Jaiswal et al. (2025) More accurate forecasting of drought indices using a decomposition-based hybrid machine learning model

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-11-01
• Authors: Ronit Jaiswal, Kapil Choudhary, Rajeev Ranjan Kumar, Girish Kumar Jha, Vijay Kamal Meena, N S Sudhakara, Mahesh Kumar Poonia
• DOI: 10.1007/s00704-025-05848-7

Research Groups

• ICAR- Central Institute of Temperate Horticulture, Srinagar, India
• Agriculture University, Jodhpur, Rajasthan, India
• ICAR-Indian Agricultural Statistics Research Institute, New Delhi, India

Short Summary

This study develops a decomposition-based hybrid machine learning framework for more accurate forecasting of precipitation-based drought indices (Effective Drought Index (EDI), Standardized Precipitation Index (SPI) at 3- and 6-month scales) in two drought-prone districts of Maharashtra, India. The Ensemble Empirical Mode Decomposition-Time Delay Neural Network (EEMD-TDNN) hybrid model emerged as the most effective, achieving a 15–30% reduction in Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE) compared to conventional and other hybrid models.

Objective

• To develop and systematically evaluate a comprehensive modelling framework for forecasting precipitation-based drought indices (EDI, SPI-3, and SPI-6) in Ahmednagar and Jalgaon districts, Maharashtra, India.
• To compare the performance of various forecasting models, including Autoregressive Integrated Moving Average (ARIMA), Time Delay Neural Network (TDNN), and Extreme Learning Machine (ELM), with their hybrid variants incorporating Seasonal-Trend Decomposition using Loess (STL) and Ensemble Empirical Mode Decomposition (EEMD).

Study Configuration

• Spatial Scale: Two drought-prone districts in Maharashtra, India: Ahmednagar and Jalgaon.
• Temporal Scale: Monthly rainfall data spanning 42 years (January 1981 to December 2022). Drought indices (EDI, SPI-3, SPI-6) were computed at daily and monthly scales.

Methodology and Data

• Models used:
  • Individual models: Autoregressive Integrated Moving Average (ARIMA), Time Delay Neural Network (TDNN), Extreme Learning Machine (ELM).
  • Decomposition techniques: Seasonal-Trend Decomposition using Loess (STL), Ensemble Empirical Mode Decomposition (EEMD).
  • Hybrid models: STL-ARIMA, STL-ELM, STL-TDNN, EEMD-ARIMA, EEMD-ELM, EEMD-TDNN.
  • Evaluation metrics: Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Diebold-Mariano (DM) test for statistical significance.
• Data sources:
  • Historical monthly rainfall data (January 1981 to December 2022) for Ahmednagar and Jalgaon districts, sourced from the NASA POWER database (https://power.larc.nasa.gov).
  • Derived drought indices: Effective Drought Index (EDI), Standardized Precipitation Index at 3-month scale (SPI-3), and Standardized Precipitation Index at 6-month scale (SPI-6).

Main Results

• The EEMD-TDNN hybrid model consistently demonstrated superior predictive capability across all three drought indices (EDI, SPI-3, SPI-6) and both study districts (Ahmednagar and Jalgaon).
• EEMD-TDNN achieved a quantifiable improvement of 15–30% reduction in RMSE and MAPE across all three drought indices when compared to conventional (ARIMA, ELM, TDNN) and other hybrid models (STL-based, EEMD-ELM).
• Specifically, for Ahmednagar, EEMD-TDNN reduced RMSE by up to 51.92% (D1), 41.89% (D2), and 6.35% (D3) compared to standalone TDNN. For Jalgaon, RMSE reductions were 70.21% (D4), 31.11% (D5), and 33.33% (D6).
• EEMD-based hybrid models significantly outperformed STL-based hybrid models, with EEMD-TDNN showing RMSE reductions of up to 52.83% (D1), 39.44% (D2), and 13.24% (D3) over stlTDNN for Ahmednagar, and 61.64% (D4), 30.34% (D5), and 4.76% (D6) for Jalgaon.
• The Diebold-Mariano test confirmed the statistical significance of EEMD-TDNN's improvement over other models in most cases, particularly for D2, D5, and D6.
• The superior performance is attributed to the synergy between EEMD’s ability to decompose complex, non-stationary time series into intrinsic mode functions and TDNN’s strength in capturing temporal dependencies.

Contributions

• Proposes and systematically evaluates two distinct classes of hybrid forecasting frameworks (STL-based and EEMD-based) integrated with ARIMA, ELM, and TDNN, enabling a comparative analysis of component-based versus frequency-based decomposition techniques.
• Forecasts three widely accepted precipitation-based drought indices (EDI, SPI-3, and SPI-6) to capture short-, medium-, and dynamically evolving drought conditions, providing a holistic view of drought severity.
• Applies hybrid models to district-level datasets from two drought-prone regions in Maharashtra, India (Ahmednagar and Jalgaon), enhancing the operational relevance for climate-vulnerable agroecological zones.
• Identifies the EEMD-TDNN model as consistently demonstrating superior predictive capability, particularly in capturing nonlinear and non-stationary behavior of drought index time series.
• Demonstrates improved model generalizability and forecasting accuracy by effectively addressing high variability and nonlinear patterns in precipitation data through decomposition techniques.
• Enhances decision-support capabilities for agricultural planning, water resource allocation, and drought early warning systems (DEWS) by accurately forecasting short-term, medium-term, and dynamic drought conditions.

Funding

The author(s) received no financial support for this article's research, authorship and/or publication.

Citation

@article{Jaiswal2025More,
  author = {Jaiswal, Ronit and Choudhary, Kapil and Kumar, Rajeev Ranjan and Jha, Girish Kumar and Meena, Vijay Kamal and Sudhakara, N S and Poonia, Mahesh Kumar},
  title = {More accurate forecasting of drought indices using a decomposition-based hybrid machine learning model},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05848-7},
  url = {https://doi.org/10.1007/s00704-025-05848-7}
}