTM) across all study sites and time scales.
• It achieved an R² of 0.997 and a Mean Squared Error (MSE) of 0.001 for 12-month SPEI forecasting at the Anyang station.
• Average R² values exceeded 0.97 across all three cities (Anyang, Zhengzhou, and Xinyang).
• The model demonstrated superior fitting capacity, closely tracking actual values and accurately capturing extreme drought events.
• It consistently achieved the highest Drought-class Accuracy (DAC), reliably classifying drought events into correct severity categories.
• Diebold–Mariano test confirmed statistically significant improvements (p < 0.05) across all stations and scales compared to other models.

Contributions

• Proposes a novel frequency-adaptive hybrid prediction framework (VMD–BiLSTM–XGBoost) that addresses the limitation of uniformly handling decomposed sub-signals by adaptively allocating low-frequency components to BiLSTM and high-frequency components to XGBoost.
• Demonstrates superior accuracy, stability, and robustness in multi-scale meteorological drought forecasting in complex urbanizing monsoon transition zones.
• Provides a valuable decision-support tool for enhancing urban climate resilience, informing adaptive water governance, and advancing drought early-warning systems in regions vulnerable to hydroclimatic extremes.

Funding

• CAS-ANSO Sustainable Development Research Project (Grant No. CAS-ANSO-SDRP-2024-01)
• Remote sensing monitoring of ecosystems in the Sanjiangyuan and surrounding areas (Grant No. XNZX-2023-D0085)
• Open Fund of State Key Laboratory of Remote Sensing Science (Grant No. 6142A01210404)
• Hubei Key Laboratory of Intelligent Geo-Information Processing (Grant No. KLIGIP-2022-B03)
• Metallogenic patterns and mineralization predictions for the Daping gold deposit in Yuanyang County, Yunnan Province (Grant No. 2022026821)

Citation

@article{Zhu2026hybrid,
  author = {Zhu, Zhe and Lin, W. and Liu, Fujiang and Guo, Yan and Li, Bo},
  title = {A hybrid VMD–BiLSTM–XGBoost approach for multi-scale drought forecasting in urbanizing monsoon transition zones},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04621-3},
  url = {https://doi.org/10.1007/s11269-026-04621-3}
}