Hissan et al. (2025) Predicting long-term meteorological drought using random forest and multi-scale drought indices

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-10-01
• Authors: Riaz Ul Hissan, Nusrat Parveen
• DOI: 10.1007/s00704-025-05784-6

Research Groups

• Department of Geography, Government College University Faisalabad, Punjab, Pakistan

Short Summary

This study assessed meteorological drought patterns in Pakistan from 1960–2023 using SPI and SPEI at multiple timescales and applied a Random Forest model to predict long-term drought, finding that longer-term indices (SPI-12 and SPEI-12) are the most informative for forecasting.

Objective

• To systematically analyze the spatial and temporal trends of meteorological drought in Pakistan using both the SPI and SPEI at five distinct timescales (1, 3, 6, 9, and 12 months) over a 63-year period (1961–2023).
• To employ the Random Forest (RF) model to predict long-term drought conditions and evaluate the predictive importance of each index and timescale.
• To provide a scientific basis for developing drought early warning systems and mitigation strategies specific to the diverse climatic zones of Pakistan.

Study Configuration

• Spatial Scale: Pakistan, covering 796,096 square kilometers, using data from 44 meteorological stations across five climatic zones (arid, semi-arid, extremely arid, humid, and very humid).
• Temporal Scale: 63 years (1961–2023) for monthly meteorological data, analyzed at 1, 3, 6, 9, and 12-month timescales for drought indices.

Methodology and Data

• Models used:
  • Standardized Precipitation Index (SPI)
  • Standardized Precipitation Evapotranspiration Index (SPEI)
  • Mann–Kendall (MK) test for trend analysis
  • Sen’s slope estimator (SSE) for trend magnitude
  • Random Forest (RF) machine learning model for drought prediction and feature importance
• Data sources: Monthly precipitation, minimum temperature, and maximum temperature data from 44 meteorological stations collected from the Pakistan Meteorological Department (PMD).

Main Results

• Time series analysis revealed pronounced variability in meteorological drought patterns, with the impact of temperature on drought severity becoming more pronounced in recent decades, as evidenced by the divergence between SPI and SPEI.
• Trend analysis using Mann–Kendall and Sen’s slope showed significant wetting trends in central and northeastern Pakistan (e.g., Bahawalpur, Muzaffarabad, Sialkot), especially at longer timescales (SPI-12 and SPEI-12). Conversely, western and northern areas (e.g., Dalbandin, Saidu Sharif, Chilas) exhibited marked negative trends, particularly in SPEI, indicating drying conditions.
• The Random Forest model demonstrated strong predictive power for meteorological drought, with R² values for the testing dataset ranging from 0.70 to 0.90.
• SPI-12 and SPEI-12 consistently emerged as the most informative indicators for predicting overall drought conditions, carrying the highest predictive weight across diverse climatic zones. Shorter timescales (SPI-1, SPEI-1) contributed less to model accuracy.
• Spatial patterns of feature importance varied, with SPEI gaining significance in areas more sensitive to evapotranspiration (e.g., northern highlands like Gilgit and Chilas), while SPI-12 and SPEI-12 dominated in arid southern and western stations.

Contributions

• Employs a large and comprehensive dataset spanning 63 years and 44 meteorological stations, which is rare in existing studies for Pakistan.
• Utilizes both non-parametric Mann–Kendall and Sen’s slope tests for robust trend estimations.
• Uses both SPI and SPEI indices across five different timescales (1, 3, 6, 9, and 12 months) to provide a more holistic view of drought characteristics.
• Applies a machine learning-based Random Forest model for meteorological drought prediction with such a large dataset, which is scarce in available literature.

Funding

No Funding received.

Citation

@article{Hissan2025Predicting,
  author = {Hissan, Riaz Ul and Parveen, Nusrat},
  title = {Predicting long-term meteorological drought using random forest and multi-scale drought indices},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05784-6},
  url = {https://doi.org/10.1007/s00704-025-05784-6}
}