Gogineni et al. (2026) An integrated machine learning and decomposition framework for enhanced drought prediction

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-03-24
• Authors: Abhilash Gogineni, Madhusudana Rao Chintalacheruvu
• DOI: 10.1007/s00704-026-06144-8

Research Groups

• Centre for Promotion of Research, Graphic Era (Deemed to be University), Dehradun, Uttarakhand, India
• Department of Civil Engineering, National Institute of Technology, Jamshedpur, India

Short Summary

This study introduces a novel integration-prediction framework combining multiple signal decomposition algorithms with machine learning models for enhanced drought prediction. It found that hybrid decomposition models significantly improved accuracy over standalone models, with the VMD-SVR model consistently demonstrating superior performance across the studied drought-prone regions.

Objective

• Evaluate the performance of standalone, hybrid (decomposition-based), and integrated machine learning models for drought prediction.
• Assess drought prediction skill for short-term (SPI-6) and long-term (SPI-12) indices using long-term data (1975–2020).
• Quantify the performance gains achieved through model integration.

Study Configuration

• Spatial Scale: Four districts in South Bihar, India: Nalanda, Nawada, Jamui, and Gaya. The state of Bihar covers approximately 94.2 thousand square kilometres.
• Temporal Scale: Monthly time series data from 1975 to 2020 (45 years).

Methodology and Data

• Models used:
  • Standalone Machine Learning Models: Support Vector Regression (SVR), Random Forest (RF), XGBoost.
  • Signal Decomposition Techniques: Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD).
  • Hybrid Decomposition Models: EMD-SVR, EEMD-SVR, VMD-SVR, EMD-RF, EEMD-RF, VMD-RF, EMD-XGBoost, EEMD-XGBoost, VMD-XGBoost.
  • Integrated Models: INT-SVR, INT-RF, INT-XGBoost.
  • Drought Index: Standardized Precipitation Index (SPI) at 6-month (SPI-6) and 12-month (SPI-12) timescales.
  • Hyperparameter Optimization: Bayesian optimization.
• Data sources: Monthly precipitation data from multiple gauging stations managed by the India Meteorological Department (IMD) at Shivajinagar, Pune, accessed via the IMD Pune Data Service Portal.

Main Results

• Hybrid decomposition models significantly improved prediction accuracy compared to standalone models, with integration models providing even more significant enhancement.
• The VMD-SVR model consistently demonstrated superior performance across Nawada, Jamui, and Gaya districts, achieving Nash-Sutcliffe Efficiency (NSE) values of 0.968, 0.958, and 0.955, respectively, along with the lowest prediction errors (MAE, RMSE, MSE).
• In Nalanda district, the Integrated SVR (INT-SVR) model performed slightly better, showing the highest NSE of 0.964.
• SVR-based models, particularly VMD-SVR and INT-SVR, proved more effective than RF and XGBoost models.
• Standalone models exhibited the lowest accuracy, especially in predicting drought valley values, with noticeable errors and time shifts.
• Signal decomposition techniques, particularly VMD, significantly enhanced the performance of ML models by isolating critical components and reducing noise. Hybrid models showed an average improvement in NSE values of 10–15% compared to standalone models.

Contributions

• Introduction of a novel ‘integration-prediction’ model that systematically integrates multiple signal decomposition algorithms (EMD, EEMD, VMD) with machine learning models (SVR, RF, XGBoost) for enhanced drought prediction.
• Development of a comprehensive decomposition–machine learning framework that fuses outputs from all decomposed components through integrated models (INT-SVR, INT-RF, INT-XGBoost) to improve prediction robustness and stability.
• Demonstration that hybrid and integrated models significantly outperform standalone machine learning models in capturing complex, nonlinear drought patterns, particularly highlighting the superior performance of VMD-SVR.
• Provision of a more reliable and stable approach for drought prediction, offering significant insights for improving drought management strategies in semi-arid regions like South Bihar.

Funding

No funding was received for conducting this study.

Citation

@article{Gogineni2026integrated,
  author = {Gogineni, Abhilash and Chintalacheruvu, Madhusudana Rao},
  title = {An integrated machine learning and decomposition framework for enhanced drought prediction},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06144-8},
  url = {https://doi.org/10.1007/s00704-026-06144-8}
}