WANG et al. (2026) Extreme value combination prediction of summer–autumn drought–flood abrupt alternation in the upper Hanjiang River based on multiple machine learning ensembles
Identification
- Journal: Scientific Reports
- Year: 2026
- Date: 2026-11-10
- Authors: Le WANG, Yubin Chen, Yimin Ma, Hui Qiu
- DOI: 10.1038/s41598-026-79576-x
Research Groups
- Bureau of Hydrology, Changjiang Water Resources Commission, Wuhan, China
- Innovation Team for Flood and Drought Disaster Prevention of Changjiang Water Resources Commission, Wuhan, China
Short Summary
This study developed an extreme-value combination strategy based on multiple machine learning ensembles to improve seasonal prediction of summer-autumn drought-flood abrupt alternation (DFAA) in the upper Hanjiang River. The proposed method achieved superior performance (Nash–Sutcliffe efficiency coefficient (NSE) of 0.62, Root Mean Square Error (RMSE) of 0.17) compared to other ensemble methods and raw dynamical models, particularly in capturing extreme events.
Objective
- To improve the seasonal prediction accuracy of summer-autumn drought-flood abrupt alternation (DFAA) events in the upper Hanjiang River using a novel extreme-value combination strategy based on multiple machine learning ensembles.
Study Configuration
- Spatial Scale: Upper Hanjiang River Basin, China.
- Temporal Scale:
- Data period: 1961–2022.
- Training period: 1962–1998.
- Validation period: 1999–2022 (further split into calibration: 1999–2010 and test: 2011–2022).
- Prediction target: Summer-autumn (July–August to September–October) DFAA events.
Methodology and Data
- Models used:
- Machine Learning: Random Forest (RF), Support Vector Machine (SVM), Long Short-Term Memory (LSTM).
- Optimization: Particle Swarm Optimization (PSO) for hyperparameter tuning.
- Ensemble methods: Extreme-value combination strategy (proposed), Ensemble Mean, Bayesian Model Averaging (BMA), Best Single Member, Quantile Regression (QR), Stacking.
- Factor selection: Maximal Information Coefficient (MIC), cross-correlation tests.
- DFAA index: Modified Standardized Drought-Flood Abrupt Alternation Index (MSDFAI).
- Statistical tests: Mann-Kendall (MK) trend test, KPASS test, Hurst exponent, Gaussian kernel maximum mean discrepancy (MMD), Diebold–Mariano (DM) test, paired t-test, Wilcoxon signed-rank test.
- Data sources:
- Climate factors: Monthly data of 130 global climate factors (1961–2022) from the National Climate Center of China.
- Precipitation data: Daily precipitation data (1961–2022) from 34 rain-gauge stations in the upper Hanjiang River, aggregated to monthly totals; areal precipitation calculated using the Thiessen polygon method.
- Climate model data (for comparison): ECMWF SEAS5, BCC-CPSv3, CFSv2 hindcasts (2010–2022, initialized in early August for August–October forecasts).
Main Results
- The summer-autumn MSDFAI series in the upper Hanjiang River is stationary with no significant trend, but exhibits leptokurtic, fat-tailed, and negatively skewed characteristics, indicating strong extreme events. The Hurst exponent (0.782 for training, 0.668 for full period) suggests limited predictability through series extrapolation.
- MIC screening identified 30 key climate factors, with the Eurasian Meridional Circulation Index (November of previous year) having the highest MIC (0.51). Other factors showed relatively balanced influences (MIC values between 0.35 and 0.45).
- During the training period (1962–1998), the PSO-RF, PSO-SVM, and PSO-LSTM models achieved high NSE values (0.76–0.87) and low RMSE values (0.11–0.15).
- Model performance declined during the validation period (1999–2022), with NSE values ranging from 0.47 to 0.56 and RMSE values from 0.24 to 0.28, primarily due to non-stationarity in the factor-predictand relationship and scarcity of extreme event samples. PSO-SVM showed relatively optimal performance among single models.
- The proposed extreme-value combination strategy achieved a test-period (2011–2022) NSE of 0.62 and RMSE of 0.17, significantly outperforming other ensemble methods: ensemble mean (NSE 0.56, RMSE 0.18), best single member (NSE 0.48, RMSE 0.20), Bayesian Model Averaging (NSE 0.36, RMSE 0.22), quantile regression (NSE 0.33, RMSE 0.23), and stacking (NSE 0.12, RMSE 0.26).
- The extreme-value combination strategy also substantially outperformed raw dynamical models (ECMWF SEAS5: NSE 0.28, RMSE 0.31; BCC-CPSv3: NSE 0.26, RMSE 0.33; CFSv2: NSE 0.25, RMSE 0.33) during the test period.
- Statistical significance tests (paired t-test and Wilcoxon signed-rank test) confirmed that the extreme-value combination strategy significantly outperformed other benchmark methods during the test period (p < 0.05).
Contributions
- Proposed and validated an extreme-value combination strategy for multi-model ensemble prediction, effectively mitigating the systematic underestimation of extreme DFAA events.
- Demonstrated that this strategy significantly improves seasonal DFAA prediction accuracy in the upper Hanjiang River, particularly for extreme events, compared to conventional machine learning ensembles and raw dynamical climate models.
- Provided a robust technical approach for seasonal DFAA prediction, enhancing early warning capabilities for extreme events and offering practical value for regional water resources management and flood control.
- Identified and analyzed the importance of key climate factors influencing summer-autumn DFAA using advanced screening methods, highlighting the complex, multi-factor nature of these events.
Funding
- The National Natural Science Foundation of China (No. U2340205, U2340213)
- Major Science and Technology Program of the Ministry of Water Resources of China (SKS-2025029)
Citation
@article{WANG2026Extreme,
author = {WANG, Le and Chen, Yubin and Ma, Yimin and Qiu, Hui},
title = {Extreme value combination prediction of summer–autumn drought–flood abrupt alternation in the upper Hanjiang River based on multiple machine learning ensembles},
journal = {Scientific Reports},
year = {2026},
doi = {10.1038/s41598-026-79576-x},
url = {https://doi.org/10.1038/s41598-026-79576-x}
}
Original Source: https://doi.org/10.1038/s41598-026-79576-x