Lee et al. (2025) LSTM-Based Prediction and Evaluation of Meteorological Drought Indices Considering Cumulative Precipitation Timescale Combinations

Identification

• Journal: Korean Society of Hazard Mitigation
• Year: 2025
• Date: 2025-12-24
• Authors: Seo Yun Lee, Changhyun Jun, Do Guen Yoo
• DOI: 10.9798/kosham.2025.25.6.159

Research Groups

• The University of Suwon, Department of Civil Engineering / Department of Civil and Environmental Engineering
• Korea University, School of Civil, Environmental and Architectural Engineering

Short Summary

This study developed and evaluated Long Short-Term Memory (LSTM) models for predicting meteorological drought indices (SPI6 and SPEI6) in Gwangju Metropolitan City, comparing univariate and multivariate input configurations. The models demonstrated stable predictive performance, with multivariate inputs generally showing improved statistical accuracy and SPI exhibiting stronger agreement with actual drought stages.

Objective

• To develop and evaluate LSTM-based models for forecasting the 6-month Standardized Precipitation Index (SPI6) and Standardized Precipitation Evapotranspiration Index (SPEI6).
• To compare the predictive performance of univariate models (using a single-timescale drought index as input) against multivariate models (integrating multiple timescale drought indices as input).
• To assess the practical applicability of the predicted drought stages by comparing them with the official drought warning system of the National Drought Information Statistics System.

Study Configuration

• Spatial Scale: Gwangju Metropolitan City, Republic of Korea.
• Temporal Scale:
  • Data Period: SPI data from 1 January 1991 to 31 December 2023 (33 years); SPEI data from 1 January 2018 to 31 December 2023 (6 years).
  • Prediction Lead Time: 7 days into the future.
  • Input Window: Past 80 days of data for SPI models; past 60 days for SPEI univariate model; past 40 days for SPEI multivariate model.
  • Target Indices: 6-month Standardized Precipitation Index (SPI6) and 6-month Standardized Precipitation Evapotranspiration Index (SPEI6).
  • Input Indices for Multivariate Models: 3-month, 6-month, 9-month, and 12-month SPI/SPEI.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) deep learning models.
  • SPI Univariate Model: 2 LSTM layers (128 units, tanh activation), dropout, batch normalization.
  • SPI Multivariate Model: 2 LSTM layers (128 units, tanh activation), dropout, batch normalization.
  • SPEI Univariate Model: 1 LSTM layer (64 units, tanh activation), dropout.
  • SPEI Multivariate Model: 2 LSTM layers (100 units, tanh activation), dropout, batch normalization.
  • Optimization: Adam optimizer.
  • Loss Function: Mean Squared Error (MSE).
  • Regularization: Early stopping.
• Data sources: Daily SPI and SPEI data from the Korea Meteorological Administration (KMA) Meteorological Data Open Portal.
• Data Pre-processing: K-nearest neighbor (KNN) imputation (k=5) for missing values, Min-Max normalization (scaling data to 0-1 range).
• Performance Metrics: Coefficient of Determination (R²), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), Nash-Sutcliffe Model Efficiency coefficient (NSME), Correlation, Accuracy, F1-score (F1-macro, F1-weighted), and Cohen's Kappa.
• Practical Validation: Comparison of predicted drought stages with the official drought warning system from the National Drought Information Statistics System (Joint Ministries, 2025), which classifies drought into 5 stages (Normal, Mild, Moderate, Severe, Extreme) based on SPI6 thresholds.

Main Results

• All developed LSTM models demonstrated high predictive accuracy for SPI6 and SPEI6, with R² values consistently above 0.89, NSME values above 0.89, and correlation coefficients above 0.99 between predicted and actual values.
• Multivariate input configurations generally yielded slightly better statistical performance compared to univariate models, particularly for SPI. For SPI, the multivariate model showed an R² of 0.9052 (vs. 0.8987 for univariate) and reduced RMSE by approximately 3.3%. For SPEI, the multivariate model showed an R² of 0.9126 (vs. 0.9013 for univariate) and reduced RMSE by approximately 4.0%.
• When comparing predicted drought stages against actual observed stages, the SPI univariate model showed perfect agreement (Accuracy = 1.0, Kappa = 1.0), while other models also showed high agreement (e.g., SPI multivariate Accuracy = 0.917, Kappa = 0.833).
• When comparing predicted drought stages against official drought warning stages, SPEI models showed relatively higher agreement than SPI models. The SPEI multivariate model achieved an Accuracy of 0.833 and Kappa of 0.294, suggesting better alignment with the operational system's patterns.
• The study highlights that while multivariate models offer improved statistical accuracy, univariate models can also provide high reliability and stability, especially when considering consistency with existing operational systems.

Contributions

• Systematically compared the performance of univariate and multivariate input configurations for LSTM-based meteorological drought index (SPI and SPEI) prediction.
• Demonstrated the effectiveness of using pre-calculated drought indices as direct inputs to deep learning models, thereby simplifying the modeling process and focusing on time-series pattern learning.
• Provided a practical validation of deep learning drought prediction models by comparing their output with the official national drought warning system, offering insights into their operational applicability.
• Offered foundational data and guidance for designing inputs and implementing deep learning-based drought prediction models in real-world applications.

Funding

• Ministry of Environment (Republic of Korea)
• Korea Environmental Industry & Technology Institute (KEITI)
• Drought Response Water Management Innovation Technology Development Project (Reference Code: RS-2022-KE002032)

Citation

@article{Lee2025LSTMBased,
  author = {Lee, Seo Yun and Jun, Changhyun and Yoo, Do Guen},
  title = {LSTM-Based Prediction and Evaluation of Meteorological Drought Indices Considering Cumulative Precipitation Timescale Combinations},
  journal = {Korean Society of Hazard Mitigation},
  year = {2025},
  doi = {10.9798/kosham.2025.25.6.159},
  url = {https://doi.org/10.9798/kosham.2025.25.6.159}
}